diff --git a/docs/journals/2026-10-12_heldout_token_benchmark.md b/docs/journals/2026-10-12_heldout_token_benchmark.md
new file mode 100644
index 0000000..e7d764e
--- /dev/null
+++ b/docs/journals/2026-10-12_heldout_token_benchmark.md
@@ -0,0 +1,62 @@
+# Held-out token benchmark (2026-10-12)
+
+## Goal
+
+Test whether the 0.983x-against-Python result from #135 holds on tasks the
+built-ins were not shaped for (Lain's rules). No language changes.
+
+## Freeze
+
+The spec commit is **`f510cd0199326e4aacf206e0b08981978808b042`**:
+"heldout-v0: freeze 32 held-out task specs (no solutions)". It was made
+before any reference was written or anything was measured.
+
+| File | SHA-256 |
+|---|---|
+| tasks/README.md | `4441e1e5c4e906a55fe595c382a7d2ef784045e9c1b1230f38887d70bf8318f1` |
+| tasks/build_specs.py | `dcea7d494d8b146ae032456b1f8822364d488c3d8596f7d829aecb21a75148f8` |
+| tasks/manifest.json | `962fa8198b11ac3ff1066abf35be6dcb87ad8a852574edfc46fee17a0ab9e67e` |
+
+`TestHeldoutSpecsFrozen` enforces these hashes.
+
+## Built after the freeze
+
+- `cmd/heldoutverify`:
+  - runs each reference in a fresh sandbox;
+  - HowlFrame uses a governed build, `-run-bc` and scoped `filesystem:read`/`write` grants;
+  - requires the file set to be exactly the inputs plus the output, the inputs unchanged, and the output matching.
+- `count_tokens.py` imports v0's stripping and offline guard unchanged.
+- `run.sh`, `gaps.json` and `verification_test.go`.
+- 32 references each in Python, JS and Go, and 30 in HowlFrame. The other 2 are recorded gaps.
+
+## Result
+
+| | HowlFrame / Python, median per-task ratio | Wins / ties / losses |
+|---|---|---|
+| cl100k_base | **1.309x** | 0 / 1 / 29 |
+| o200k_base | **1.299x** | 0 / 1 / 29 |
+
+Against JS: 1.087x / 1.016x. Against Go: 0.499x / 0.493x.
+
+**The v0 result (0.983x) does not generalize.** CSV is worst at 1.84x and reshape best at 1.08x.
+
+## Gaps and findings
+
+- **h31/h32:** no directory walking. h32 would also need string ordering.
+- **Sugar not expanded in some arguments:** sugar inside `write_file`/`to_lower`
+  arguments fails to lower, because those parents are missing from the sugar contexts.
+- **`or` is not short-circuit.**
+- **`map_get` on a missing key returns `""`.**
+- **Value semantics:** lists taken out of a map must be written back.
+
+## Next (not done here, needs its own poll)
+
+These are the hot spots, but adding them would again shape features on a
+corpus. Any language work should be judged on a *new* held-out set.
+
+- `index_of`/slice;
+- counting or default `map_get`;
+- `map`/`filter`/`sum`/`min`/`max`/`reverse`;
+- dir listing.
+
+There were no live LLM runs and no H1 decision; #90 is unchanged, and so are HFIR and `-compile-bc`.
diff --git a/experiments/token-efficiency-heldout-v0/REPORT.md b/experiments/token-efficiency-heldout-v0/REPORT.md
new file mode 100644
index 0000000..c7e34ed
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/REPORT.md
@@ -0,0 +1,206 @@
+# Held-out token-efficiency benchmark (heldout-v0): measure only
+
+**HowlFrame's v0 token advantage does not generalize.** On a fresh, frozen
+corpus of 32 tasks:
+
+- **Python:** HowlFrame uses **more tokens than Python on 29 of the 30 tasks it can
+  express** and ties on 1, with both tokenizers. The median per-task ratio is
+  **1.309x** (`cl100k_base`) and **1.299x** (`o200k_base`). That compares with
+  **0.983x** on the v0 corpus, which the #135 built-ins were shaped for.
+- **JavaScript:** roughly at parity, at 1.087x and 1.016x.
+- **Go:** HowlFrame still beats standalone Go on every solved task, at about 0.50x.
+- **Gaps:** two tasks, both file walking, cannot be expressed in HowlFrame on
+  main at all.
+
+This measures source size only: no model success, no generation cost, and no
+live LLM/API calls. There is no H1 decision, and #90 stays Partial. No
+built-in, syntax, compiler or VM change is part of this PR.
+
+## Corpus and freeze
+
+- **Size:** 32 tasks in 7 categories (text, csv, aggregate, parse, validate,
+  reshape, walk). None reuses the three v0 release workflows. See
+  [tasks/README.md](tasks/README.md).
+- **Freeze:** the specs (`tasks/manifest.json`, `tasks/build_specs.py`,
+  `tasks/README.md`) were committed in **`f510cd01`** before any reference was
+  written or anything was measured. Their SHA-256 hashes are recorded in the
+  journal and enforced by `TestHeldoutSpecsFrozen`.
+- **Expected outputs:** these come from a spec-time oracle in `build_specs.py`. It is
+  part of the freeze and is never counted or verified as a reference.
+- **Not frozen:** the verifier pre-creates an empty `out/` directory, and the
+  sandbox contains only the task inputs.
+
+## Method
+
+| Element | Setup |
+|---|---|
+| Tokenizers | `tiktoken==0.8.0`, `cl100k_base` and `o200k_base`, offline cache only |
+| Counting and stripping | Imports v0's `count_tokens.py` (`strip_comments`, offline guard) unchanged |
+| Python, JS, Go references | Idiomatic, standard library only, not token-minimized, same style as v0. Go files are standalone `package main` with `//go:build ignore` |
+| HowlFrame references | Features on main `24f6a10f` only |
+
+`cmd/heldoutverify` replays every reference in a fresh sandbox:
+
+- **HowlFrame** goes through `howlframe build --profile governed`, then `-run-bc` with
+  only `filesystem:read=<sandbox>/in` and `filesystem:write=<sandbox>/out`.
+- **Python/Node** run as host processes. **Go** is compiled from the single file and then run.
+
+After each run, the verifier requires all of the following:
+
+- the sandbox's file set is exactly the inputs plus the one output;
+- the inputs are unchanged;
+- the output matches: exact text after removing trailing newlines, or decoded-JSON equality.
+
+Results are written to `results/verification.json`, which records the SHA-256 of each source.
+The counter refuses unverified or changed sources. `TestHeldoutVerificationCoverage` enforces:
+
+- every task × language is either a passing entry with a matching hash or a HowlFrame gap in
+  `gaps.json` that has a reason and no solution file;
+- there are no duplicates.
+
+Reproduce with `experiments/token-efficiency-heldout-v0/run.sh` (`TOKBENCH_PYTHON` selects the interpreter).
+
+Pass counts: **HowlFrame 30/30 expressible (2 gaps), Python 32/32, JavaScript 32/32, Go 32/32.**
+
+## Results
+
+Medians are over the 30 HowlFrame-solved tasks. The all-32 median is shown in parentheses.
+
+| Encoding | HowlFrame | Python | JavaScript | Go |
+| --- | ---: | ---: | ---: | ---: |
+| cl100k_base median | 72.5 | 63.0 (62.5) | 69.0 (69.0) | 162.5 (162.5) |
+| cl100k_base mean | 85.5 | 60.8 | 71.6 | 162.0 |
+| o200k_base median | 72.5 | 63.0 (62.5) | 73.0 (73.0) | 164.5 (164.5) |
+| o200k_base mean | 85.6 | 61.1 | 76.3 | 163.6 |
+
+HowlFrame against each language, over the 30 solved tasks. A ratio above 1 means HowlFrame uses more tokens.
+
+| Encoding | vs | Median per-task ratio | Ratio of medians | HowlFrame wins / ties / losses | Gaps |
+| --- | --- | ---: | ---: | ---: | ---: |
+| cl100k_base | Python | **1.309** | 1.151 | **0 / 1 / 29** | 2 |
+| cl100k_base | JavaScript | 1.087 | 1.051 | 6 / 1 / 23 | 2 |
+| cl100k_base | Go | 0.499 | 0.446 | 30 / 0 / 0 | 2 |
+| o200k_base | Python | **1.299** | 1.151 | **0 / 1 / 29** | 2 |
+| o200k_base | JavaScript | 1.016 | 0.993 | 13 / 0 / 17 | 2 |
+| o200k_base | Go | 0.493 | 0.441 | 30 / 0 / 0 | 2 |
+
+If the two gaps are counted as losses, the result against Python is 0 / 1 / 31 out of 32.
+Comment and blank-line stripping changes only the Go numbers, which go to 0.519/0.513.
+
+**Compared with v0:**
+
+| Corpus | HowlFrame / Python, median per-task ratio | HowlFrame wins against Python |
+|---|---|---|
+| v0 (#135) | 0.983x on both encodings | 44 wins, 16 ties, 0 losses |
+| Held-out | 1.309x / 1.299x | 0 wins, 1 tie, 29 losses |
+
+The v0 result was produced by built-ins tuned on that corpus. It does not
+carry over to unseen workflows.
+
+### By category (`cl100k_base`; `o200k_base` is within 0.1 except where shown)
+
+| Category | n | HowlFrame solved | HowlFrame mean | Python mean | JavaScript mean | Go mean | Median HF/Py ratio |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
+| csv | 5 | 5 | 135.0 | 73.0 | 89.0 | 185.6 | **1.843** (1.817) |
+| aggregate | 5 | 5 | 81.8 | 53.6 | 71.8 | 152.4 | 1.476 |
+| parse | 5 | 5 | 81.2 | 57.8 | 72.0 | 152.8 | 1.397 |
+| text | 5 | 5 | 58.4 | 47.0 | 50.8 | 119.2 | 1.227 (1.200) |
+| validate | 5 | 5 | 84.0 | 71.8 | 80.4 | 187.8 | 1.111 |
+| reshape | 5 | 5 | 72.4 | 61.8 | 65.6 | 174.2 | 1.079 |
+| walk | 2 | **0** | gap | n/a | n/a | n/a | n/a |
+
+### Per task (raw tokens, cl100k_base/o200k_base)
+
+| Task | Category | HowlFrame | Python | JS | Go | HF/Py |
+|---|---|---:|---:|---:|---:|---:|
+| h01 | text | 54/54 | 44/44 | 49/55 | 112/114 | 1.23 |
+| h02 | text | 70/70 | 40/40 | 43/47 | 138/139 | 1.75 |
+| h03 | text | 66/66 | 62/62 | 61/65 | 125/126 | 1.06 |
+| h04 | text | 66/66 | 53/55 | 61/65 | 133/138 | 1.25 |
+| h05 | text | 36/36 | 36/36 | 40/45 | 88/89 | 1.00 |
+| h06 | csv | 122/122 | 71/72 | 74/79 | 165/166 | 1.72 |
+| h07 | csv | 144/144 | 82/84 | 89/96 | 206/211 | 1.76 |
+| h08 | csv | 127/127 | 64/64 | 81/87 | 186/187 | 1.98 |
+| h09 | csv | 129/129 | 70/71 | 96/100 | 185/186 | 1.84 |
+| h10 | csv | 153/153 | 78/78 | 105/110 | 186/187 | 1.96 |
+| h11 | aggregate | 137/137 | 71/71 | 96/101 | 224/225 | 1.93 |
+| h12 | aggregate | 92/92 | 67/67 | 91/97 | 154/156 | 1.37 |
+| h13 | aggregate | 62/62 | 42/42 | 60/64 | 133/134 | 1.48 |
+| h14 | aggregate | 56/56 | 50/50 | 53/57 | 124/125 | 1.12 |
+| h15 | aggregate | 62/62 | 38/38 | 59/63 | 127/128 | 1.63 |
+| h16 | parse | 109/109 | 59/59 | 84/89 | 152/153 | 1.85 |
+| h17 | parse | 88/88 | 63/63 | 86/90 | 145/146 | 1.40 |
+| h18 | parse | 62/62 | 54/54 | 57/61 | 126/127 | 1.15 |
+| h19 | parse | 88/88 | 63/63 | 81/86 | 181/182 | 1.40 |
+| h20 | parse | 59/60 | 50/51 | 52/57 | 160/163 | 1.18 |
+| h21 | validate | 64/65 | 61/61 | 68/72 | 173/174 | 1.05 |
+| h22 | validate | 99/100 | 72/73 | 73/78 | 197/201 | 1.38 |
+| h23 | validate | 80/80 | 72/72 | 83/87 | 154/156 | 1.11 |
+| h24 | validate | 73/73 | 69/69 | 74/80 | 218/220 | 1.06 |
+| h25 | validate | 104/104 | 85/85 | 104/109 | 197/198 | 1.22 |
+| h26 | reshape | 63/63 | 59/59 | 64/68 | 175/176 | 1.07 |
+| h27 | reshape | 68/68 | 63/63 | 61/65 | 181/182 | 1.08 |
+| h28 | reshape | 97/97 | 64/64 | 70/74 | 183/184 | 1.52 |
+| h29 | reshape | 72/72 | 65/66 | 65/69 | 177/178 | 1.11 |
+| h30 | reshape | 62/62 | 58/58 | 68/72 | 155/156 | 1.07 |
+| h31 | walk | gap/gap | 41/41 | 54/58 | 141/142 | gap |
+| h32 | walk | gap/gap | 57/57 | 109/114 | 196/198 | gap |
+
+## Hot spots
+
+These are `cl100k_base` counts of source fragments. They are not an additive attribution.
+
+1. **No index or slice primitives (csv: 1.84x).** Finding a column takes
+   about 19 tokens for the header binding plus a 28-token loop of
+   `for`/`when`/`set`/`inc`. Python does it in 15 tokens with `header.index("email")`.
+   Skipping the header needs a second counter and `(when (> i 0) …)`, 14 tokens
+   against 4 for `lines[1:]`.
+2. **Counting into a map (aggregate, parse, csv).** Each count is
+   `(map_set counts k (+ (if (map_has counts k) (map_get counts k) 0) 1))`,
+   about 25 tokens, against Python's `Counter(` at 4. A missing `map_get` returns
+   `""`, not nil, so `nz` cannot provide a default here.
+3. **Imperative accumulation instead of map/filter/reduce (every category).**
+   - Each transform pays `out (list)` + `(for x xs (append out …))` + `(str_join out "\n")`,
+     about 21 tokens. Python's generator expression inside `"\n".join(…)` costs 12.
+   - Reversal is a 22-token `while`/`dec` loop against `reversed(lines)` at 4.
+   - min/max/sum are manual loops.
+   - There is no `map`, `filter`, `sum`, `min`, `max`, `reverse` or `index_of`.
+
+## Gaps and findings (recorded, not fixed)
+
+- **h31 and h32 (walk):** HowlFrame has no directory listing or walking construct,
+  so neither task can be expressed. h32 would also need string ordering:
+  `<` rejects strings with `TYPE_ERROR`, and there is no sort.
+- **Sugar inside some call arguments:**
+  - Inside `write_file` or `to_lower` arguments, `str_replace`/`read_text` is not
+    expanded. The governed build fails with "no lowering for construct".
+  - Their parents are missing from `isSugarExprContext`.
+  - Workaround used in h05 and h12: bind the value first.
+- **`or` evaluates both operands:** h07's `(or (= i 0) (>= (to_int …) 30))` raised a
+  `CONVERSION_ERROR` on the header row. Rewritten with `if`.
+- **Value semantics:** `append` on a list taken out of a map does not update the
+  map, so h28 has to write the list back with `map_set`.
+
+## Threats to validity
+
+- **One author:** the same author wrote the specs, the oracle and all 122 references.
+  The specs were frozen first, but that author has seen HowlFrame's
+  weaknesses, and the task choice may lean towards Python idioms (CSV,
+  `Counter`, slicing).
+- **Fixtures:** small, valid inputs only. CSV has no quoting; text has no Unicode edge cases.
+- **Proxy only:** these are tokenizer proxies, not model generation cost or correctness.
+- **Isolation:** Python, JS and Go run as host processes, with the same lack of host
+  isolation as v0. The verifier only checks the sandbox file set afterwards.
+  HowlFrame runs under scoped VM filesystem grants.
+- **Small samples:** each category has n = 5, so category means are noisy.
+
+## Repository validation status
+
+All of these were run locally on the box, outside a sandbox, with `GOCACHE=/tmp/hf-token-go-cache`:
+
+- `go build ./... && go vet ./... && go test -count=1 ./...`: exit 0.
+- `go test -count=1 ./experiments/...`: passes. This includes `TestHeldoutVerificationCoverage` and `TestHeldoutSpecsFrozen`.
+- `experiments/token-efficiency-heldout-v0/run.sh`: exit 0, with 122 of 122 verified references passing.
+- `experiments/token-efficiency-v0/run.sh`: exit 0. v0 is unchanged.
+
+See the [journal](../../docs/journals/2026-10-12_heldout_token_benchmark.md).
diff --git a/experiments/token-efficiency-heldout-v0/cmd/heldoutverify/main.go b/experiments/token-efficiency-heldout-v0/cmd/heldoutverify/main.go
new file mode 100644
index 0000000..ce5535c
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/cmd/heldoutverify/main.go
@@ -0,0 +1,213 @@
+// Command heldoutverify replays every held-out reference in a fresh sandbox
+// and compares its single output with the frozen expected value.
+package main
+
+import (
+	"bytes"
+	"context"
+	"crypto/sha256"
+	"encoding/json"
+	"flag"
+	"fmt"
+	"io/fs"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"reflect"
+	"sort"
+	"strings"
+	"time"
+)
+
+type Task struct {
+	ID       string            `json:"id"`
+	Category string            `json:"category"`
+	Inputs   map[string]string `json:"inputs"`
+	Output   string            `json:"output"`
+	Kind     string            `json:"kind"`
+	Expected json.RawMessage   `json:"expected"`
+}
+
+type Gap struct {
+	Task   string `json:"task"`
+	Lang   string `json:"lang"`
+	Reason string `json:"reason"`
+}
+
+type Result struct {
+	Task   string `json:"task"`
+	Lang   string `json:"lang"`
+	Pass   bool   `json:"pass"`
+	SHA256 string `json:"sha256"`
+	Detail string `json:"detail,omitempty"`
+}
+
+var exts = map[string]string{"howl": "howl", "python": "py", "js": "js", "go": "go"}
+
+func main() {
+	root := flag.String("root", "experiments/token-efficiency-heldout-v0", "benchmark root")
+	howl := flag.String("howlframe", "/tmp/howlframe", "HowlFrame binary")
+	flag.Parse()
+	var manifest struct{ Tasks []Task }
+	must(readJSON(filepath.Join(*root, "tasks/manifest.json"), &manifest))
+	var gaps []Gap
+	must(readJSON(filepath.Join(*root, "gaps.json"), &gaps))
+	gapped := map[string]bool{}
+	for _, g := range gaps {
+		gapped[g.Task+"/"+g.Lang] = true
+	}
+	work, err := os.MkdirTemp("", "hf-heldout-verify-")
+	must(err)
+	fmt.Println("Replay artifacts:", work)
+	var results []Result
+	failed := false
+	for _, lang := range []string{"howl", "python", "js", "go"} {
+		passed, total := 0, 0
+		for _, t := range manifest.Tasks {
+			if gapped[t.ID+"/"+lang] {
+				continue
+			}
+			total++
+			src, _ := filepath.Abs(filepath.Join(*root, "solutions", lang, t.ID+"."+exts[lang]))
+			data, readErr := os.ReadFile(src)
+			r := Result{Task: t.ID, Lang: lang, SHA256: fmt.Sprintf("%x", sha256.Sum256(data))}
+			if readErr != nil {
+				r.Detail = readErr.Error()
+			} else {
+				r.Detail = verify(filepath.Join(work, lang+"-"+t.ID), src, lang, *howl, t)
+				r.Pass = r.Detail == ""
+			}
+			if r.Pass {
+				passed++
+			} else {
+				failed = true
+			}
+			fmt.Printf("%s %s pass=%v %s\n", lang, t.ID, r.Pass, r.Detail)
+			results = append(results, r)
+		}
+		fmt.Printf("%s: %d/%d passed\n", lang, passed, total)
+	}
+	out, _ := json.MarshalIndent(results, "", "  ")
+	must(os.MkdirAll(filepath.Join(*root, "results"), 0o755))
+	must(os.WriteFile(filepath.Join(*root, "results/verification.json"), append(out, '\n'), 0o644))
+	if failed {
+		os.Exit(1)
+	}
+}
+
+func verify(dir, src, lang, howl string, t Task) string {
+	if err := os.MkdirAll(filepath.Join(dir, "out"), 0o755); err != nil {
+		return err.Error()
+	}
+	for p, content := range t.Inputs {
+		full := filepath.Join(dir, p)
+		os.MkdirAll(filepath.Dir(full), 0o755)
+		if err := os.WriteFile(full, []byte(content), 0o644); err != nil {
+			return err.Error()
+		}
+	}
+	absDir, _ := filepath.Abs(dir)
+	var cmd []string
+	switch lang {
+	case "howl":
+		art := filepath.Join(filepath.Dir(absDir), filepath.Base(absDir)+".hfbc")
+		if o, err := run(absDir, 30*time.Second, howl, "build", "--profile", "governed", src, "-o", art); err != nil {
+			return "build: " + o
+		}
+		cmd = []string{howl, "-run-bc", "-allow-caps", "filesystem:read=" + filepath.Join(absDir, "in") + ",filesystem:write=" + filepath.Join(absDir, "out"), art}
+	case "python":
+		cmd = []string{"python3", src}
+	case "js":
+		cmd = []string{"node", src}
+	case "go":
+		bin := absDir + ".bin"
+		if o, err := run(filepath.Dir(absDir), 60*time.Second, "go", "build", "-o", bin, src); err != nil {
+			return "go build: " + o
+		}
+		cmd = []string{bin}
+	}
+	if o, err := run(absDir, 10*time.Second, cmd[0], cmd[1:]...); err != nil {
+		return "run: " + o
+	}
+	// Exactly the inputs (unchanged) plus the one output may exist.
+	var files []string
+	filepath.WalkDir(absDir, func(p string, d fs.DirEntry, err error) error {
+		if err == nil && !d.IsDir() {
+			rel, _ := filepath.Rel(absDir, p)
+			files = append(files, filepath.ToSlash(rel))
+		}
+		return nil
+	})
+	want := []string{t.Output}
+	for p := range t.Inputs {
+		want = append(want, p)
+	}
+	sort.Strings(files)
+	sort.Strings(want)
+	if !reflect.DeepEqual(files, want) {
+		return fmt.Sprintf("file set %v, want %v", files, want)
+	}
+	for p, content := range t.Inputs {
+		got, _ := os.ReadFile(filepath.Join(absDir, p))
+		if string(got) != content {
+			return "input modified: " + p
+		}
+	}
+	got, err := os.ReadFile(filepath.Join(absDir, t.Output))
+	if err != nil {
+		return err.Error()
+	}
+	return compare(got, t)
+}
+
+func compare(got []byte, t Task) string {
+	if t.Kind == "text" {
+		var want string
+		json.Unmarshal(t.Expected, &want)
+		g := strings.TrimRight(string(got), "\n")
+		if g != strings.TrimRight(want, "\n") {
+			return fmt.Sprintf("text mismatch: got %q want %q", g, want)
+		}
+		return ""
+	}
+	var g, w any
+	if err := json.Unmarshal(got, &g); err != nil {
+		return "invalid JSON output: " + err.Error()
+	}
+	json.Unmarshal(t.Expected, &w)
+	if !reflect.DeepEqual(g, w) {
+		return fmt.Sprintf("json mismatch: got %s want %s", bytes.TrimSpace(got), t.Expected)
+	}
+	return ""
+}
+
+func run(dir string, timeout time.Duration, name string, args ...string) (string, error) {
+	ctx, cancel := context.WithTimeout(context.Background(), timeout)
+	defer cancel()
+	c := exec.CommandContext(ctx, name, args...)
+	c.Dir = dir
+	out, err := c.CombinedOutput()
+	if ctx.Err() != nil {
+		return "timeout", ctx.Err()
+	}
+	s := string(out)
+	if len(s) > 400 {
+		s = s[:400]
+	}
+	return s, err
+}
+
+func readJSON(path string, v any) error {
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return err
+	}
+	return json.Unmarshal(data, v)
+}
+
+func must(err error) {
+	if err != nil {
+		fmt.Fprintln(os.Stderr, err)
+		os.Exit(2)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/count_tokens.py b/experiments/token-efficiency-heldout-v0/count_tokens.py
new file mode 100644
index 0000000..44bda83
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/count_tokens.py
@@ -0,0 +1,82 @@
+"""Count verified held-out sources with the v0 tokenizers and stripping rules."""
+import csv, hashlib, importlib.util, json, statistics
+from pathlib import Path
+
+import tiktoken
+import tiktoken.load
+
+ROOT = Path(__file__).resolve().parent
+_spec = importlib.util.spec_from_file_location("v0count", ROOT.parent / "token-efficiency-v0/count_tokens.py")
+v0 = importlib.util.module_from_spec(_spec)
+_spec.loader.exec_module(v0)  # reuse strip_comments and offline guard unchanged
+LANGUAGES = v0.LANGUAGES
+
+
+def main():
+    if tiktoken.__version__ != "0.8.0":
+        raise RuntimeError("Install requirements.txt: tiktoken==0.8.0 required")
+    tiktoken.load.read_file = v0.offline_read
+    encodings = {n: tiktoken.get_encoding(n) for n in ("cl100k_base", "o200k_base")}
+    tasks = json.loads((ROOT / "tasks/manifest.json").read_text())["tasks"]
+    cats = {t["id"]: t["category"] for t in tasks}
+    gaps = {(g["task"], g["lang"]) for g in json.loads((ROOT / "gaps.json").read_text())}
+    checks = {(r["task"], r["lang"]): r for r in json.loads((ROOT / "results/verification.json").read_text())}
+    expected = {(t["id"], l) for t in tasks for l in LANGUAGES} - gaps
+    if set(checks) != expected:
+        raise RuntimeError("verification coverage must be every task x language minus gaps.json")
+    rows = []
+    for t in tasks:
+        for lang, ext in LANGUAGES.items():
+            if (t["id"], lang) in gaps:
+                continue
+            raw = (ROOT / f"solutions/{lang}/{t['id']}.{ext}").read_bytes()
+            c = checks[t["id"], lang]
+            if hashlib.sha256(raw).hexdigest() != c["sha256"] or not c["pass"]:
+                raise RuntimeError(f"unverified or changed source {t['id']}/{lang}")
+            src = raw.decode()
+            stripped = v0.strip_comments(src, lang)
+            for name, enc in encodings.items():
+                rows.append({"task": t["id"], "category": t["category"], "lang": lang, "encoding": name,
+                             "tokens": len(enc.encode(src, disallowed_special=())),
+                             "tokens_nocomments": len(enc.encode(stripped, disallowed_special=())),
+                             "bytes": len(raw), "pass": c["pass"]})
+    (ROOT / "results/tokens.json").write_text(json.dumps(rows, indent=2) + "\n")
+    with (ROOT / "results/tokens.csv").open("w", newline="") as f:
+        w = csv.DictWriter(f, fieldnames=list(rows[0]))
+        w.writeheader()
+        w.writerows(rows)
+    summary = {}
+    for enc in encodings:
+        summary[enc] = {}
+        for metric in ("tokens", "tokens_nocomments"):
+            by = {l: {r["task"]: r[metric] for r in rows if r["encoding"] == enc and r["lang"] == l} for l in LANGUAGES}
+            solved = sorted(by["howl"])
+            s = {"languages": {l: {"median_all": statistics.median(v.values()),
+                                   "median_howl_solved": statistics.median(v[t] for t in solved),
+                                   "mean_howl_solved": statistics.mean(v[t] for t in solved)} for l, v in by.items()},
+                 "howl_vs": {}, "categories": {}}
+            for l in ("python", "js", "go"):
+                ratios = [by["howl"][t] / by[l][t] for t in solved]
+                s["howl_vs"][l] = {
+                    "median_per_task_ratio": statistics.median(ratios),
+                    "ratio_of_medians": statistics.median(by["howl"].values()) / statistics.median(by[l][t] for t in solved),
+                    "wins": sum(by["howl"][t] < by[l][t] for t in solved),
+                    "ties": sum(by["howl"][t] == by[l][t] for t in solved),
+                    "losses": sum(by["howl"][t] > by[l][t] for t in solved),
+                    "gaps": len(by[l]) - len(solved),
+                }
+            for cat in sorted(set(cats.values())):
+                ids = [t for t in by["python"] if cats[t] == cat]
+                hs = [t for t in ids if t in by["howl"]]
+                s["categories"][cat] = {
+                    "n": len(ids), "howl_solved": len(hs),
+                    **{f"mean_{l}": (statistics.mean(by[l][t] for t in hs) if hs else None) for l in LANGUAGES},
+                    "median_ratio_vs_python": statistics.median(by["howl"][t] / by["python"][t] for t in hs) if hs else None,
+                }
+            summary[enc][metric] = s
+    (ROOT / "results/summary.json").write_text(json.dumps(summary, indent=2) + "\n")
+    print(json.dumps({e: summary[e]["tokens"]["howl_vs"] for e in summary}, indent=1))
+
+
+if __name__ == "__main__":
+    main()
diff --git a/experiments/token-efficiency-heldout-v0/gaps.json b/experiments/token-efficiency-heldout-v0/gaps.json
new file mode 100644
index 0000000..7eaaeb8
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/gaps.json
@@ -0,0 +1,4 @@
+[
+  {"task": "h31", "lang": "howl", "reason": "no directory listing or walking construct on main (only read_file/write_file/mkdir on fixed paths)"},
+  {"task": "h32", "lang": "howl", "reason": "no directory listing or walking construct on main; also no string ordering comparison for sorting"}
+]
diff --git a/experiments/token-efficiency-heldout-v0/requirements.txt b/experiments/token-efficiency-heldout-v0/requirements.txt
new file mode 100644
index 0000000..135d4f5
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/requirements.txt
@@ -0,0 +1 @@
+tiktoken==0.8.0
diff --git a/experiments/token-efficiency-heldout-v0/results/summary.json b/experiments/token-efficiency-heldout-v0/results/summary.json
new file mode 100644
index 0000000..546376f
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/results/summary.json
@@ -0,0 +1,466 @@
+{
+  "cl100k_base": {
+    "tokens": {
+      "languages": {
+        "howl": {
+          "median_all": 72.5,
+          "median_howl_solved": 72.5,
+          "mean_howl_solved": 85.46666666666667
+        },
+        "python": {
+          "median_all": 62.5,
+          "median_howl_solved": 63.0,
+          "mean_howl_solved": 60.833333333333336
+        },
+        "js": {
+          "median_all": 69.0,
+          "median_howl_solved": 69.0,
+          "mean_howl_solved": 71.6
+        },
+        "go": {
+          "median_all": 162.5,
+          "median_howl_solved": 162.5,
+          "mean_howl_solved": 162
+        }
+      },
+      "howl_vs": {
+        "python": {
+          "median_per_task_ratio": 1.3092086736130666,
+          "ratio_of_medians": 1.1507936507936507,
+          "wins": 0,
+          "ties": 1,
+          "losses": 29,
+          "gaps": 2
+        },
+        "js": {
+          "median_per_task_ratio": 1.087069525666017,
+          "ratio_of_medians": 1.0507246376811594,
+          "wins": 6,
+          "ties": 1,
+          "losses": 23,
+          "gaps": 2
+        },
+        "go": {
+          "median_per_task_ratio": 0.4993893362848746,
+          "ratio_of_medians": 0.4461538461538462,
+          "wins": 30,
+          "ties": 0,
+          "losses": 0,
+          "gaps": 2
+        }
+      },
+      "categories": {
+        "aggregate": {
+          "n": 5,
+          "howl_solved": 5,
+          "mean_howl": 81.8,
+          "mean_python": 53.6,
+          "mean_js": 71.8,
+          "mean_go": 152.4,
+          "median_ratio_vs_python": 1.4761904761904763
+        },
+        "csv": {
+          "n": 5,
+          "howl_solved": 5,
+          "mean_howl": 135,
+          "mean_python": 73,
+          "mean_js": 89,
+          "mean_go": 185.6,
+          "median_ratio_vs_python": 1.8428571428571427
+        },
+        "parse": {
+          "n": 5,
+          "howl_solved": 5,
+          "mean_howl": 81.2,
+          "mean_python": 57.8,
+          "mean_js": 72,
+          "mean_go": 152.8,
+          "median_ratio_vs_python": 1.3968253968253967
+        },
+        "reshape": {
+          "n": 5,
+          "howl_solved": 5,
+          "mean_howl": 72.4,
+          "mean_python": 61.8,
+          "mean_js": 65.6,
+          "mean_go": 174.2,
+          "median_ratio_vs_python": 1.0793650793650793
+        },
+        "text": {
+          "n": 5,
+          "howl_solved": 5,
+          "mean_howl": 58.4,
+          "mean_python": 47,
+          "mean_js": 50.8,
+          "mean_go": 119.2,
+          "median_ratio_vs_python": 1.2272727272727273
+        },
+        "validate": {
+          "n": 5,
+          "howl_solved": 5,
+          "mean_howl": 84,
+          "mean_python": 71.8,
+          "mean_js": 80.4,
+          "mean_go": 187.8,
+          "median_ratio_vs_python": 1.1111111111111112
+        },
+        "walk": {
+          "n": 2,
+          "howl_solved": 0,
+          "mean_howl": null,
+          "mean_python": null,
+          "mean_js": null,
+          "mean_go": null,
+          "median_ratio_vs_python": null
+        }
+      }
+    },
+    "tokens_nocomments": {
+      "languages": {
+        "howl": {
+          "median_all": 72.5,
+          "median_howl_solved": 72.5,
+          "mean_howl_solved": 85.46666666666667
+        },
+        "python": {
+          "median_all": 62.5,
+          "median_howl_solved": 63.0,
+          "mean_howl_solved": 60.833333333333336
+        },
+        "js": {
+          "median_all": 69.0,
+          "median_howl_solved": 69.0,
+          "mean_howl_solved": 71.6
+        },
+        "go": {
+          "median_all": 156.5,
+          "median_howl_solved": 156.5,
+          "mean_howl_solved": 156
+        }
+      },
+      "howl_vs": {
+        "python": {
+          "median_per_task_ratio": 1.3092086736130666,
+          "ratio_of_medians": 1.1507936507936507,
+          "wins": 0,
+          "ties": 1,
+          "losses": 29,
+          "gaps": 2
+        },
+        "js": {
+          "median_per_task_ratio": 1.087069525666017,
+          "ratio_of_medians": 1.0507246376811594,
+          "wins": 6,
+          "ties": 1,
+          "losses": 23,
+          "gaps": 2
+        },
+        "go": {
+          "median_per_task_ratio": 0.5190048233499609,
+          "ratio_of_medians": 0.46325878594249204,
+          "wins": 30,
+          "ties": 0,
+          "losses": 0,
+          "gaps": 2
+        }
+      },
+      "categories": {
+        "aggregate": {
+          "n": 5,
+          "howl_solved": 5,
+          "mean_howl": 81.8,
+          "mean_python": 53.6,
+          "mean_js": 71.8,
+          "mean_go": 146.4,
+          "median_ratio_vs_python": 1.4761904761904763
+        },
+        "csv": {
+          "n": 5,
+          "howl_solved": 5,
+          "mean_howl": 135,
+          "mean_python": 73,
+          "mean_js": 89,
+          "mean_go": 179.6,
+          "median_ratio_vs_python": 1.8428571428571427
+        },
+        "parse": {
+          "n": 5,
+          "howl_solved": 5,
+          "mean_howl": 81.2,
+          "mean_python": 57.8,
+          "mean_js": 72,
+          "mean_go": 146.8,
+          "median_ratio_vs_python": 1.3968253968253967
+        },
+        "reshape": {
+          "n": 5,
+          "howl_solved": 5,
+          "mean_howl": 72.4,
+          "mean_python": 61.8,
+          "mean_js": 65.6,
+          "mean_go": 168.2,
+          "median_ratio_vs_python": 1.0793650793650793
+        },
+        "text": {
+          "n": 5,
+          "howl_solved": 5,
+          "mean_howl": 58.4,
+          "mean_python": 47,
+          "mean_js": 50.8,
+          "mean_go": 113.2,
+          "median_ratio_vs_python": 1.2272727272727273
+        },
+        "validate": {
+          "n": 5,
+          "howl_solved": 5,
+          "mean_howl": 84,
+          "mean_python": 71.8,
+          "mean_js": 80.4,
+          "mean_go": 181.8,
+          "median_ratio_vs_python": 1.1111111111111112
+        },
+        "walk": {
+          "n": 2,
+          "howl_solved": 0,
+          "mean_howl": null,
+          "mean_python": null,
+          "mean_js": null,
+          "mean_go": null,
+          "median_ratio_vs_python": null
+        }
+      }
+    }
+  },
+  "o200k_base": {
+    "tokens": {
+      "languages": {
+        "howl": {
+          "median_all": 72.5,
+          "median_howl_solved": 72.5,
+          "mean_howl_solved": 85.56666666666666
+        },
+        "python": {
+          "median_all": 62.5,
+          "median_howl_solved": 63.0,
+          "mean_howl_solved": 61.13333333333333
+        },
+        "js": {
+          "median_all": 73.0,
+          "median_howl_solved": 73.0,
+          "mean_howl_solved": 76.26666666666667
+        },
+        "go": {
+          "median_all": 164.5,
+          "median_howl_solved": 164.5,
+          "mean_howl_solved": 163.56666666666666
+        }
+      },
+      "howl_vs": {
+        "python": {
+          "median_per_task_ratio": 1.2985678704856787,
+          "ratio_of_medians": 1.1507936507936507,
+          "wins": 0,
+          "ties": 1,
+          "losses": 29,
+          "gaps": 2
+        },
+        "js": {
+          "median_per_task_ratio": 1.015889029003783,
+          "ratio_of_medians": 0.9931506849315068,
+          "wins": 13,
+          "ties": 0,
+          "losses": 17,
+          "gaps": 2
+        },
+        "go": {
+          "median_per_task_ratio": 0.492850707094449,
+          "ratio_of_medians": 0.44072948328267475,
+          "wins": 30,
+          "ties": 0,
+          "losses": 0,
+          "gaps": 2
+        }
+      },
+      "categories": {
+        "aggregate": {
+          "n": 5,
+          "howl_solved": 5,
+          "mean_howl": 81.8,
+          "mean_python": 53.6,
+          "mean_js": 76.4,
+          "mean_go": 153.6,
+          "median_ratio_vs_python": 1.4761904761904763
+        },
+        "csv": {
+          "n": 5,
+          "howl_solved": 5,
+          "mean_howl": 135,
+          "mean_python": 73.8,
+          "mean_js": 94.4,
+          "mean_go": 187.4,
+          "median_ratio_vs_python": 1.8169014084507042
+        },
+        "parse": {
+          "n": 5,
+          "howl_solved": 5,
+          "mean_howl": 81.4,
+          "mean_python": 58,
+          "mean_js": 76.6,
+          "mean_go": 154.2,
+          "median_ratio_vs_python": 1.3968253968253967
+        },
+        "reshape": {
+          "n": 5,
+          "howl_solved": 5,
+          "mean_howl": 72.4,
+          "mean_python": 62,
+          "mean_js": 69.6,
+          "mean_go": 175.2,
+          "median_ratio_vs_python": 1.0793650793650793
+        },
+        "text": {
+          "n": 5,
+          "howl_solved": 5,
+          "mean_howl": 58.4,
+          "mean_python": 47.4,
+          "mean_js": 55.4,
+          "mean_go": 121.2,
+          "median_ratio_vs_python": 1.2
+        },
+        "validate": {
+          "n": 5,
+          "howl_solved": 5,
+          "mean_howl": 84.4,
+          "mean_python": 72,
+          "mean_js": 85.2,
+          "mean_go": 189.8,
+          "median_ratio_vs_python": 1.1111111111111112
+        },
+        "walk": {
+          "n": 2,
+          "howl_solved": 0,
+          "mean_howl": null,
+          "mean_python": null,
+          "mean_js": null,
+          "mean_go": null,
+          "median_ratio_vs_python": null
+        }
+      }
+    },
+    "tokens_nocomments": {
+      "languages": {
+        "howl": {
+          "median_all": 72.5,
+          "median_howl_solved": 72.5,
+          "mean_howl_solved": 85.56666666666666
+        },
+        "python": {
+          "median_all": 62.5,
+          "median_howl_solved": 63.0,
+          "mean_howl_solved": 61.13333333333333
+        },
+        "js": {
+          "median_all": 73.0,
+          "median_howl_solved": 73.0,
+          "mean_howl_solved": 76.26666666666667
+        },
+        "go": {
+          "median_all": 158.5,
+          "median_howl_solved": 158.5,
+          "mean_howl_solved": 157.56666666666666
+        }
+      },
+      "howl_vs": {
+        "python": {
+          "median_per_task_ratio": 1.2985678704856787,
+          "ratio_of_medians": 1.1507936507936507,
+          "wins": 0,
+          "ties": 1,
+          "losses": 29,
+          "gaps": 2
+        },
+        "js": {
+          "median_per_task_ratio": 1.015889029003783,
+          "ratio_of_medians": 0.9931506849315068,
+          "wins": 13,
+          "ties": 0,
+          "losses": 17,
+          "gaps": 2
+        },
+        "go": {
+          "median_per_task_ratio": 0.5126086035176944,
+          "ratio_of_medians": 0.45741324921135645,
+          "wins": 30,
+          "ties": 0,
+          "losses": 0,
+          "gaps": 2
+        }
+      },
+      "categories": {
+        "aggregate": {
+          "n": 5,
+          "howl_solved": 5,
+          "mean_howl": 81.8,
+          "mean_python": 53.6,
+          "mean_js": 76.4,
+          "mean_go": 147.6,
+          "median_ratio_vs_python": 1.4761904761904763
+        },
+        "csv": {
+          "n": 5,
+          "howl_solved": 5,
+          "mean_howl": 135,
+          "mean_python": 73.8,
+          "mean_js": 94.4,
+          "mean_go": 181.4,
+          "median_ratio_vs_python": 1.8169014084507042
+        },
+        "parse": {
+          "n": 5,
+          "howl_solved": 5,
+          "mean_howl": 81.4,
+          "mean_python": 58,
+          "mean_js": 76.6,
+          "mean_go": 148.2,
+          "median_ratio_vs_python": 1.3968253968253967
+        },
+        "reshape": {
+          "n": 5,
+          "howl_solved": 5,
+          "mean_howl": 72.4,
+          "mean_python": 62,
+          "mean_js": 69.6,
+          "mean_go": 169.2,
+          "median_ratio_vs_python": 1.0793650793650793
+        },
+        "text": {
+          "n": 5,
+          "howl_solved": 5,
+          "mean_howl": 58.4,
+          "mean_python": 47.4,
+          "mean_js": 55.4,
+          "mean_go": 115.2,
+          "median_ratio_vs_python": 1.2
+        },
+        "validate": {
+          "n": 5,
+          "howl_solved": 5,
+          "mean_howl": 84.4,
+          "mean_python": 72,
+          "mean_js": 85.2,
+          "mean_go": 183.8,
+          "median_ratio_vs_python": 1.1111111111111112
+        },
+        "walk": {
+          "n": 2,
+          "howl_solved": 0,
+          "mean_howl": null,
+          "mean_python": null,
+          "mean_js": null,
+          "mean_go": null,
+          "median_ratio_vs_python": null
+        }
+      }
+    }
+  }
+}
diff --git a/experiments/token-efficiency-heldout-v0/results/tokens.csv b/experiments/token-efficiency-heldout-v0/results/tokens.csv
new file mode 100644
index 0000000..93f057c
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/results/tokens.csv
@@ -0,0 +1,253 @@
+task,category,lang,encoding,tokens,tokens_nocomments,bytes,pass
+h01,text,howl,cl100k_base,54,54,168,True
+h01,text,howl,o200k_base,54,54,168,True
+h01,text,python,cl100k_base,44,44,162,True
+h01,text,python,o200k_base,44,44,162,True
+h01,text,js,cl100k_base,49,49,188,True
+h01,text,js,o200k_base,55,55,188,True
+h01,text,go,cl100k_base,112,106,383,True
+h01,text,go,o200k_base,114,108,383,True
+h02,text,howl,cl100k_base,70,70,208,True
+h02,text,howl,o200k_base,70,70,208,True
+h02,text,python,cl100k_base,40,40,147,True
+h02,text,python,o200k_base,40,40,147,True
+h02,text,js,cl100k_base,43,43,164,True
+h02,text,js,o200k_base,47,47,164,True
+h02,text,go,cl100k_base,138,132,415,True
+h02,text,go,o200k_base,139,133,415,True
+h03,text,howl,cl100k_base,66,66,187,True
+h03,text,howl,o200k_base,66,66,187,True
+h03,text,python,cl100k_base,62,62,209,True
+h03,text,python,o200k_base,62,62,209,True
+h03,text,js,cl100k_base,61,61,219,True
+h03,text,js,o200k_base,65,65,219,True
+h03,text,go,cl100k_base,125,119,401,True
+h03,text,go,o200k_base,126,120,401,True
+h04,text,howl,cl100k_base,66,66,199,True
+h04,text,howl,o200k_base,66,66,199,True
+h04,text,python,cl100k_base,53,53,192,True
+h04,text,python,o200k_base,55,55,192,True
+h04,text,js,cl100k_base,61,61,230,True
+h04,text,js,o200k_base,65,65,230,True
+h04,text,go,cl100k_base,133,127,459,True
+h04,text,go,o200k_base,138,132,459,True
+h05,text,howl,cl100k_base,36,36,116,True
+h05,text,howl,o200k_base,36,36,116,True
+h05,text,python,cl100k_base,36,36,134,True
+h05,text,python,o200k_base,36,36,134,True
+h05,text,js,cl100k_base,40,40,154,True
+h05,text,js,o200k_base,45,45,154,True
+h05,text,go,cl100k_base,88,82,292,True
+h05,text,go,o200k_base,89,83,292,True
+h06,csv,howl,cl100k_base,122,122,349,True
+h06,csv,howl,o200k_base,122,122,349,True
+h06,csv,python,cl100k_base,71,71,247,True
+h06,csv,python,o200k_base,72,72,247,True
+h06,csv,js,cl100k_base,74,74,270,True
+h06,csv,js,o200k_base,79,79,270,True
+h06,csv,go,cl100k_base,165,159,534,True
+h06,csv,go,o200k_base,166,160,534,True
+h07,csv,howl,cl100k_base,144,144,404,True
+h07,csv,howl,o200k_base,144,144,404,True
+h07,csv,python,cl100k_base,82,82,272,True
+h07,csv,python,o200k_base,84,84,272,True
+h07,csv,js,cl100k_base,89,89,302,True
+h07,csv,js,o200k_base,96,96,302,True
+h07,csv,go,cl100k_base,206,200,634,True
+h07,csv,go,o200k_base,211,205,634,True
+h08,csv,howl,cl100k_base,127,127,367,True
+h08,csv,howl,o200k_base,127,127,367,True
+h08,csv,python,cl100k_base,64,64,227,True
+h08,csv,python,o200k_base,64,64,227,True
+h08,csv,js,cl100k_base,81,81,289,True
+h08,csv,js,o200k_base,87,87,289,True
+h08,csv,go,cl100k_base,186,180,597,True
+h08,csv,go,o200k_base,187,181,597,True
+h09,csv,howl,cl100k_base,129,129,395,True
+h09,csv,howl,o200k_base,129,129,395,True
+h09,csv,python,cl100k_base,70,70,253,True
+h09,csv,python,o200k_base,71,71,253,True
+h09,csv,js,cl100k_base,96,96,357,True
+h09,csv,js,o200k_base,100,100,357,True
+h09,csv,go,cl100k_base,185,179,638,True
+h09,csv,go,o200k_base,186,180,638,True
+h10,csv,howl,cl100k_base,153,153,447,True
+h10,csv,howl,o200k_base,153,153,447,True
+h10,csv,python,cl100k_base,78,78,293,True
+h10,csv,python,o200k_base,78,78,293,True
+h10,csv,js,cl100k_base,105,105,354,True
+h10,csv,js,o200k_base,110,110,354,True
+h10,csv,go,cl100k_base,186,180,604,True
+h10,csv,go,o200k_base,187,181,604,True
+h11,aggregate,howl,cl100k_base,137,137,375,True
+h11,aggregate,howl,o200k_base,137,137,375,True
+h11,aggregate,python,cl100k_base,71,71,264,True
+h11,aggregate,python,o200k_base,71,71,264,True
+h11,aggregate,js,cl100k_base,96,96,330,True
+h11,aggregate,js,o200k_base,101,101,330,True
+h11,aggregate,go,cl100k_base,224,218,717,True
+h11,aggregate,go,o200k_base,225,219,717,True
+h12,aggregate,howl,cl100k_base,92,92,283,True
+h12,aggregate,howl,o200k_base,92,92,283,True
+h12,aggregate,python,cl100k_base,67,67,270,True
+h12,aggregate,python,o200k_base,67,67,270,True
+h12,aggregate,js,cl100k_base,91,91,320,True
+h12,aggregate,js,o200k_base,97,97,320,True
+h12,aggregate,go,cl100k_base,154,148,530,True
+h12,aggregate,go,o200k_base,156,150,530,True
+h13,aggregate,howl,cl100k_base,62,62,206,True
+h13,aggregate,howl,o200k_base,62,62,206,True
+h13,aggregate,python,cl100k_base,42,42,163,True
+h13,aggregate,python,o200k_base,42,42,163,True
+h13,aggregate,js,cl100k_base,60,60,224,True
+h13,aggregate,js,o200k_base,64,64,224,True
+h13,aggregate,go,cl100k_base,133,127,440,True
+h13,aggregate,go,o200k_base,134,128,440,True
+h14,aggregate,howl,cl100k_base,56,56,168,True
+h14,aggregate,howl,o200k_base,56,56,168,True
+h14,aggregate,python,cl100k_base,50,50,177,True
+h14,aggregate,python,o200k_base,50,50,177,True
+h14,aggregate,js,cl100k_base,53,53,211,True
+h14,aggregate,js,o200k_base,57,57,211,True
+h14,aggregate,go,cl100k_base,124,118,416,True
+h14,aggregate,go,o200k_base,125,119,416,True
+h15,aggregate,howl,cl100k_base,62,62,193,True
+h15,aggregate,howl,o200k_base,62,62,193,True
+h15,aggregate,python,cl100k_base,38,38,140,True
+h15,aggregate,python,o200k_base,38,38,140,True
+h15,aggregate,js,cl100k_base,59,59,234,True
+h15,aggregate,js,o200k_base,63,63,234,True
+h15,aggregate,go,cl100k_base,127,121,409,True
+h15,aggregate,go,o200k_base,128,122,409,True
+h16,parse,howl,cl100k_base,109,109,336,True
+h16,parse,howl,o200k_base,109,109,336,True
+h16,parse,python,cl100k_base,59,59,215,True
+h16,parse,python,o200k_base,59,59,215,True
+h16,parse,js,cl100k_base,84,84,299,True
+h16,parse,js,o200k_base,89,89,299,True
+h16,parse,go,cl100k_base,152,146,524,True
+h16,parse,go,o200k_base,153,147,524,True
+h17,parse,howl,cl100k_base,88,88,285,True
+h17,parse,howl,o200k_base,88,88,285,True
+h17,parse,python,cl100k_base,63,63,251,True
+h17,parse,python,o200k_base,63,63,251,True
+h17,parse,js,cl100k_base,86,86,301,True
+h17,parse,js,o200k_base,90,90,301,True
+h17,parse,go,cl100k_base,145,139,498,True
+h17,parse,go,o200k_base,146,140,498,True
+h18,parse,howl,cl100k_base,62,62,188,True
+h18,parse,howl,o200k_base,62,62,188,True
+h18,parse,python,cl100k_base,54,54,194,True
+h18,parse,python,o200k_base,54,54,194,True
+h18,parse,js,cl100k_base,57,57,219,True
+h18,parse,js,o200k_base,61,61,219,True
+h18,parse,go,cl100k_base,126,120,433,True
+h18,parse,go,o200k_base,127,121,433,True
+h19,parse,howl,cl100k_base,88,88,259,True
+h19,parse,howl,o200k_base,88,88,259,True
+h19,parse,python,cl100k_base,63,63,223,True
+h19,parse,python,o200k_base,63,63,223,True
+h19,parse,js,cl100k_base,81,81,279,True
+h19,parse,js,o200k_base,86,86,279,True
+h19,parse,go,cl100k_base,181,175,579,True
+h19,parse,go,o200k_base,182,176,579,True
+h20,parse,howl,cl100k_base,59,59,193,True
+h20,parse,howl,o200k_base,60,60,193,True
+h20,parse,python,cl100k_base,50,50,195,True
+h20,parse,python,o200k_base,51,51,195,True
+h20,parse,js,cl100k_base,52,52,203,True
+h20,parse,js,o200k_base,57,57,203,True
+h20,parse,go,cl100k_base,160,154,533,True
+h20,parse,go,o200k_base,163,157,533,True
+h21,validate,howl,cl100k_base,64,64,211,True
+h21,validate,howl,o200k_base,65,65,211,True
+h21,validate,python,cl100k_base,61,61,229,True
+h21,validate,python,o200k_base,61,61,229,True
+h21,validate,js,cl100k_base,68,68,256,True
+h21,validate,js,o200k_base,72,72,256,True
+h21,validate,go,cl100k_base,173,167,574,True
+h21,validate,go,o200k_base,174,168,574,True
+h22,validate,howl,cl100k_base,99,99,288,True
+h22,validate,howl,o200k_base,100,100,288,True
+h22,validate,python,cl100k_base,72,72,241,True
+h22,validate,python,o200k_base,73,73,241,True
+h22,validate,js,cl100k_base,73,73,254,True
+h22,validate,js,o200k_base,78,78,254,True
+h22,validate,go,cl100k_base,197,191,637,True
+h22,validate,go,o200k_base,201,195,637,True
+h23,validate,howl,cl100k_base,80,80,261,True
+h23,validate,howl,o200k_base,80,80,261,True
+h23,validate,python,cl100k_base,72,72,272,True
+h23,validate,python,o200k_base,72,72,272,True
+h23,validate,js,cl100k_base,83,83,296,True
+h23,validate,js,o200k_base,87,87,296,True
+h23,validate,go,cl100k_base,154,148,493,True
+h23,validate,go,o200k_base,156,150,493,True
+h24,validate,howl,cl100k_base,73,73,265,True
+h24,validate,howl,o200k_base,73,73,265,True
+h24,validate,python,cl100k_base,69,69,278,True
+h24,validate,python,o200k_base,69,69,278,True
+h24,validate,js,cl100k_base,74,74,303,True
+h24,validate,js,o200k_base,80,80,303,True
+h24,validate,go,cl100k_base,218,212,750,True
+h24,validate,go,o200k_base,220,214,750,True
+h25,validate,howl,cl100k_base,104,104,306,True
+h25,validate,howl,o200k_base,104,104,306,True
+h25,validate,python,cl100k_base,85,85,289,True
+h25,validate,python,o200k_base,85,85,289,True
+h25,validate,js,cl100k_base,104,104,332,True
+h25,validate,js,o200k_base,109,109,332,True
+h25,validate,go,cl100k_base,197,191,614,True
+h25,validate,go,o200k_base,198,192,614,True
+h26,reshape,howl,cl100k_base,63,63,205,True
+h26,reshape,howl,o200k_base,63,63,205,True
+h26,reshape,python,cl100k_base,59,59,216,True
+h26,reshape,python,o200k_base,59,59,216,True
+h26,reshape,js,cl100k_base,64,64,246,True
+h26,reshape,js,o200k_base,68,68,246,True
+h26,reshape,go,cl100k_base,175,169,579,True
+h26,reshape,go,o200k_base,176,170,579,True
+h27,reshape,howl,cl100k_base,68,68,241,True
+h27,reshape,howl,o200k_base,68,68,241,True
+h27,reshape,python,cl100k_base,63,63,237,True
+h27,reshape,python,o200k_base,63,63,237,True
+h27,reshape,js,cl100k_base,61,61,242,True
+h27,reshape,js,o200k_base,65,65,242,True
+h27,reshape,go,cl100k_base,181,175,627,True
+h27,reshape,go,o200k_base,182,176,627,True
+h28,reshape,howl,cl100k_base,97,97,331,True
+h28,reshape,howl,o200k_base,97,97,331,True
+h28,reshape,python,cl100k_base,64,64,252,True
+h28,reshape,python,o200k_base,64,64,252,True
+h28,reshape,js,cl100k_base,70,70,263,True
+h28,reshape,js,o200k_base,74,74,263,True
+h28,reshape,go,cl100k_base,183,177,618,True
+h28,reshape,go,o200k_base,184,178,618,True
+h29,reshape,howl,cl100k_base,72,72,234,True
+h29,reshape,howl,o200k_base,72,72,234,True
+h29,reshape,python,cl100k_base,65,65,231,True
+h29,reshape,python,o200k_base,66,66,231,True
+h29,reshape,js,cl100k_base,65,65,241,True
+h29,reshape,js,o200k_base,69,69,241,True
+h29,reshape,go,cl100k_base,177,171,580,True
+h29,reshape,go,o200k_base,178,172,580,True
+h30,reshape,howl,cl100k_base,62,62,197,True
+h30,reshape,howl,o200k_base,62,62,197,True
+h30,reshape,python,cl100k_base,58,58,222,True
+h30,reshape,python,o200k_base,58,58,222,True
+h30,reshape,js,cl100k_base,68,68,270,True
+h30,reshape,js,o200k_base,72,72,270,True
+h30,reshape,go,cl100k_base,155,149,536,True
+h30,reshape,go,o200k_base,156,150,536,True
+h31,walk,python,cl100k_base,41,41,146,True
+h31,walk,python,o200k_base,41,41,146,True
+h31,walk,js,cl100k_base,54,54,226,True
+h31,walk,js,o200k_base,58,58,226,True
+h31,walk,go,cl100k_base,141,135,444,True
+h31,walk,go,o200k_base,142,136,444,True
+h32,walk,python,cl100k_base,57,57,199,True
+h32,walk,python,o200k_base,57,57,199,True
+h32,walk,js,cl100k_base,109,109,429,True
+h32,walk,js,o200k_base,114,114,429,True
+h32,walk,go,cl100k_base,196,190,645,True
+h32,walk,go,o200k_base,198,192,645,True
diff --git a/experiments/token-efficiency-heldout-v0/results/tokens.json b/experiments/token-efficiency-heldout-v0/results/tokens.json
new file mode 100644
index 0000000..d6ffd02
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/results/tokens.json
@@ -0,0 +1,2522 @@
+[
+  {
+    "task": "h01",
+    "category": "text",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 54,
+    "tokens_nocomments": 54,
+    "bytes": 168,
+    "pass": true
+  },
+  {
+    "task": "h01",
+    "category": "text",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 54,
+    "tokens_nocomments": 54,
+    "bytes": 168,
+    "pass": true
+  },
+  {
+    "task": "h01",
+    "category": "text",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 44,
+    "tokens_nocomments": 44,
+    "bytes": 162,
+    "pass": true
+  },
+  {
+    "task": "h01",
+    "category": "text",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 44,
+    "tokens_nocomments": 44,
+    "bytes": 162,
+    "pass": true
+  },
+  {
+    "task": "h01",
+    "category": "text",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 49,
+    "tokens_nocomments": 49,
+    "bytes": 188,
+    "pass": true
+  },
+  {
+    "task": "h01",
+    "category": "text",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 55,
+    "tokens_nocomments": 55,
+    "bytes": 188,
+    "pass": true
+  },
+  {
+    "task": "h01",
+    "category": "text",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 112,
+    "tokens_nocomments": 106,
+    "bytes": 383,
+    "pass": true
+  },
+  {
+    "task": "h01",
+    "category": "text",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 114,
+    "tokens_nocomments": 108,
+    "bytes": 383,
+    "pass": true
+  },
+  {
+    "task": "h02",
+    "category": "text",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 70,
+    "tokens_nocomments": 70,
+    "bytes": 208,
+    "pass": true
+  },
+  {
+    "task": "h02",
+    "category": "text",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 70,
+    "tokens_nocomments": 70,
+    "bytes": 208,
+    "pass": true
+  },
+  {
+    "task": "h02",
+    "category": "text",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 40,
+    "tokens_nocomments": 40,
+    "bytes": 147,
+    "pass": true
+  },
+  {
+    "task": "h02",
+    "category": "text",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 40,
+    "tokens_nocomments": 40,
+    "bytes": 147,
+    "pass": true
+  },
+  {
+    "task": "h02",
+    "category": "text",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 43,
+    "tokens_nocomments": 43,
+    "bytes": 164,
+    "pass": true
+  },
+  {
+    "task": "h02",
+    "category": "text",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 47,
+    "tokens_nocomments": 47,
+    "bytes": 164,
+    "pass": true
+  },
+  {
+    "task": "h02",
+    "category": "text",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 138,
+    "tokens_nocomments": 132,
+    "bytes": 415,
+    "pass": true
+  },
+  {
+    "task": "h02",
+    "category": "text",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 139,
+    "tokens_nocomments": 133,
+    "bytes": 415,
+    "pass": true
+  },
+  {
+    "task": "h03",
+    "category": "text",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 66,
+    "tokens_nocomments": 66,
+    "bytes": 187,
+    "pass": true
+  },
+  {
+    "task": "h03",
+    "category": "text",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 66,
+    "tokens_nocomments": 66,
+    "bytes": 187,
+    "pass": true
+  },
+  {
+    "task": "h03",
+    "category": "text",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 62,
+    "tokens_nocomments": 62,
+    "bytes": 209,
+    "pass": true
+  },
+  {
+    "task": "h03",
+    "category": "text",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 62,
+    "tokens_nocomments": 62,
+    "bytes": 209,
+    "pass": true
+  },
+  {
+    "task": "h03",
+    "category": "text",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 61,
+    "tokens_nocomments": 61,
+    "bytes": 219,
+    "pass": true
+  },
+  {
+    "task": "h03",
+    "category": "text",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 65,
+    "tokens_nocomments": 65,
+    "bytes": 219,
+    "pass": true
+  },
+  {
+    "task": "h03",
+    "category": "text",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 125,
+    "tokens_nocomments": 119,
+    "bytes": 401,
+    "pass": true
+  },
+  {
+    "task": "h03",
+    "category": "text",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 126,
+    "tokens_nocomments": 120,
+    "bytes": 401,
+    "pass": true
+  },
+  {
+    "task": "h04",
+    "category": "text",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 66,
+    "tokens_nocomments": 66,
+    "bytes": 199,
+    "pass": true
+  },
+  {
+    "task": "h04",
+    "category": "text",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 66,
+    "tokens_nocomments": 66,
+    "bytes": 199,
+    "pass": true
+  },
+  {
+    "task": "h04",
+    "category": "text",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 53,
+    "tokens_nocomments": 53,
+    "bytes": 192,
+    "pass": true
+  },
+  {
+    "task": "h04",
+    "category": "text",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 55,
+    "tokens_nocomments": 55,
+    "bytes": 192,
+    "pass": true
+  },
+  {
+    "task": "h04",
+    "category": "text",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 61,
+    "tokens_nocomments": 61,
+    "bytes": 230,
+    "pass": true
+  },
+  {
+    "task": "h04",
+    "category": "text",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 65,
+    "tokens_nocomments": 65,
+    "bytes": 230,
+    "pass": true
+  },
+  {
+    "task": "h04",
+    "category": "text",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 133,
+    "tokens_nocomments": 127,
+    "bytes": 459,
+    "pass": true
+  },
+  {
+    "task": "h04",
+    "category": "text",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 138,
+    "tokens_nocomments": 132,
+    "bytes": 459,
+    "pass": true
+  },
+  {
+    "task": "h05",
+    "category": "text",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 36,
+    "tokens_nocomments": 36,
+    "bytes": 116,
+    "pass": true
+  },
+  {
+    "task": "h05",
+    "category": "text",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 36,
+    "tokens_nocomments": 36,
+    "bytes": 116,
+    "pass": true
+  },
+  {
+    "task": "h05",
+    "category": "text",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 36,
+    "tokens_nocomments": 36,
+    "bytes": 134,
+    "pass": true
+  },
+  {
+    "task": "h05",
+    "category": "text",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 36,
+    "tokens_nocomments": 36,
+    "bytes": 134,
+    "pass": true
+  },
+  {
+    "task": "h05",
+    "category": "text",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 40,
+    "tokens_nocomments": 40,
+    "bytes": 154,
+    "pass": true
+  },
+  {
+    "task": "h05",
+    "category": "text",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 45,
+    "tokens_nocomments": 45,
+    "bytes": 154,
+    "pass": true
+  },
+  {
+    "task": "h05",
+    "category": "text",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 88,
+    "tokens_nocomments": 82,
+    "bytes": 292,
+    "pass": true
+  },
+  {
+    "task": "h05",
+    "category": "text",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 89,
+    "tokens_nocomments": 83,
+    "bytes": 292,
+    "pass": true
+  },
+  {
+    "task": "h06",
+    "category": "csv",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 122,
+    "tokens_nocomments": 122,
+    "bytes": 349,
+    "pass": true
+  },
+  {
+    "task": "h06",
+    "category": "csv",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 122,
+    "tokens_nocomments": 122,
+    "bytes": 349,
+    "pass": true
+  },
+  {
+    "task": "h06",
+    "category": "csv",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 71,
+    "tokens_nocomments": 71,
+    "bytes": 247,
+    "pass": true
+  },
+  {
+    "task": "h06",
+    "category": "csv",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 72,
+    "tokens_nocomments": 72,
+    "bytes": 247,
+    "pass": true
+  },
+  {
+    "task": "h06",
+    "category": "csv",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 74,
+    "tokens_nocomments": 74,
+    "bytes": 270,
+    "pass": true
+  },
+  {
+    "task": "h06",
+    "category": "csv",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 79,
+    "tokens_nocomments": 79,
+    "bytes": 270,
+    "pass": true
+  },
+  {
+    "task": "h06",
+    "category": "csv",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 165,
+    "tokens_nocomments": 159,
+    "bytes": 534,
+    "pass": true
+  },
+  {
+    "task": "h06",
+    "category": "csv",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 166,
+    "tokens_nocomments": 160,
+    "bytes": 534,
+    "pass": true
+  },
+  {
+    "task": "h07",
+    "category": "csv",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 144,
+    "tokens_nocomments": 144,
+    "bytes": 404,
+    "pass": true
+  },
+  {
+    "task": "h07",
+    "category": "csv",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 144,
+    "tokens_nocomments": 144,
+    "bytes": 404,
+    "pass": true
+  },
+  {
+    "task": "h07",
+    "category": "csv",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 82,
+    "tokens_nocomments": 82,
+    "bytes": 272,
+    "pass": true
+  },
+  {
+    "task": "h07",
+    "category": "csv",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 84,
+    "tokens_nocomments": 84,
+    "bytes": 272,
+    "pass": true
+  },
+  {
+    "task": "h07",
+    "category": "csv",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 89,
+    "tokens_nocomments": 89,
+    "bytes": 302,
+    "pass": true
+  },
+  {
+    "task": "h07",
+    "category": "csv",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 96,
+    "tokens_nocomments": 96,
+    "bytes": 302,
+    "pass": true
+  },
+  {
+    "task": "h07",
+    "category": "csv",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 206,
+    "tokens_nocomments": 200,
+    "bytes": 634,
+    "pass": true
+  },
+  {
+    "task": "h07",
+    "category": "csv",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 211,
+    "tokens_nocomments": 205,
+    "bytes": 634,
+    "pass": true
+  },
+  {
+    "task": "h08",
+    "category": "csv",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 127,
+    "tokens_nocomments": 127,
+    "bytes": 367,
+    "pass": true
+  },
+  {
+    "task": "h08",
+    "category": "csv",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 127,
+    "tokens_nocomments": 127,
+    "bytes": 367,
+    "pass": true
+  },
+  {
+    "task": "h08",
+    "category": "csv",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 64,
+    "tokens_nocomments": 64,
+    "bytes": 227,
+    "pass": true
+  },
+  {
+    "task": "h08",
+    "category": "csv",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 64,
+    "tokens_nocomments": 64,
+    "bytes": 227,
+    "pass": true
+  },
+  {
+    "task": "h08",
+    "category": "csv",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 81,
+    "tokens_nocomments": 81,
+    "bytes": 289,
+    "pass": true
+  },
+  {
+    "task": "h08",
+    "category": "csv",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 87,
+    "tokens_nocomments": 87,
+    "bytes": 289,
+    "pass": true
+  },
+  {
+    "task": "h08",
+    "category": "csv",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 186,
+    "tokens_nocomments": 180,
+    "bytes": 597,
+    "pass": true
+  },
+  {
+    "task": "h08",
+    "category": "csv",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 187,
+    "tokens_nocomments": 181,
+    "bytes": 597,
+    "pass": true
+  },
+  {
+    "task": "h09",
+    "category": "csv",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 129,
+    "tokens_nocomments": 129,
+    "bytes": 395,
+    "pass": true
+  },
+  {
+    "task": "h09",
+    "category": "csv",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 129,
+    "tokens_nocomments": 129,
+    "bytes": 395,
+    "pass": true
+  },
+  {
+    "task": "h09",
+    "category": "csv",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 70,
+    "tokens_nocomments": 70,
+    "bytes": 253,
+    "pass": true
+  },
+  {
+    "task": "h09",
+    "category": "csv",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 71,
+    "tokens_nocomments": 71,
+    "bytes": 253,
+    "pass": true
+  },
+  {
+    "task": "h09",
+    "category": "csv",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 96,
+    "tokens_nocomments": 96,
+    "bytes": 357,
+    "pass": true
+  },
+  {
+    "task": "h09",
+    "category": "csv",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 100,
+    "tokens_nocomments": 100,
+    "bytes": 357,
+    "pass": true
+  },
+  {
+    "task": "h09",
+    "category": "csv",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 185,
+    "tokens_nocomments": 179,
+    "bytes": 638,
+    "pass": true
+  },
+  {
+    "task": "h09",
+    "category": "csv",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 186,
+    "tokens_nocomments": 180,
+    "bytes": 638,
+    "pass": true
+  },
+  {
+    "task": "h10",
+    "category": "csv",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 153,
+    "tokens_nocomments": 153,
+    "bytes": 447,
+    "pass": true
+  },
+  {
+    "task": "h10",
+    "category": "csv",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 153,
+    "tokens_nocomments": 153,
+    "bytes": 447,
+    "pass": true
+  },
+  {
+    "task": "h10",
+    "category": "csv",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 78,
+    "tokens_nocomments": 78,
+    "bytes": 293,
+    "pass": true
+  },
+  {
+    "task": "h10",
+    "category": "csv",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 78,
+    "tokens_nocomments": 78,
+    "bytes": 293,
+    "pass": true
+  },
+  {
+    "task": "h10",
+    "category": "csv",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 105,
+    "tokens_nocomments": 105,
+    "bytes": 354,
+    "pass": true
+  },
+  {
+    "task": "h10",
+    "category": "csv",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 110,
+    "tokens_nocomments": 110,
+    "bytes": 354,
+    "pass": true
+  },
+  {
+    "task": "h10",
+    "category": "csv",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 186,
+    "tokens_nocomments": 180,
+    "bytes": 604,
+    "pass": true
+  },
+  {
+    "task": "h10",
+    "category": "csv",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 187,
+    "tokens_nocomments": 181,
+    "bytes": 604,
+    "pass": true
+  },
+  {
+    "task": "h11",
+    "category": "aggregate",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 137,
+    "tokens_nocomments": 137,
+    "bytes": 375,
+    "pass": true
+  },
+  {
+    "task": "h11",
+    "category": "aggregate",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 137,
+    "tokens_nocomments": 137,
+    "bytes": 375,
+    "pass": true
+  },
+  {
+    "task": "h11",
+    "category": "aggregate",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 71,
+    "tokens_nocomments": 71,
+    "bytes": 264,
+    "pass": true
+  },
+  {
+    "task": "h11",
+    "category": "aggregate",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 71,
+    "tokens_nocomments": 71,
+    "bytes": 264,
+    "pass": true
+  },
+  {
+    "task": "h11",
+    "category": "aggregate",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 96,
+    "tokens_nocomments": 96,
+    "bytes": 330,
+    "pass": true
+  },
+  {
+    "task": "h11",
+    "category": "aggregate",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 101,
+    "tokens_nocomments": 101,
+    "bytes": 330,
+    "pass": true
+  },
+  {
+    "task": "h11",
+    "category": "aggregate",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 224,
+    "tokens_nocomments": 218,
+    "bytes": 717,
+    "pass": true
+  },
+  {
+    "task": "h11",
+    "category": "aggregate",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 225,
+    "tokens_nocomments": 219,
+    "bytes": 717,
+    "pass": true
+  },
+  {
+    "task": "h12",
+    "category": "aggregate",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 92,
+    "tokens_nocomments": 92,
+    "bytes": 283,
+    "pass": true
+  },
+  {
+    "task": "h12",
+    "category": "aggregate",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 92,
+    "tokens_nocomments": 92,
+    "bytes": 283,
+    "pass": true
+  },
+  {
+    "task": "h12",
+    "category": "aggregate",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 67,
+    "tokens_nocomments": 67,
+    "bytes": 270,
+    "pass": true
+  },
+  {
+    "task": "h12",
+    "category": "aggregate",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 67,
+    "tokens_nocomments": 67,
+    "bytes": 270,
+    "pass": true
+  },
+  {
+    "task": "h12",
+    "category": "aggregate",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 91,
+    "tokens_nocomments": 91,
+    "bytes": 320,
+    "pass": true
+  },
+  {
+    "task": "h12",
+    "category": "aggregate",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 97,
+    "tokens_nocomments": 97,
+    "bytes": 320,
+    "pass": true
+  },
+  {
+    "task": "h12",
+    "category": "aggregate",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 154,
+    "tokens_nocomments": 148,
+    "bytes": 530,
+    "pass": true
+  },
+  {
+    "task": "h12",
+    "category": "aggregate",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 156,
+    "tokens_nocomments": 150,
+    "bytes": 530,
+    "pass": true
+  },
+  {
+    "task": "h13",
+    "category": "aggregate",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 62,
+    "tokens_nocomments": 62,
+    "bytes": 206,
+    "pass": true
+  },
+  {
+    "task": "h13",
+    "category": "aggregate",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 62,
+    "tokens_nocomments": 62,
+    "bytes": 206,
+    "pass": true
+  },
+  {
+    "task": "h13",
+    "category": "aggregate",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 42,
+    "tokens_nocomments": 42,
+    "bytes": 163,
+    "pass": true
+  },
+  {
+    "task": "h13",
+    "category": "aggregate",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 42,
+    "tokens_nocomments": 42,
+    "bytes": 163,
+    "pass": true
+  },
+  {
+    "task": "h13",
+    "category": "aggregate",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 60,
+    "tokens_nocomments": 60,
+    "bytes": 224,
+    "pass": true
+  },
+  {
+    "task": "h13",
+    "category": "aggregate",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 64,
+    "tokens_nocomments": 64,
+    "bytes": 224,
+    "pass": true
+  },
+  {
+    "task": "h13",
+    "category": "aggregate",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 133,
+    "tokens_nocomments": 127,
+    "bytes": 440,
+    "pass": true
+  },
+  {
+    "task": "h13",
+    "category": "aggregate",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 134,
+    "tokens_nocomments": 128,
+    "bytes": 440,
+    "pass": true
+  },
+  {
+    "task": "h14",
+    "category": "aggregate",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 56,
+    "tokens_nocomments": 56,
+    "bytes": 168,
+    "pass": true
+  },
+  {
+    "task": "h14",
+    "category": "aggregate",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 56,
+    "tokens_nocomments": 56,
+    "bytes": 168,
+    "pass": true
+  },
+  {
+    "task": "h14",
+    "category": "aggregate",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 50,
+    "tokens_nocomments": 50,
+    "bytes": 177,
+    "pass": true
+  },
+  {
+    "task": "h14",
+    "category": "aggregate",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 50,
+    "tokens_nocomments": 50,
+    "bytes": 177,
+    "pass": true
+  },
+  {
+    "task": "h14",
+    "category": "aggregate",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 53,
+    "tokens_nocomments": 53,
+    "bytes": 211,
+    "pass": true
+  },
+  {
+    "task": "h14",
+    "category": "aggregate",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 57,
+    "tokens_nocomments": 57,
+    "bytes": 211,
+    "pass": true
+  },
+  {
+    "task": "h14",
+    "category": "aggregate",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 124,
+    "tokens_nocomments": 118,
+    "bytes": 416,
+    "pass": true
+  },
+  {
+    "task": "h14",
+    "category": "aggregate",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 125,
+    "tokens_nocomments": 119,
+    "bytes": 416,
+    "pass": true
+  },
+  {
+    "task": "h15",
+    "category": "aggregate",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 62,
+    "tokens_nocomments": 62,
+    "bytes": 193,
+    "pass": true
+  },
+  {
+    "task": "h15",
+    "category": "aggregate",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 62,
+    "tokens_nocomments": 62,
+    "bytes": 193,
+    "pass": true
+  },
+  {
+    "task": "h15",
+    "category": "aggregate",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 38,
+    "tokens_nocomments": 38,
+    "bytes": 140,
+    "pass": true
+  },
+  {
+    "task": "h15",
+    "category": "aggregate",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 38,
+    "tokens_nocomments": 38,
+    "bytes": 140,
+    "pass": true
+  },
+  {
+    "task": "h15",
+    "category": "aggregate",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 59,
+    "tokens_nocomments": 59,
+    "bytes": 234,
+    "pass": true
+  },
+  {
+    "task": "h15",
+    "category": "aggregate",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 63,
+    "tokens_nocomments": 63,
+    "bytes": 234,
+    "pass": true
+  },
+  {
+    "task": "h15",
+    "category": "aggregate",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 127,
+    "tokens_nocomments": 121,
+    "bytes": 409,
+    "pass": true
+  },
+  {
+    "task": "h15",
+    "category": "aggregate",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 128,
+    "tokens_nocomments": 122,
+    "bytes": 409,
+    "pass": true
+  },
+  {
+    "task": "h16",
+    "category": "parse",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 109,
+    "tokens_nocomments": 109,
+    "bytes": 336,
+    "pass": true
+  },
+  {
+    "task": "h16",
+    "category": "parse",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 109,
+    "tokens_nocomments": 109,
+    "bytes": 336,
+    "pass": true
+  },
+  {
+    "task": "h16",
+    "category": "parse",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 59,
+    "tokens_nocomments": 59,
+    "bytes": 215,
+    "pass": true
+  },
+  {
+    "task": "h16",
+    "category": "parse",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 59,
+    "tokens_nocomments": 59,
+    "bytes": 215,
+    "pass": true
+  },
+  {
+    "task": "h16",
+    "category": "parse",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 84,
+    "tokens_nocomments": 84,
+    "bytes": 299,
+    "pass": true
+  },
+  {
+    "task": "h16",
+    "category": "parse",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 89,
+    "tokens_nocomments": 89,
+    "bytes": 299,
+    "pass": true
+  },
+  {
+    "task": "h16",
+    "category": "parse",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 152,
+    "tokens_nocomments": 146,
+    "bytes": 524,
+    "pass": true
+  },
+  {
+    "task": "h16",
+    "category": "parse",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 153,
+    "tokens_nocomments": 147,
+    "bytes": 524,
+    "pass": true
+  },
+  {
+    "task": "h17",
+    "category": "parse",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 88,
+    "tokens_nocomments": 88,
+    "bytes": 285,
+    "pass": true
+  },
+  {
+    "task": "h17",
+    "category": "parse",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 88,
+    "tokens_nocomments": 88,
+    "bytes": 285,
+    "pass": true
+  },
+  {
+    "task": "h17",
+    "category": "parse",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 63,
+    "tokens_nocomments": 63,
+    "bytes": 251,
+    "pass": true
+  },
+  {
+    "task": "h17",
+    "category": "parse",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 63,
+    "tokens_nocomments": 63,
+    "bytes": 251,
+    "pass": true
+  },
+  {
+    "task": "h17",
+    "category": "parse",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 86,
+    "tokens_nocomments": 86,
+    "bytes": 301,
+    "pass": true
+  },
+  {
+    "task": "h17",
+    "category": "parse",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 90,
+    "tokens_nocomments": 90,
+    "bytes": 301,
+    "pass": true
+  },
+  {
+    "task": "h17",
+    "category": "parse",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 145,
+    "tokens_nocomments": 139,
+    "bytes": 498,
+    "pass": true
+  },
+  {
+    "task": "h17",
+    "category": "parse",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 146,
+    "tokens_nocomments": 140,
+    "bytes": 498,
+    "pass": true
+  },
+  {
+    "task": "h18",
+    "category": "parse",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 62,
+    "tokens_nocomments": 62,
+    "bytes": 188,
+    "pass": true
+  },
+  {
+    "task": "h18",
+    "category": "parse",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 62,
+    "tokens_nocomments": 62,
+    "bytes": 188,
+    "pass": true
+  },
+  {
+    "task": "h18",
+    "category": "parse",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 54,
+    "tokens_nocomments": 54,
+    "bytes": 194,
+    "pass": true
+  },
+  {
+    "task": "h18",
+    "category": "parse",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 54,
+    "tokens_nocomments": 54,
+    "bytes": 194,
+    "pass": true
+  },
+  {
+    "task": "h18",
+    "category": "parse",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 57,
+    "tokens_nocomments": 57,
+    "bytes": 219,
+    "pass": true
+  },
+  {
+    "task": "h18",
+    "category": "parse",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 61,
+    "tokens_nocomments": 61,
+    "bytes": 219,
+    "pass": true
+  },
+  {
+    "task": "h18",
+    "category": "parse",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 126,
+    "tokens_nocomments": 120,
+    "bytes": 433,
+    "pass": true
+  },
+  {
+    "task": "h18",
+    "category": "parse",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 127,
+    "tokens_nocomments": 121,
+    "bytes": 433,
+    "pass": true
+  },
+  {
+    "task": "h19",
+    "category": "parse",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 88,
+    "tokens_nocomments": 88,
+    "bytes": 259,
+    "pass": true
+  },
+  {
+    "task": "h19",
+    "category": "parse",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 88,
+    "tokens_nocomments": 88,
+    "bytes": 259,
+    "pass": true
+  },
+  {
+    "task": "h19",
+    "category": "parse",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 63,
+    "tokens_nocomments": 63,
+    "bytes": 223,
+    "pass": true
+  },
+  {
+    "task": "h19",
+    "category": "parse",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 63,
+    "tokens_nocomments": 63,
+    "bytes": 223,
+    "pass": true
+  },
+  {
+    "task": "h19",
+    "category": "parse",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 81,
+    "tokens_nocomments": 81,
+    "bytes": 279,
+    "pass": true
+  },
+  {
+    "task": "h19",
+    "category": "parse",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 86,
+    "tokens_nocomments": 86,
+    "bytes": 279,
+    "pass": true
+  },
+  {
+    "task": "h19",
+    "category": "parse",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 181,
+    "tokens_nocomments": 175,
+    "bytes": 579,
+    "pass": true
+  },
+  {
+    "task": "h19",
+    "category": "parse",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 182,
+    "tokens_nocomments": 176,
+    "bytes": 579,
+    "pass": true
+  },
+  {
+    "task": "h20",
+    "category": "parse",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 59,
+    "tokens_nocomments": 59,
+    "bytes": 193,
+    "pass": true
+  },
+  {
+    "task": "h20",
+    "category": "parse",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 60,
+    "tokens_nocomments": 60,
+    "bytes": 193,
+    "pass": true
+  },
+  {
+    "task": "h20",
+    "category": "parse",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 50,
+    "tokens_nocomments": 50,
+    "bytes": 195,
+    "pass": true
+  },
+  {
+    "task": "h20",
+    "category": "parse",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 51,
+    "tokens_nocomments": 51,
+    "bytes": 195,
+    "pass": true
+  },
+  {
+    "task": "h20",
+    "category": "parse",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 52,
+    "tokens_nocomments": 52,
+    "bytes": 203,
+    "pass": true
+  },
+  {
+    "task": "h20",
+    "category": "parse",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 57,
+    "tokens_nocomments": 57,
+    "bytes": 203,
+    "pass": true
+  },
+  {
+    "task": "h20",
+    "category": "parse",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 160,
+    "tokens_nocomments": 154,
+    "bytes": 533,
+    "pass": true
+  },
+  {
+    "task": "h20",
+    "category": "parse",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 163,
+    "tokens_nocomments": 157,
+    "bytes": 533,
+    "pass": true
+  },
+  {
+    "task": "h21",
+    "category": "validate",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 64,
+    "tokens_nocomments": 64,
+    "bytes": 211,
+    "pass": true
+  },
+  {
+    "task": "h21",
+    "category": "validate",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 65,
+    "tokens_nocomments": 65,
+    "bytes": 211,
+    "pass": true
+  },
+  {
+    "task": "h21",
+    "category": "validate",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 61,
+    "tokens_nocomments": 61,
+    "bytes": 229,
+    "pass": true
+  },
+  {
+    "task": "h21",
+    "category": "validate",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 61,
+    "tokens_nocomments": 61,
+    "bytes": 229,
+    "pass": true
+  },
+  {
+    "task": "h21",
+    "category": "validate",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 68,
+    "tokens_nocomments": 68,
+    "bytes": 256,
+    "pass": true
+  },
+  {
+    "task": "h21",
+    "category": "validate",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 72,
+    "tokens_nocomments": 72,
+    "bytes": 256,
+    "pass": true
+  },
+  {
+    "task": "h21",
+    "category": "validate",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 173,
+    "tokens_nocomments": 167,
+    "bytes": 574,
+    "pass": true
+  },
+  {
+    "task": "h21",
+    "category": "validate",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 174,
+    "tokens_nocomments": 168,
+    "bytes": 574,
+    "pass": true
+  },
+  {
+    "task": "h22",
+    "category": "validate",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 99,
+    "tokens_nocomments": 99,
+    "bytes": 288,
+    "pass": true
+  },
+  {
+    "task": "h22",
+    "category": "validate",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 100,
+    "tokens_nocomments": 100,
+    "bytes": 288,
+    "pass": true
+  },
+  {
+    "task": "h22",
+    "category": "validate",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 72,
+    "tokens_nocomments": 72,
+    "bytes": 241,
+    "pass": true
+  },
+  {
+    "task": "h22",
+    "category": "validate",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 73,
+    "tokens_nocomments": 73,
+    "bytes": 241,
+    "pass": true
+  },
+  {
+    "task": "h22",
+    "category": "validate",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 73,
+    "tokens_nocomments": 73,
+    "bytes": 254,
+    "pass": true
+  },
+  {
+    "task": "h22",
+    "category": "validate",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 78,
+    "tokens_nocomments": 78,
+    "bytes": 254,
+    "pass": true
+  },
+  {
+    "task": "h22",
+    "category": "validate",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 197,
+    "tokens_nocomments": 191,
+    "bytes": 637,
+    "pass": true
+  },
+  {
+    "task": "h22",
+    "category": "validate",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 201,
+    "tokens_nocomments": 195,
+    "bytes": 637,
+    "pass": true
+  },
+  {
+    "task": "h23",
+    "category": "validate",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 80,
+    "tokens_nocomments": 80,
+    "bytes": 261,
+    "pass": true
+  },
+  {
+    "task": "h23",
+    "category": "validate",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 80,
+    "tokens_nocomments": 80,
+    "bytes": 261,
+    "pass": true
+  },
+  {
+    "task": "h23",
+    "category": "validate",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 72,
+    "tokens_nocomments": 72,
+    "bytes": 272,
+    "pass": true
+  },
+  {
+    "task": "h23",
+    "category": "validate",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 72,
+    "tokens_nocomments": 72,
+    "bytes": 272,
+    "pass": true
+  },
+  {
+    "task": "h23",
+    "category": "validate",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 83,
+    "tokens_nocomments": 83,
+    "bytes": 296,
+    "pass": true
+  },
+  {
+    "task": "h23",
+    "category": "validate",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 87,
+    "tokens_nocomments": 87,
+    "bytes": 296,
+    "pass": true
+  },
+  {
+    "task": "h23",
+    "category": "validate",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 154,
+    "tokens_nocomments": 148,
+    "bytes": 493,
+    "pass": true
+  },
+  {
+    "task": "h23",
+    "category": "validate",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 156,
+    "tokens_nocomments": 150,
+    "bytes": 493,
+    "pass": true
+  },
+  {
+    "task": "h24",
+    "category": "validate",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 73,
+    "tokens_nocomments": 73,
+    "bytes": 265,
+    "pass": true
+  },
+  {
+    "task": "h24",
+    "category": "validate",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 73,
+    "tokens_nocomments": 73,
+    "bytes": 265,
+    "pass": true
+  },
+  {
+    "task": "h24",
+    "category": "validate",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 69,
+    "tokens_nocomments": 69,
+    "bytes": 278,
+    "pass": true
+  },
+  {
+    "task": "h24",
+    "category": "validate",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 69,
+    "tokens_nocomments": 69,
+    "bytes": 278,
+    "pass": true
+  },
+  {
+    "task": "h24",
+    "category": "validate",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 74,
+    "tokens_nocomments": 74,
+    "bytes": 303,
+    "pass": true
+  },
+  {
+    "task": "h24",
+    "category": "validate",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 80,
+    "tokens_nocomments": 80,
+    "bytes": 303,
+    "pass": true
+  },
+  {
+    "task": "h24",
+    "category": "validate",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 218,
+    "tokens_nocomments": 212,
+    "bytes": 750,
+    "pass": true
+  },
+  {
+    "task": "h24",
+    "category": "validate",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 220,
+    "tokens_nocomments": 214,
+    "bytes": 750,
+    "pass": true
+  },
+  {
+    "task": "h25",
+    "category": "validate",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 104,
+    "tokens_nocomments": 104,
+    "bytes": 306,
+    "pass": true
+  },
+  {
+    "task": "h25",
+    "category": "validate",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 104,
+    "tokens_nocomments": 104,
+    "bytes": 306,
+    "pass": true
+  },
+  {
+    "task": "h25",
+    "category": "validate",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 85,
+    "tokens_nocomments": 85,
+    "bytes": 289,
+    "pass": true
+  },
+  {
+    "task": "h25",
+    "category": "validate",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 85,
+    "tokens_nocomments": 85,
+    "bytes": 289,
+    "pass": true
+  },
+  {
+    "task": "h25",
+    "category": "validate",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 104,
+    "tokens_nocomments": 104,
+    "bytes": 332,
+    "pass": true
+  },
+  {
+    "task": "h25",
+    "category": "validate",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 109,
+    "tokens_nocomments": 109,
+    "bytes": 332,
+    "pass": true
+  },
+  {
+    "task": "h25",
+    "category": "validate",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 197,
+    "tokens_nocomments": 191,
+    "bytes": 614,
+    "pass": true
+  },
+  {
+    "task": "h25",
+    "category": "validate",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 198,
+    "tokens_nocomments": 192,
+    "bytes": 614,
+    "pass": true
+  },
+  {
+    "task": "h26",
+    "category": "reshape",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 63,
+    "tokens_nocomments": 63,
+    "bytes": 205,
+    "pass": true
+  },
+  {
+    "task": "h26",
+    "category": "reshape",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 63,
+    "tokens_nocomments": 63,
+    "bytes": 205,
+    "pass": true
+  },
+  {
+    "task": "h26",
+    "category": "reshape",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 59,
+    "tokens_nocomments": 59,
+    "bytes": 216,
+    "pass": true
+  },
+  {
+    "task": "h26",
+    "category": "reshape",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 59,
+    "tokens_nocomments": 59,
+    "bytes": 216,
+    "pass": true
+  },
+  {
+    "task": "h26",
+    "category": "reshape",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 64,
+    "tokens_nocomments": 64,
+    "bytes": 246,
+    "pass": true
+  },
+  {
+    "task": "h26",
+    "category": "reshape",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 68,
+    "tokens_nocomments": 68,
+    "bytes": 246,
+    "pass": true
+  },
+  {
+    "task": "h26",
+    "category": "reshape",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 175,
+    "tokens_nocomments": 169,
+    "bytes": 579,
+    "pass": true
+  },
+  {
+    "task": "h26",
+    "category": "reshape",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 176,
+    "tokens_nocomments": 170,
+    "bytes": 579,
+    "pass": true
+  },
+  {
+    "task": "h27",
+    "category": "reshape",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 68,
+    "tokens_nocomments": 68,
+    "bytes": 241,
+    "pass": true
+  },
+  {
+    "task": "h27",
+    "category": "reshape",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 68,
+    "tokens_nocomments": 68,
+    "bytes": 241,
+    "pass": true
+  },
+  {
+    "task": "h27",
+    "category": "reshape",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 63,
+    "tokens_nocomments": 63,
+    "bytes": 237,
+    "pass": true
+  },
+  {
+    "task": "h27",
+    "category": "reshape",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 63,
+    "tokens_nocomments": 63,
+    "bytes": 237,
+    "pass": true
+  },
+  {
+    "task": "h27",
+    "category": "reshape",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 61,
+    "tokens_nocomments": 61,
+    "bytes": 242,
+    "pass": true
+  },
+  {
+    "task": "h27",
+    "category": "reshape",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 65,
+    "tokens_nocomments": 65,
+    "bytes": 242,
+    "pass": true
+  },
+  {
+    "task": "h27",
+    "category": "reshape",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 181,
+    "tokens_nocomments": 175,
+    "bytes": 627,
+    "pass": true
+  },
+  {
+    "task": "h27",
+    "category": "reshape",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 182,
+    "tokens_nocomments": 176,
+    "bytes": 627,
+    "pass": true
+  },
+  {
+    "task": "h28",
+    "category": "reshape",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 97,
+    "tokens_nocomments": 97,
+    "bytes": 331,
+    "pass": true
+  },
+  {
+    "task": "h28",
+    "category": "reshape",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 97,
+    "tokens_nocomments": 97,
+    "bytes": 331,
+    "pass": true
+  },
+  {
+    "task": "h28",
+    "category": "reshape",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 64,
+    "tokens_nocomments": 64,
+    "bytes": 252,
+    "pass": true
+  },
+  {
+    "task": "h28",
+    "category": "reshape",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 64,
+    "tokens_nocomments": 64,
+    "bytes": 252,
+    "pass": true
+  },
+  {
+    "task": "h28",
+    "category": "reshape",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 70,
+    "tokens_nocomments": 70,
+    "bytes": 263,
+    "pass": true
+  },
+  {
+    "task": "h28",
+    "category": "reshape",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 74,
+    "tokens_nocomments": 74,
+    "bytes": 263,
+    "pass": true
+  },
+  {
+    "task": "h28",
+    "category": "reshape",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 183,
+    "tokens_nocomments": 177,
+    "bytes": 618,
+    "pass": true
+  },
+  {
+    "task": "h28",
+    "category": "reshape",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 184,
+    "tokens_nocomments": 178,
+    "bytes": 618,
+    "pass": true
+  },
+  {
+    "task": "h29",
+    "category": "reshape",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 72,
+    "tokens_nocomments": 72,
+    "bytes": 234,
+    "pass": true
+  },
+  {
+    "task": "h29",
+    "category": "reshape",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 72,
+    "tokens_nocomments": 72,
+    "bytes": 234,
+    "pass": true
+  },
+  {
+    "task": "h29",
+    "category": "reshape",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 65,
+    "tokens_nocomments": 65,
+    "bytes": 231,
+    "pass": true
+  },
+  {
+    "task": "h29",
+    "category": "reshape",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 66,
+    "tokens_nocomments": 66,
+    "bytes": 231,
+    "pass": true
+  },
+  {
+    "task": "h29",
+    "category": "reshape",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 65,
+    "tokens_nocomments": 65,
+    "bytes": 241,
+    "pass": true
+  },
+  {
+    "task": "h29",
+    "category": "reshape",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 69,
+    "tokens_nocomments": 69,
+    "bytes": 241,
+    "pass": true
+  },
+  {
+    "task": "h29",
+    "category": "reshape",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 177,
+    "tokens_nocomments": 171,
+    "bytes": 580,
+    "pass": true
+  },
+  {
+    "task": "h29",
+    "category": "reshape",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 178,
+    "tokens_nocomments": 172,
+    "bytes": 580,
+    "pass": true
+  },
+  {
+    "task": "h30",
+    "category": "reshape",
+    "lang": "howl",
+    "encoding": "cl100k_base",
+    "tokens": 62,
+    "tokens_nocomments": 62,
+    "bytes": 197,
+    "pass": true
+  },
+  {
+    "task": "h30",
+    "category": "reshape",
+    "lang": "howl",
+    "encoding": "o200k_base",
+    "tokens": 62,
+    "tokens_nocomments": 62,
+    "bytes": 197,
+    "pass": true
+  },
+  {
+    "task": "h30",
+    "category": "reshape",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 58,
+    "tokens_nocomments": 58,
+    "bytes": 222,
+    "pass": true
+  },
+  {
+    "task": "h30",
+    "category": "reshape",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 58,
+    "tokens_nocomments": 58,
+    "bytes": 222,
+    "pass": true
+  },
+  {
+    "task": "h30",
+    "category": "reshape",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 68,
+    "tokens_nocomments": 68,
+    "bytes": 270,
+    "pass": true
+  },
+  {
+    "task": "h30",
+    "category": "reshape",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 72,
+    "tokens_nocomments": 72,
+    "bytes": 270,
+    "pass": true
+  },
+  {
+    "task": "h30",
+    "category": "reshape",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 155,
+    "tokens_nocomments": 149,
+    "bytes": 536,
+    "pass": true
+  },
+  {
+    "task": "h30",
+    "category": "reshape",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 156,
+    "tokens_nocomments": 150,
+    "bytes": 536,
+    "pass": true
+  },
+  {
+    "task": "h31",
+    "category": "walk",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 41,
+    "tokens_nocomments": 41,
+    "bytes": 146,
+    "pass": true
+  },
+  {
+    "task": "h31",
+    "category": "walk",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 41,
+    "tokens_nocomments": 41,
+    "bytes": 146,
+    "pass": true
+  },
+  {
+    "task": "h31",
+    "category": "walk",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 54,
+    "tokens_nocomments": 54,
+    "bytes": 226,
+    "pass": true
+  },
+  {
+    "task": "h31",
+    "category": "walk",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 58,
+    "tokens_nocomments": 58,
+    "bytes": 226,
+    "pass": true
+  },
+  {
+    "task": "h31",
+    "category": "walk",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 141,
+    "tokens_nocomments": 135,
+    "bytes": 444,
+    "pass": true
+  },
+  {
+    "task": "h31",
+    "category": "walk",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 142,
+    "tokens_nocomments": 136,
+    "bytes": 444,
+    "pass": true
+  },
+  {
+    "task": "h32",
+    "category": "walk",
+    "lang": "python",
+    "encoding": "cl100k_base",
+    "tokens": 57,
+    "tokens_nocomments": 57,
+    "bytes": 199,
+    "pass": true
+  },
+  {
+    "task": "h32",
+    "category": "walk",
+    "lang": "python",
+    "encoding": "o200k_base",
+    "tokens": 57,
+    "tokens_nocomments": 57,
+    "bytes": 199,
+    "pass": true
+  },
+  {
+    "task": "h32",
+    "category": "walk",
+    "lang": "js",
+    "encoding": "cl100k_base",
+    "tokens": 109,
+    "tokens_nocomments": 109,
+    "bytes": 429,
+    "pass": true
+  },
+  {
+    "task": "h32",
+    "category": "walk",
+    "lang": "js",
+    "encoding": "o200k_base",
+    "tokens": 114,
+    "tokens_nocomments": 114,
+    "bytes": 429,
+    "pass": true
+  },
+  {
+    "task": "h32",
+    "category": "walk",
+    "lang": "go",
+    "encoding": "cl100k_base",
+    "tokens": 196,
+    "tokens_nocomments": 190,
+    "bytes": 645,
+    "pass": true
+  },
+  {
+    "task": "h32",
+    "category": "walk",
+    "lang": "go",
+    "encoding": "o200k_base",
+    "tokens": 198,
+    "tokens_nocomments": 192,
+    "bytes": 645,
+    "pass": true
+  }
+]
diff --git a/experiments/token-efficiency-heldout-v0/results/verification.json b/experiments/token-efficiency-heldout-v0/results/verification.json
new file mode 100644
index 0000000..37eab5b
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/results/verification.json
@@ -0,0 +1,758 @@
+[
+  {
+    "task": "h01",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "b4de690a94feafe6fee53dace2c98c1b54e3df174f28831cb079979be816b59e"
+  },
+  {
+    "task": "h02",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "6cdaac65ab7e5cf8fe48bd36f7566f824ec9cf809d05d5d840e308d70bf9ae2e"
+  },
+  {
+    "task": "h03",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "28fe212b9313bf0a87ac292098626b862d3746ce8f37e1e6183848703b6cd9c7"
+  },
+  {
+    "task": "h04",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "a86c6fb7c64a9127f2f002db65af8dcf737c9b269192e73c4d7b0cd9e5fdad05"
+  },
+  {
+    "task": "h05",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "286b716b9155ba003e7dd618ad42436e16f387ba8ff0a5e2b6ec4048756402a7"
+  },
+  {
+    "task": "h06",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "1850ea29d9337bf1912d6b9105ec09ba8afd39ffc37b62db0d0bc7852cf276ea"
+  },
+  {
+    "task": "h07",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "c371b635596e230839e25e9aac14bd6b0db14844564e5b772985c022117c29bb"
+  },
+  {
+    "task": "h08",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "91330ac95a1cf8745bd2040042d449a0bdf77b3b76b96cd3e1e377dc5863d054"
+  },
+  {
+    "task": "h09",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "b1749f7053e6511d0ebaa8d2dc050acaa3d4f48e31cbbe2f953d88822b79ff54"
+  },
+  {
+    "task": "h10",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "62be5b650b51ddebefeaefef2181e2898dd97d90b3ec8186c6f1756102f95ad7"
+  },
+  {
+    "task": "h11",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "266248df5f57d46c001045d0c1b90ed51f649b0a2c442a7ebc9cb35eef28fd48"
+  },
+  {
+    "task": "h12",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "e06880a540f1b2f853b00cc0d366b22ccd30ab22b79a3c5b90aaca4acb369792"
+  },
+  {
+    "task": "h13",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "939b6aeaeef5e96f0dc07a7524507f95e88b54ce497aeb004251c6c88f7ef1fa"
+  },
+  {
+    "task": "h14",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "851d0ed640e674c73a1c244ea9fe109db706532c928764013013e3c3ea30d7f7"
+  },
+  {
+    "task": "h15",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "73c5cbbe59036fb1217d545bc76665bf5f58daaa1c0a472c9c96a38447acc777"
+  },
+  {
+    "task": "h16",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "1f22506e18b96edf2a2b8eb99094dccff1940d1f6fc8f9786de9e17f2729c580"
+  },
+  {
+    "task": "h17",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "1b2c6bbde05473dd1dc539468be7ae3d15a43d6d0b8971813634c29be717c146"
+  },
+  {
+    "task": "h18",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "e92bf7ed57ed7b1dbf864122c4d5f9fba4892313e24ff65a1fbfbfe728ec1c40"
+  },
+  {
+    "task": "h19",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "b7bd999315e5a09f94b89bd3e24218b21978295ffae16935a9e1072c41b65340"
+  },
+  {
+    "task": "h20",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "0fc488b6133d71fafc955ff1f6e1f34b4dc18a8fd2859a5882a436d65e3833f6"
+  },
+  {
+    "task": "h21",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "131d369bd2ec71182aed19a80ef6bcba25018dcd498ca6428fbc7b23049d5061"
+  },
+  {
+    "task": "h22",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "348bf96094d72de914a2347556a87423d1cb69f54fb16a40273e2305976db81f"
+  },
+  {
+    "task": "h23",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "f998d8189a63721548335c6ab412c7c84caf3cf3e32d01a75cb7d1fa778209b3"
+  },
+  {
+    "task": "h24",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "9b9195ed368ae5d0045d015fbf5d8a7ac70a7fff3a994c008c791e53a3758152"
+  },
+  {
+    "task": "h25",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "0d81dc7acde6316c6b8886c54c38aab91568ac5c185d7a275a1c9b6fb1156836"
+  },
+  {
+    "task": "h26",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "bd0deb275b6afeffd3f16ee96fcc69b2ab25d2554ac4b44857cb2d5c93de088d"
+  },
+  {
+    "task": "h27",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "593af43a1306b0aaf5aa8fe356497dcc1f76a5247a33d28f158d6a08de829f45"
+  },
+  {
+    "task": "h28",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "6603f4e0e88857a619f902df4f7a731329dbf4b64074abcf56d997fe429c9b43"
+  },
+  {
+    "task": "h29",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "0953d2707c42c3e00ff597c1e2c23bea4abcd882e14a733e68b29f26f3bdcd02"
+  },
+  {
+    "task": "h30",
+    "lang": "howl",
+    "pass": true,
+    "sha256": "ae952468b04822b901d73767f9e240e4e374c588b1925782a66772b602b099ea"
+  },
+  {
+    "task": "h01",
+    "lang": "python",
+    "pass": true,
+    "sha256": "4f2f858c937c95c8ebd433083a9cfcf091b9ef0775569389009a4b96741e91d2"
+  },
+  {
+    "task": "h02",
+    "lang": "python",
+    "pass": true,
+    "sha256": "f42b028456fa49a3f14a4d19e93df065e7e8fad964ec88a2f63b673b60a2f37f"
+  },
+  {
+    "task": "h03",
+    "lang": "python",
+    "pass": true,
+    "sha256": "9a5ca0de7370dff2af6e0e1083874f90e025c438a1968fd4325c8326e2797eb2"
+  },
+  {
+    "task": "h04",
+    "lang": "python",
+    "pass": true,
+    "sha256": "d0b7ab420461c352c77bdfde9fccf49898407dcdd096751486087ae7ddc97dd6"
+  },
+  {
+    "task": "h05",
+    "lang": "python",
+    "pass": true,
+    "sha256": "baeefb32824ce6be6f02a8b239bb1b6c76f772ca41eb595f54bc8bfe02d237e7"
+  },
+  {
+    "task": "h06",
+    "lang": "python",
+    "pass": true,
+    "sha256": "40f4dd86ba39f55af7f327de0717ca96768df36d30a22f4c4692c48a102afe87"
+  },
+  {
+    "task": "h07",
+    "lang": "python",
+    "pass": true,
+    "sha256": "6420a906c13d0b49409dff5edef52f73d854d5a3ece2a2565cd38d180ed49ca1"
+  },
+  {
+    "task": "h08",
+    "lang": "python",
+    "pass": true,
+    "sha256": "e988e38d6531a0282542dc2346c2bb02ddca1609f563b8678c105bc5b768acb6"
+  },
+  {
+    "task": "h09",
+    "lang": "python",
+    "pass": true,
+    "sha256": "487c78a6fb57e477cf0ddb9e9f4b183dce9d2fa5886c87dbb77a86f9b07418de"
+  },
+  {
+    "task": "h10",
+    "lang": "python",
+    "pass": true,
+    "sha256": "027928342905eac12c6f8a4b4a0779a4b69c33c7bf41fa2dba464866390b6452"
+  },
+  {
+    "task": "h11",
+    "lang": "python",
+    "pass": true,
+    "sha256": "6378cb337a1a4a4b30a0379c13c6f41b798831d001b49eae9fe05a218b765f28"
+  },
+  {
+    "task": "h12",
+    "lang": "python",
+    "pass": true,
+    "sha256": "ef873c7fbfae70e8f0de98ef08d4d4835c592c4258b498be60e55ed50b3e8c65"
+  },
+  {
+    "task": "h13",
+    "lang": "python",
+    "pass": true,
+    "sha256": "94197926ffa52cfa5ce3ef6f43301910a58d59b43a83b9c70d74897c85c47496"
+  },
+  {
+    "task": "h14",
+    "lang": "python",
+    "pass": true,
+    "sha256": "5a0d4c6cd40d504f4e6bac87d03998b3db8e8ed387682bdbfdce1ec76d398eba"
+  },
+  {
+    "task": "h15",
+    "lang": "python",
+    "pass": true,
+    "sha256": "d71970c859e45d298284f87c4b207bbbf9c243924987ea9af3fe30690ee3bd0e"
+  },
+  {
+    "task": "h16",
+    "lang": "python",
+    "pass": true,
+    "sha256": "5370a3a8b05fc6a6d949274fced699d9b20267be3dbaae03ed586c86608e2ac0"
+  },
+  {
+    "task": "h17",
+    "lang": "python",
+    "pass": true,
+    "sha256": "3a5ce38bd09da174b880bf692dd6b61af5cf2bc76461bff4e93ac06bf06a6b5e"
+  },
+  {
+    "task": "h18",
+    "lang": "python",
+    "pass": true,
+    "sha256": "1044de98234037baf6d54c9717c7e3a1c94f3f26b6c7a01bb53f07ff039495da"
+  },
+  {
+    "task": "h19",
+    "lang": "python",
+    "pass": true,
+    "sha256": "1679efba38f6c6ed238305b37e6f9061826a135f27970e615524aa576e511fac"
+  },
+  {
+    "task": "h20",
+    "lang": "python",
+    "pass": true,
+    "sha256": "1efeb040fe4242ad846f56367417124f561b97e10b3d65d4683d27dafb0e7afb"
+  },
+  {
+    "task": "h21",
+    "lang": "python",
+    "pass": true,
+    "sha256": "6f940610e5558d7855dd32e2c1db2e486db98476ad9c88fc43530e1f5164a74f"
+  },
+  {
+    "task": "h22",
+    "lang": "python",
+    "pass": true,
+    "sha256": "73be9f869de3b4adc3f9152eedec3d1ab13b6f2a092384910c453a18ff4e4ed6"
+  },
+  {
+    "task": "h23",
+    "lang": "python",
+    "pass": true,
+    "sha256": "026134d6ff7fe1af88e9293fcd9a7b242733e3cf0c8929ab7f64f017349c828c"
+  },
+  {
+    "task": "h24",
+    "lang": "python",
+    "pass": true,
+    "sha256": "106dee982a5cc21599aaf9456c9ad4caf8192d3d1e7c751a1d8342a5d6d78e75"
+  },
+  {
+    "task": "h25",
+    "lang": "python",
+    "pass": true,
+    "sha256": "bc0316ef6d7737d41a8e90fc51b00fb51dfdb941020f07a60c0853c6578637b6"
+  },
+  {
+    "task": "h26",
+    "lang": "python",
+    "pass": true,
+    "sha256": "d6474388680a2f12e9320272c3e22512389254c58f3935be61639f1c63e6367c"
+  },
+  {
+    "task": "h27",
+    "lang": "python",
+    "pass": true,
+    "sha256": "ab574fe3795318a1221f3062ca65509dedb4dc5a6cb29b1e8acbef0dfd05ca3e"
+  },
+  {
+    "task": "h28",
+    "lang": "python",
+    "pass": true,
+    "sha256": "0fc05f767a70162adb89046931a37fcf4d99f8aa14ac36c7373d7619b91515ae"
+  },
+  {
+    "task": "h29",
+    "lang": "python",
+    "pass": true,
+    "sha256": "7fcfa8b0919f92ad56543f3c8df695415393a471e8ec6bc87ba81db3ac8e3d46"
+  },
+  {
+    "task": "h30",
+    "lang": "python",
+    "pass": true,
+    "sha256": "26f4b7bdfbbb243358f379802551f8617035565d6d73ffd3572ab5b8229c6287"
+  },
+  {
+    "task": "h31",
+    "lang": "python",
+    "pass": true,
+    "sha256": "28295d265e2b3ebb2e1772ce8a71557fecb2093e203ed81e57f25a8e6a2bce80"
+  },
+  {
+    "task": "h32",
+    "lang": "python",
+    "pass": true,
+    "sha256": "35bcc7cd514a8e9e76fff30c286c640dc0bd7ca8c48b6a4c912f0c1a10cb8893"
+  },
+  {
+    "task": "h01",
+    "lang": "js",
+    "pass": true,
+    "sha256": "ed85768b3c7331f4ab8fce7afb6f23cea40ad4861e08bf606d857d49fdca671a"
+  },
+  {
+    "task": "h02",
+    "lang": "js",
+    "pass": true,
+    "sha256": "22d776b0fc78c3b1338e03cc45f36d1e0297507bac0161da7c55c6df6abce1eb"
+  },
+  {
+    "task": "h03",
+    "lang": "js",
+    "pass": true,
+    "sha256": "d487a4c5368009eaac2aba60f03cd5345230ce230aa57372ad4d5bff877557b9"
+  },
+  {
+    "task": "h04",
+    "lang": "js",
+    "pass": true,
+    "sha256": "2ccae5bb12be92bea499ea1385342e72c3acebdfefc4f369336e63ffa973107e"
+  },
+  {
+    "task": "h05",
+    "lang": "js",
+    "pass": true,
+    "sha256": "ec7f32f93db53aa2022a2e3ef417a7ffbcb5656f23f01d3956369b6a8e89ae48"
+  },
+  {
+    "task": "h06",
+    "lang": "js",
+    "pass": true,
+    "sha256": "9566570463f16372eeb6d72f913ee0471245f5da4fe7bb939dabb88604ff0657"
+  },
+  {
+    "task": "h07",
+    "lang": "js",
+    "pass": true,
+    "sha256": "55cf2b6074540a05c9888f438537e7adaff140e4c75c7b023d147b8a586030d4"
+  },
+  {
+    "task": "h08",
+    "lang": "js",
+    "pass": true,
+    "sha256": "982457bac94c08f58c06e3e2572137b8cd53245845782d1dbf3ba0720a64ef9d"
+  },
+  {
+    "task": "h09",
+    "lang": "js",
+    "pass": true,
+    "sha256": "68ba81a2020a4daa2a2f1c96edd732f7a68c2071a988aaa37880debe8f5c4010"
+  },
+  {
+    "task": "h10",
+    "lang": "js",
+    "pass": true,
+    "sha256": "82fd8d9c135b3154eea0815a4cf80a860ed2106745cd9e8e07a13fad79ab102c"
+  },
+  {
+    "task": "h11",
+    "lang": "js",
+    "pass": true,
+    "sha256": "b24eaca279d260c1f7d22de3f111c99b900df0a9050ae46eec3b03abf173f4eb"
+  },
+  {
+    "task": "h12",
+    "lang": "js",
+    "pass": true,
+    "sha256": "a799801ada1a36244eb78e6519a938e877c026bc4f84db2f25cf719a1b79a65e"
+  },
+  {
+    "task": "h13",
+    "lang": "js",
+    "pass": true,
+    "sha256": "a4dad63f39559e88aadf0c3163c359964fb0f430f8e88bd7d380771a04feafa7"
+  },
+  {
+    "task": "h14",
+    "lang": "js",
+    "pass": true,
+    "sha256": "3d45243fad89a36d37d3cd70404af039f57fe03c34c491493345b7a8475b7fb3"
+  },
+  {
+    "task": "h15",
+    "lang": "js",
+    "pass": true,
+    "sha256": "aa24e2f02ea5ee570a0b6701656778693f67cdb7cba37bd6faaebc0c0f4f34bb"
+  },
+  {
+    "task": "h16",
+    "lang": "js",
+    "pass": true,
+    "sha256": "2f068c42a0e9feb554127ed33e3b14330f5120d67f06a03833f8af8fc5991acc"
+  },
+  {
+    "task": "h17",
+    "lang": "js",
+    "pass": true,
+    "sha256": "2063e580378002d9ab92000fb3ee9e1ebc5d406161658274562c54be4d726db4"
+  },
+  {
+    "task": "h18",
+    "lang": "js",
+    "pass": true,
+    "sha256": "3741104ec8552fd070f99756f350e4cc39010c41076f51be1b6ba3190c65f8c7"
+  },
+  {
+    "task": "h19",
+    "lang": "js",
+    "pass": true,
+    "sha256": "c4215c3f9e48c09891d791d1bb6245adf902e0194d49083b42698d96ab4ce5e6"
+  },
+  {
+    "task": "h20",
+    "lang": "js",
+    "pass": true,
+    "sha256": "f04179414e8ae40577158eaa6dbba58ebc852a92ff235b216a0ad7b9c5e437a7"
+  },
+  {
+    "task": "h21",
+    "lang": "js",
+    "pass": true,
+    "sha256": "7c78238feb28d25a7d27b11968c7f68634ff9886478a57f532b773f37f39cc4c"
+  },
+  {
+    "task": "h22",
+    "lang": "js",
+    "pass": true,
+    "sha256": "cb8ef0964bc2cd8c376f9f748a360733c45e7b8cd36f162e9826d19201ba59ba"
+  },
+  {
+    "task": "h23",
+    "lang": "js",
+    "pass": true,
+    "sha256": "7d7dcc055dc0a50b0ecd0e49d79e22a326f6f192850d851be845ae8103d129b6"
+  },
+  {
+    "task": "h24",
+    "lang": "js",
+    "pass": true,
+    "sha256": "494650fed76da0200392c91eb4db61632251692a4b8c69d944f316d1ea8a8d0e"
+  },
+  {
+    "task": "h25",
+    "lang": "js",
+    "pass": true,
+    "sha256": "09a8a50c268f48724bdcd9e1a84d7fae58e295d39fdd913c4703fdca0f6359b7"
+  },
+  {
+    "task": "h26",
+    "lang": "js",
+    "pass": true,
+    "sha256": "fd053154d4eb4a67f7adf22665dd3c88593319d732209bde0290e3a5c108c2b6"
+  },
+  {
+    "task": "h27",
+    "lang": "js",
+    "pass": true,
+    "sha256": "496b5247495ddbe082033e450313d1f32a44058177113ac833296242c0220af5"
+  },
+  {
+    "task": "h28",
+    "lang": "js",
+    "pass": true,
+    "sha256": "cd9ca67668100a0ed0d507a52ebb1cb6ec96af29a2e53e06d03e9c58237cf8c6"
+  },
+  {
+    "task": "h29",
+    "lang": "js",
+    "pass": true,
+    "sha256": "d3a8d32a5f139157b17f725ea3d77b6542d354604ac4edc70337dd4852bb2573"
+  },
+  {
+    "task": "h30",
+    "lang": "js",
+    "pass": true,
+    "sha256": "7d14dc6c7fd758ed87699c167bdf8815a74fb54a453ce37c6198ce4325a948bf"
+  },
+  {
+    "task": "h31",
+    "lang": "js",
+    "pass": true,
+    "sha256": "907d594579552cb74fd8eebc92bfac84b2461ba06830788decc7b637c48c702e"
+  },
+  {
+    "task": "h32",
+    "lang": "js",
+    "pass": true,
+    "sha256": "4ee11e20975ef48d58fba9be3156d2725e18cd467c1411d9a664d1f0c35f1735"
+  },
+  {
+    "task": "h01",
+    "lang": "go",
+    "pass": true,
+    "sha256": "3ef78ec6cef665d45009bf955ec355cb3202cf107745ca0b4089ac5d3ae85a35"
+  },
+  {
+    "task": "h02",
+    "lang": "go",
+    "pass": true,
+    "sha256": "de4b2dcf329b6873970280d091822761df02f9757ca5e656a1c5af07b7613d4a"
+  },
+  {
+    "task": "h03",
+    "lang": "go",
+    "pass": true,
+    "sha256": "56dac09eea18669edeac5ed673a9c97189d02e1d44ddbf6bf4aa45d1386cb25f"
+  },
+  {
+    "task": "h04",
+    "lang": "go",
+    "pass": true,
+    "sha256": "5d787a897183bbff89b9e62415d748c8834bde41e940b80abb2c71143b34d8c5"
+  },
+  {
+    "task": "h05",
+    "lang": "go",
+    "pass": true,
+    "sha256": "c40b4bae5c7846e23e4526c264487c3a0847d8de3b272eedd1141ec193183e8f"
+  },
+  {
+    "task": "h06",
+    "lang": "go",
+    "pass": true,
+    "sha256": "573319e1ed7c078e22392117f8ee4cf13f5f959882d159c6c351da67c2ed6529"
+  },
+  {
+    "task": "h07",
+    "lang": "go",
+    "pass": true,
+    "sha256": "7548872e4ff08c74998fcbc0de1d0194456c12dcb8334bb993471c9fbe94dc28"
+  },
+  {
+    "task": "h08",
+    "lang": "go",
+    "pass": true,
+    "sha256": "6db3310e9cd829dd0c824b29fb1e67024f9fa834ca1b158d2c60f2135fb90f8e"
+  },
+  {
+    "task": "h09",
+    "lang": "go",
+    "pass": true,
+    "sha256": "256421d5a08b40a9609f59aac6215998bdfa32fd9cbf2dea062c3558930fd6ff"
+  },
+  {
+    "task": "h10",
+    "lang": "go",
+    "pass": true,
+    "sha256": "948358441c4bd3033688c8caf8f468d01a69753acbd968203623168681e38873"
+  },
+  {
+    "task": "h11",
+    "lang": "go",
+    "pass": true,
+    "sha256": "fa80a151719f8d5b7dea700e3ae065fe38c072c8ce5ce3ba6fe3222a1ea18b38"
+  },
+  {
+    "task": "h12",
+    "lang": "go",
+    "pass": true,
+    "sha256": "1a536cfc670938fc9dc870b4cb855df3b3c196b5c7c48b406b4758c96688e9d4"
+  },
+  {
+    "task": "h13",
+    "lang": "go",
+    "pass": true,
+    "sha256": "90d71aeeb40ef71422b1b2f6d8b4ea3d5b21ff33d119c613dfc231e194cfcdb8"
+  },
+  {
+    "task": "h14",
+    "lang": "go",
+    "pass": true,
+    "sha256": "bd2dd389e05bb5a7278b0088727fbd5e90c215d793966d2f9319dc4d8db8653e"
+  },
+  {
+    "task": "h15",
+    "lang": "go",
+    "pass": true,
+    "sha256": "933aee58ff41565613c1c1d4a74c753e043357262f7e3436acc75c9d7ef1133c"
+  },
+  {
+    "task": "h16",
+    "lang": "go",
+    "pass": true,
+    "sha256": "f37cc554c73690f0f9a6c6f79de58d10436f144df5ac0165c7aede9633fb01fd"
+  },
+  {
+    "task": "h17",
+    "lang": "go",
+    "pass": true,
+    "sha256": "e66efabe47bc98d5ec5ce46d6abf346466bafe228d4055062dbd447767946bf1"
+  },
+  {
+    "task": "h18",
+    "lang": "go",
+    "pass": true,
+    "sha256": "a4c5a7a84de4bf7e9923182adf24ae925d8f505b3d5baf70f94ba24f1c3721bd"
+  },
+  {
+    "task": "h19",
+    "lang": "go",
+    "pass": true,
+    "sha256": "6166969bdcf4c96adacbeb2d5e4bcc77ef10a04fc14eb6fe297a87eef554b7a0"
+  },
+  {
+    "task": "h20",
+    "lang": "go",
+    "pass": true,
+    "sha256": "f3a1dc8644d457e4c7b45802ac5b338470dbb6348af82e93de32972704fb2cf1"
+  },
+  {
+    "task": "h21",
+    "lang": "go",
+    "pass": true,
+    "sha256": "7dff3d4638679eba64bc0aa826f6df060d6194f89fb27d429d5b5a1a1a5d2e96"
+  },
+  {
+    "task": "h22",
+    "lang": "go",
+    "pass": true,
+    "sha256": "d51d2d23710e5e8fec4a7b449b7738c5bfec1f417170818dff1f6c79c9372c19"
+  },
+  {
+    "task": "h23",
+    "lang": "go",
+    "pass": true,
+    "sha256": "6a88e87572e528b7fb46e1a218fe4ec2d1e6ee7b2eeb38eb97c71cd1bc3bf8b5"
+  },
+  {
+    "task": "h24",
+    "lang": "go",
+    "pass": true,
+    "sha256": "e702f6f81daa29c4c772a9fa76834b157592125dbffc3cc62638d601b5e632da"
+  },
+  {
+    "task": "h25",
+    "lang": "go",
+    "pass": true,
+    "sha256": "2f0aa4987aa2df5d396dc931426d391da36fe1c701bffd7029da3084d405495f"
+  },
+  {
+    "task": "h26",
+    "lang": "go",
+    "pass": true,
+    "sha256": "fd3a41116a6bbb5cd31e137f0b612f5d811a6dee2596d2c0375f9684fb2bc684"
+  },
+  {
+    "task": "h27",
+    "lang": "go",
+    "pass": true,
+    "sha256": "32a052df2580f0d90e86083f0a535d5420abcb4e90595ed666b3af83367b09fe"
+  },
+  {
+    "task": "h28",
+    "lang": "go",
+    "pass": true,
+    "sha256": "0f9ac96ff9361763d8b0d6c971678f702aa790277e70f27d3bb9966dabf1ddd7"
+  },
+  {
+    "task": "h29",
+    "lang": "go",
+    "pass": true,
+    "sha256": "96f195da264dacc8aa963ad40a0a45ba65b4ca47d5897fb20e93bc6350f8be5c"
+  },
+  {
+    "task": "h30",
+    "lang": "go",
+    "pass": true,
+    "sha256": "8265c91a860879e20070185a9699043c265087d383eecc3da8e430dfde81cb2c"
+  },
+  {
+    "task": "h31",
+    "lang": "go",
+    "pass": true,
+    "sha256": "7453c3a0680e5b367e5ab7dc101119f35ae482c14bd517bc62083c4c6b23dc4a"
+  },
+  {
+    "task": "h32",
+    "lang": "go",
+    "pass": true,
+    "sha256": "025371b372077b1408245fac4c5caeb60c3767c5d15bd2659abfe49e56ef0ada"
+  }
+]
diff --git a/experiments/token-efficiency-heldout-v0/run.sh b/experiments/token-efficiency-heldout-v0/run.sh
new file mode 100755
index 0000000..679779e
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/run.sh
@@ -0,0 +1,12 @@
+#!/usr/bin/env bash
+set -euo pipefail
+cd "$(dirname "$0")/../.."
+export GOCACHE="${GOCACHE:-/tmp/hf-token-go-cache}"
+go build -o /tmp/howlframe .
+go build -o /tmp/heldoutverify ./experiments/token-efficiency-heldout-v0/cmd/heldoutverify
+status=0
+/tmp/heldoutverify -howlframe /tmp/howlframe || status=$?
+if [ "$status" -eq 0 ]; then
+  "${TOKBENCH_PYTHON:-python3}" experiments/token-efficiency-heldout-v0/count_tokens.py
+fi
+exit "$status"
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h01.go b/experiments/token-efficiency-heldout-v0/solutions/go/h01.go
new file mode 100644
index 0000000..37f61b5
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h01.go
@@ -0,0 +1,22 @@
+//go:build ignore
+
+package main
+
+import (
+	"os"
+	"strings"
+)
+
+func main() {
+	data, err := os.ReadFile("in/text.txt")
+	if err != nil {
+		panic(err)
+	}
+	lines := strings.Split(string(data), "\n")
+	for i, line := range lines {
+		lines[i] = strings.ToUpper(line)
+	}
+	if err := os.WriteFile("out/result.txt", []byte(strings.Join(lines, "\n")+"\n"), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h02.go b/experiments/token-efficiency-heldout-v0/solutions/go/h02.go
new file mode 100644
index 0000000..7b1951f
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h02.go
@@ -0,0 +1,22 @@
+//go:build ignore
+
+package main
+
+import (
+	"os"
+	"strings"
+)
+
+func main() {
+	data, err := os.ReadFile("in/text.txt")
+	if err != nil {
+		panic(err)
+	}
+	lines := strings.Split(string(data), "\n")
+	for i, j := 0, len(lines)-1; i < j; i, j = i+1, j-1 {
+		lines[i], lines[j] = lines[j], lines[i]
+	}
+	if err := os.WriteFile("out/result.txt", []byte(strings.Join(lines, "\n")+"\n"), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h03.go b/experiments/token-efficiency-heldout-v0/solutions/go/h03.go
new file mode 100644
index 0000000..26c47f8
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h03.go
@@ -0,0 +1,23 @@
+//go:build ignore
+
+package main
+
+import (
+	"fmt"
+	"os"
+	"strings"
+)
+
+func main() {
+	data, err := os.ReadFile("in/text.txt")
+	if err != nil {
+		panic(err)
+	}
+	lines := strings.Split(string(data), "\n")
+	for i, line := range lines {
+		lines[i] = fmt.Sprintf("%d: %s", i+1, line)
+	}
+	if err := os.WriteFile("out/result.txt", []byte(strings.Join(lines, "\n")+"\n"), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h04.go b/experiments/token-efficiency-heldout-v0/solutions/go/h04.go
new file mode 100644
index 0000000..d6c6bc4
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h04.go
@@ -0,0 +1,25 @@
+//go:build ignore
+
+package main
+
+import (
+	"os"
+	"strings"
+)
+
+func main() {
+	data, err := os.ReadFile("in/text.txt")
+	if err != nil {
+		panic(err)
+	}
+	lines := strings.Split(string(data), "\n")
+	var kept []string
+	for _, line := range lines {
+		if trimmed := strings.TrimSpace(line); trimmed != "" {
+			kept = append(kept, trimmed)
+		}
+	}
+	if err := os.WriteFile("out/result.txt", []byte(strings.Join(kept, "\n")+"\n"), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h05.go b/experiments/token-efficiency-heldout-v0/solutions/go/h05.go
new file mode 100644
index 0000000..ec663cc
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h05.go
@@ -0,0 +1,18 @@
+//go:build ignore
+
+package main
+
+import (
+	"os"
+	"strings"
+)
+
+func main() {
+	data, err := os.ReadFile("in/text.txt")
+	if err != nil {
+		panic(err)
+	}
+	if err := os.WriteFile("out/result.txt", []byte(strings.ReplaceAll(string(data), "foo", "bar")+"\n"), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h06.go b/experiments/token-efficiency-heldout-v0/solutions/go/h06.go
new file mode 100644
index 0000000..dc0fece
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h06.go
@@ -0,0 +1,29 @@
+//go:build ignore
+
+package main
+
+import (
+	"os"
+	"strings"
+)
+
+func main() {
+	data, err := os.ReadFile("in/data.csv")
+	if err != nil {
+		panic(err)
+	}
+	lines := strings.Split(string(data), "\n")
+	col := -1
+	for i, name := range strings.Split(lines[0], ",") {
+		if name == "email" {
+			col = i
+		}
+	}
+	var emails []string
+	for _, row := range lines[1:] {
+		emails = append(emails, strings.Split(row, ",")[col])
+	}
+	if err := os.WriteFile("out/result.txt", []byte(strings.Join(emails, "\n")+"\n"), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h07.go b/experiments/token-efficiency-heldout-v0/solutions/go/h07.go
new file mode 100644
index 0000000..10c0c46
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h07.go
@@ -0,0 +1,36 @@
+//go:build ignore
+
+package main
+
+import (
+	"os"
+	"strconv"
+	"strings"
+)
+
+func main() {
+	data, err := os.ReadFile("in/data.csv")
+	if err != nil {
+		panic(err)
+	}
+	lines := strings.Split(string(data), "\n")
+	col := -1
+	for i, name := range strings.Split(lines[0], ",") {
+		if name == "age" {
+			col = i
+		}
+	}
+	kept := []string{lines[0]}
+	for _, row := range lines[1:] {
+		age, err := strconv.Atoi(strings.Split(row, ",")[col])
+		if err != nil {
+			panic(err)
+		}
+		if age >= 30 {
+			kept = append(kept, row)
+		}
+	}
+	if err := os.WriteFile("out/result.csv", []byte(strings.Join(kept, "\n")+"\n"), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h08.go b/experiments/token-efficiency-heldout-v0/solutions/go/h08.go
new file mode 100644
index 0000000..e44dc46
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h08.go
@@ -0,0 +1,35 @@
+//go:build ignore
+
+package main
+
+import (
+	"fmt"
+	"os"
+	"strconv"
+	"strings"
+)
+
+func main() {
+	data, err := os.ReadFile("in/data.csv")
+	if err != nil {
+		panic(err)
+	}
+	lines := strings.Split(string(data), "\n")
+	col := -1
+	for i, name := range strings.Split(lines[0], ",") {
+		if name == "amount" {
+			col = i
+		}
+	}
+	total := 0
+	for _, row := range lines[1:] {
+		amount, err := strconv.Atoi(strings.Split(row, ",")[col])
+		if err != nil {
+			panic(err)
+		}
+		total += amount
+	}
+	if err := os.WriteFile("out/result.txt", []byte(fmt.Sprintf("%d\n", total)), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h09.go b/experiments/token-efficiency-heldout-v0/solutions/go/h09.go
new file mode 100644
index 0000000..5699462
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h09.go
@@ -0,0 +1,33 @@
+//go:build ignore
+
+package main
+
+import (
+	"encoding/json"
+	"os"
+	"strings"
+)
+
+func main() {
+	data, err := os.ReadFile("in/data.csv")
+	if err != nil {
+		panic(err)
+	}
+	lines := strings.Split(string(data), "\n")
+	header := strings.Split(lines[0], ",")
+	rows := []map[string]string{}
+	for _, line := range lines[1:] {
+		row := map[string]string{}
+		for i, value := range strings.Split(line, ",") {
+			row[header[i]] = value
+		}
+		rows = append(rows, row)
+	}
+	out, err := json.MarshalIndent(rows, "", "  ")
+	if err != nil {
+		panic(err)
+	}
+	if err := os.WriteFile("out/result.json", append(out, '\n'), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h10.go b/experiments/token-efficiency-heldout-v0/solutions/go/h10.go
new file mode 100644
index 0000000..5cdd129
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h10.go
@@ -0,0 +1,34 @@
+//go:build ignore
+
+package main
+
+import (
+	"encoding/json"
+	"os"
+	"strings"
+)
+
+func main() {
+	data, err := os.ReadFile("in/data.csv")
+	if err != nil {
+		panic(err)
+	}
+	lines := strings.Split(string(data), "\n")
+	col := -1
+	for i, name := range strings.Split(lines[0], ",") {
+		if name == "dept" {
+			col = i
+		}
+	}
+	counts := map[string]int{}
+	for _, row := range lines[1:] {
+		counts[strings.Split(row, ",")[col]]++
+	}
+	out, err := json.MarshalIndent(counts, "", "  ")
+	if err != nil {
+		panic(err)
+	}
+	if err := os.WriteFile("out/result.json", append(out, '\n'), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h11.go b/experiments/token-efficiency-heldout-v0/solutions/go/h11.go
new file mode 100644
index 0000000..4c36689
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h11.go
@@ -0,0 +1,40 @@
+//go:build ignore
+
+package main
+
+import (
+	"encoding/json"
+	"os"
+	"strconv"
+	"strings"
+)
+
+func main() {
+	data, err := os.ReadFile("in/numbers.txt")
+	if err != nil {
+		panic(err)
+	}
+	lines := strings.Split(string(data), "\n")
+	var minimum, maximum, sum int
+	for i, line := range lines {
+		n, err := strconv.Atoi(line)
+		if err != nil {
+			panic(err)
+		}
+		if i == 0 || n < minimum {
+			minimum = n
+		}
+		if i == 0 || n > maximum {
+			maximum = n
+		}
+		sum += n
+	}
+	result := map[string]int{"min": minimum, "max": maximum, "sum": sum}
+	out, err := json.MarshalIndent(result, "", "  ")
+	if err != nil {
+		panic(err)
+	}
+	if err := os.WriteFile("out/result.json", append(out, '\n'), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h12.go b/experiments/token-efficiency-heldout-v0/solutions/go/h12.go
new file mode 100644
index 0000000..af6b27b
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h12.go
@@ -0,0 +1,29 @@
+//go:build ignore
+
+package main
+
+import (
+	"encoding/json"
+	"os"
+	"strings"
+)
+
+func main() {
+	data, err := os.ReadFile("in/words.txt")
+	if err != nil {
+		panic(err)
+	}
+	counts := map[string]int{}
+	for _, line := range strings.Split(strings.ToLower(string(data)), "\n") {
+		for _, word := range strings.Split(line, " ") {
+			counts[word]++
+		}
+	}
+	out, err := json.MarshalIndent(counts, "", "  ")
+	if err != nil {
+		panic(err)
+	}
+	if err := os.WriteFile("out/result.json", append(out, '\n'), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h13.go b/experiments/token-efficiency-heldout-v0/solutions/go/h13.go
new file mode 100644
index 0000000..4dc2d94
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h13.go
@@ -0,0 +1,27 @@
+//go:build ignore
+
+package main
+
+import (
+	"encoding/json"
+	"fmt"
+	"os"
+)
+
+func main() {
+	data, err := os.ReadFile("in/scores.json")
+	if err != nil {
+		panic(err)
+	}
+	var scores []int
+	if err := json.Unmarshal(data, &scores); err != nil {
+		panic(err)
+	}
+	total := 0
+	for _, s := range scores {
+		total += s
+	}
+	if err := os.WriteFile("out/result.txt", []byte(fmt.Sprintf("%d\n", total/len(scores))), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h14.go b/experiments/token-efficiency-heldout-v0/solutions/go/h14.go
new file mode 100644
index 0000000..0723f41
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h14.go
@@ -0,0 +1,26 @@
+//go:build ignore
+
+package main
+
+import (
+	"fmt"
+	"os"
+	"strings"
+)
+
+func main() {
+	data, err := os.ReadFile("in/app.log")
+	if err != nil {
+		panic(err)
+	}
+	lines := strings.Split(string(data), "\n")
+	count := 0
+	for _, line := range lines {
+		if strings.Contains(line, "ERROR") {
+			count++
+		}
+	}
+	if err := os.WriteFile("out/result.txt", []byte(fmt.Sprintf("%d\n", count)), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h15.go b/experiments/token-efficiency-heldout-v0/solutions/go/h15.go
new file mode 100644
index 0000000..da71680
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h15.go
@@ -0,0 +1,25 @@
+//go:build ignore
+
+package main
+
+import (
+	"os"
+	"strings"
+)
+
+func main() {
+	data, err := os.ReadFile("in/text.txt")
+	if err != nil {
+		panic(err)
+	}
+	lines := strings.Split(string(data), "\n")
+	longest := lines[0]
+	for _, line := range lines[1:] {
+		if len(line) > len(longest) {
+			longest = line
+		}
+	}
+	if err := os.WriteFile("out/result.txt", []byte(longest+"\n"), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h16.go b/experiments/token-efficiency-heldout-v0/solutions/go/h16.go
new file mode 100644
index 0000000..01f854e
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h16.go
@@ -0,0 +1,29 @@
+//go:build ignore
+
+package main
+
+import (
+	"encoding/json"
+	"os"
+	"strings"
+)
+
+func main() {
+	data, err := os.ReadFile("in/config.kv")
+	if err != nil {
+		panic(err)
+	}
+	lines := strings.Split(string(data), "\n")
+	config := map[string]string{}
+	for _, line := range lines {
+		key, value, _ := strings.Cut(line, "=")
+		config[key] = value
+	}
+	out, err := json.MarshalIndent(config, "", "  ")
+	if err != nil {
+		panic(err)
+	}
+	if err := os.WriteFile("out/result.json", append(out, '\n'), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h17.go b/experiments/token-efficiency-heldout-v0/solutions/go/h17.go
new file mode 100644
index 0000000..8001ac2
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h17.go
@@ -0,0 +1,28 @@
+//go:build ignore
+
+package main
+
+import (
+	"encoding/json"
+	"os"
+	"strings"
+)
+
+func main() {
+	data, err := os.ReadFile("in/events.log")
+	if err != nil {
+		panic(err)
+	}
+	lines := strings.Split(string(data), "\n")
+	counts := map[string]int{}
+	for _, line := range lines {
+		counts[strings.Split(line, "|")[0]]++
+	}
+	out, err := json.MarshalIndent(counts, "", "  ")
+	if err != nil {
+		panic(err)
+	}
+	if err := os.WriteFile("out/result.json", append(out, '\n'), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h18.go b/experiments/token-efficiency-heldout-v0/solutions/go/h18.go
new file mode 100644
index 0000000..6ea9d2c
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h18.go
@@ -0,0 +1,23 @@
+//go:build ignore
+
+package main
+
+import (
+	"os"
+	"strings"
+)
+
+func main() {
+	data, err := os.ReadFile("in/emails.txt")
+	if err != nil {
+		panic(err)
+	}
+	lines := strings.Split(string(data), "\n")
+	var domains []string
+	for _, email := range lines {
+		domains = append(domains, strings.Split(email, "@")[1])
+	}
+	if err := os.WriteFile("out/result.txt", []byte(strings.Join(domains, "\n")+"\n"), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h19.go b/experiments/token-efficiency-heldout-v0/solutions/go/h19.go
new file mode 100644
index 0000000..aac5924
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h19.go
@@ -0,0 +1,34 @@
+//go:build ignore
+
+package main
+
+import (
+	"fmt"
+	"os"
+	"strconv"
+	"strings"
+)
+
+func main() {
+	data, err := os.ReadFile("in/times.txt")
+	if err != nil {
+		panic(err)
+	}
+	lines := strings.Split(string(data), "\n")
+	total := 0
+	for _, line := range lines {
+		h, m, _ := strings.Cut(line, ":")
+		hours, err := strconv.Atoi(h)
+		if err != nil {
+			panic(err)
+		}
+		minutes, err := strconv.Atoi(m)
+		if err != nil {
+			panic(err)
+		}
+		total += hours*60 + minutes
+	}
+	if err := os.WriteFile("out/result.txt", []byte(fmt.Sprintf("%d\n", total)), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h20.go b/experiments/token-efficiency-heldout-v0/solutions/go/h20.go
new file mode 100644
index 0000000..c928070
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h20.go
@@ -0,0 +1,31 @@
+//go:build ignore
+
+package main
+
+import (
+	"encoding/json"
+	"os"
+	"strings"
+)
+
+func main() {
+	data, err := os.ReadFile("in/tags.txt")
+	if err != nil {
+		panic(err)
+	}
+	seen := map[string]bool{}
+	unique := []string{}
+	for _, tag := range strings.Split(string(data), ";") {
+		if !seen[tag] {
+			seen[tag] = true
+			unique = append(unique, tag)
+		}
+	}
+	out, err := json.MarshalIndent(unique, "", "  ")
+	if err != nil {
+		panic(err)
+	}
+	if err := os.WriteFile("out/result.json", append(out, '\n'), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h21.go b/experiments/token-efficiency-heldout-v0/solutions/go/h21.go
new file mode 100644
index 0000000..474e203
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h21.go
@@ -0,0 +1,32 @@
+//go:build ignore
+
+package main
+
+import (
+	"encoding/json"
+	"os"
+)
+
+func main() {
+	data, err := os.ReadFile("in/users.json")
+	if err != nil {
+		panic(err)
+	}
+	var users []map[string]any
+	if err := json.Unmarshal(data, &users); err != nil {
+		panic(err)
+	}
+	missing := []any{}
+	for _, user := range users {
+		if _, ok := user["email"]; !ok {
+			missing = append(missing, user["id"])
+		}
+	}
+	out, err := json.MarshalIndent(missing, "", "  ")
+	if err != nil {
+		panic(err)
+	}
+	if err := os.WriteFile("out/result.json", append(out, '\n'), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h22.go b/experiments/token-efficiency-heldout-v0/solutions/go/h22.go
new file mode 100644
index 0000000..6c1078f
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h22.go
@@ -0,0 +1,35 @@
+//go:build ignore
+
+package main
+
+import (
+	"encoding/json"
+	"fmt"
+	"os"
+	"strings"
+)
+
+func main() {
+	data, err := os.ReadFile("in/people.json")
+	if err != nil {
+		panic(err)
+	}
+	var people []struct {
+		ID  string `json:"id"`
+		Age int    `json:"age"`
+	}
+	if err := json.Unmarshal(data, &people); err != nil {
+		panic(err)
+	}
+	var lines []string
+	for _, p := range people {
+		status := "invalid"
+		if p.Age >= 0 && p.Age <= 150 {
+			status = "valid"
+		}
+		lines = append(lines, fmt.Sprintf("%s %s", p.ID, status))
+	}
+	if err := os.WriteFile("out/result.txt", []byte(strings.Join(lines, "\n")+"\n"), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h23.go b/experiments/token-efficiency-heldout-v0/solutions/go/h23.go
new file mode 100644
index 0000000..174d71a
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h23.go
@@ -0,0 +1,31 @@
+//go:build ignore
+
+package main
+
+import (
+	"encoding/json"
+	"os"
+)
+
+func main() {
+	data, err := os.ReadFile("in/ids.json")
+	if err != nil {
+		panic(err)
+	}
+	var ids []string
+	if err := json.Unmarshal(data, &ids); err != nil {
+		panic(err)
+	}
+	result := "ok"
+	seen := map[string]bool{}
+	for _, id := range ids {
+		if seen[id] {
+			result = "duplicate:" + id
+			break
+		}
+		seen[id] = true
+	}
+	if err := os.WriteFile("out/result.txt", []byte(result+"\n"), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h24.go b/experiments/token-efficiency-heldout-v0/solutions/go/h24.go
new file mode 100644
index 0000000..7a29f3a
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h24.go
@@ -0,0 +1,40 @@
+//go:build ignore
+
+package main
+
+import (
+	"encoding/json"
+	"os"
+)
+
+func main() {
+	data, err := os.ReadFile("in/schema.json")
+	if err != nil {
+		panic(err)
+	}
+	var schema []string
+	if err := json.Unmarshal(data, &schema); err != nil {
+		panic(err)
+	}
+	configData, err := os.ReadFile("in/config.json")
+	if err != nil {
+		panic(err)
+	}
+	var config map[string]any
+	if err := json.Unmarshal(configData, &config); err != nil {
+		panic(err)
+	}
+	missing := []string{}
+	for _, key := range schema {
+		if _, ok := config[key]; !ok {
+			missing = append(missing, key)
+		}
+	}
+	out, err := json.MarshalIndent(missing, "", "  ")
+	if err != nil {
+		panic(err)
+	}
+	if err := os.WriteFile("out/result.json", append(out, '\n'), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h25.go b/experiments/token-efficiency-heldout-v0/solutions/go/h25.go
new file mode 100644
index 0000000..ac0461d
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h25.go
@@ -0,0 +1,36 @@
+//go:build ignore
+
+package main
+
+import (
+	"fmt"
+	"os"
+	"strconv"
+	"strings"
+)
+
+func main() {
+	data, err := os.ReadFile("in/seq.txt")
+	if err != nil {
+		panic(err)
+	}
+	lines := strings.Split(string(data), "\n")
+	values := make([]int, len(lines))
+	for i, line := range lines {
+		n, err := strconv.Atoi(line)
+		if err != nil {
+			panic(err)
+		}
+		values[i] = n
+	}
+	result := "sorted"
+	for i := 1; i < len(values); i++ {
+		if values[i] < values[i-1] {
+			result = fmt.Sprintf("unsorted at %d", i+1)
+			break
+		}
+	}
+	if err := os.WriteFile("out/result.txt", []byte(result+"\n"), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h26.go b/experiments/token-efficiency-heldout-v0/solutions/go/h26.go
new file mode 100644
index 0000000..f2fe8a6
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h26.go
@@ -0,0 +1,33 @@
+//go:build ignore
+
+package main
+
+import (
+	"encoding/json"
+	"os"
+)
+
+func main() {
+	data, err := os.ReadFile("in/items.json")
+	if err != nil {
+		panic(err)
+	}
+	var items []struct {
+		ID   string `json:"id"`
+		Name string `json:"name"`
+	}
+	if err := json.Unmarshal(data, &items); err != nil {
+		panic(err)
+	}
+	byID := map[string]string{}
+	for _, item := range items {
+		byID[item.ID] = item.Name
+	}
+	out, err := json.MarshalIndent(byID, "", "  ")
+	if err != nil {
+		panic(err)
+	}
+	if err := os.WriteFile("out/result.json", append(out, '\n'), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h27.go b/experiments/token-efficiency-heldout-v0/solutions/go/h27.go
new file mode 100644
index 0000000..4b9d1af
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h27.go
@@ -0,0 +1,32 @@
+//go:build ignore
+
+package main
+
+import (
+	"encoding/json"
+	"os"
+)
+
+func main() {
+	data, err := os.ReadFile("in/profile.json")
+	if err != nil {
+		panic(err)
+	}
+	var profile struct {
+		User struct {
+			Name string   `json:"name"`
+			Tags []string `json:"tags"`
+		} `json:"user"`
+	}
+	if err := json.Unmarshal(data, &profile); err != nil {
+		panic(err)
+	}
+	result := map[string]any{"name": profile.User.Name, "tag_count": len(profile.User.Tags)}
+	out, err := json.MarshalIndent(result, "", "  ")
+	if err != nil {
+		panic(err)
+	}
+	if err := os.WriteFile("out/result.json", append(out, '\n'), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h28.go b/experiments/token-efficiency-heldout-v0/solutions/go/h28.go
new file mode 100644
index 0000000..cf82923
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h28.go
@@ -0,0 +1,33 @@
+//go:build ignore
+
+package main
+
+import (
+	"encoding/json"
+	"os"
+)
+
+func main() {
+	data, err := os.ReadFile("in/members.json")
+	if err != nil {
+		panic(err)
+	}
+	var members []struct {
+		Team   string `json:"team"`
+		Member string `json:"member"`
+	}
+	if err := json.Unmarshal(data, &members); err != nil {
+		panic(err)
+	}
+	teams := map[string][]string{}
+	for _, m := range members {
+		teams[m.Team] = append(teams[m.Team], m.Member)
+	}
+	out, err := json.MarshalIndent(teams, "", "  ")
+	if err != nil {
+		panic(err)
+	}
+	if err := os.WriteFile("out/result.json", append(out, '\n'), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h29.go b/experiments/token-efficiency-heldout-v0/solutions/go/h29.go
new file mode 100644
index 0000000..3962b2a
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h29.go
@@ -0,0 +1,30 @@
+//go:build ignore
+
+package main
+
+import (
+	"encoding/json"
+	"os"
+)
+
+func main() {
+	data, err := os.ReadFile("in/records.json")
+	if err != nil {
+		panic(err)
+	}
+	var records []map[string]any
+	if err := json.Unmarshal(data, &records); err != nil {
+		panic(err)
+	}
+	slim := []map[string]any{}
+	for _, r := range records {
+		slim = append(slim, map[string]any{"id": r["id"], "status": r["status"]})
+	}
+	out, err := json.MarshalIndent(slim, "", "  ")
+	if err != nil {
+		panic(err)
+	}
+	if err := os.WriteFile("out/result.json", append(out, '\n'), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h30.go b/experiments/token-efficiency-heldout-v0/solutions/go/h30.go
new file mode 100644
index 0000000..ada590e
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h30.go
@@ -0,0 +1,30 @@
+//go:build ignore
+
+package main
+
+import (
+	"encoding/json"
+	"os"
+)
+
+func main() {
+	data, err := os.ReadFile("in/codes.json")
+	if err != nil {
+		panic(err)
+	}
+	var codes map[string]string
+	if err := json.Unmarshal(data, &codes); err != nil {
+		panic(err)
+	}
+	inverted := map[string]string{}
+	for key, value := range codes {
+		inverted[value] = key
+	}
+	out, err := json.MarshalIndent(inverted, "", "  ")
+	if err != nil {
+		panic(err)
+	}
+	if err := os.WriteFile("out/result.json", append(out, '\n'), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h31.go b/experiments/token-efficiency-heldout-v0/solutions/go/h31.go
new file mode 100644
index 0000000..8a3ebd2
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h31.go
@@ -0,0 +1,29 @@
+//go:build ignore
+
+package main
+
+import (
+	"fmt"
+	"io/fs"
+	"os"
+	"path/filepath"
+)
+
+func main() {
+	count := 0
+	err := filepath.WalkDir("in/docs", func(path string, d fs.DirEntry, err error) error {
+		if err != nil {
+			return err
+		}
+		if d.Type().IsRegular() {
+			count++
+		}
+		return nil
+	})
+	if err != nil {
+		panic(err)
+	}
+	if err := os.WriteFile("out/result.txt", []byte(fmt.Sprintf("%d\n", count)), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/go/h32.go b/experiments/token-efficiency-heldout-v0/solutions/go/h32.go
new file mode 100644
index 0000000..092e676
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/go/h32.go
@@ -0,0 +1,35 @@
+//go:build ignore
+
+package main
+
+import (
+	"io/fs"
+	"os"
+	"path/filepath"
+	"sort"
+	"strings"
+)
+
+func main() {
+	var paths []string
+	err := filepath.WalkDir("in/docs", func(path string, d fs.DirEntry, err error) error {
+		if err != nil {
+			return err
+		}
+		if d.Type().IsRegular() && strings.HasSuffix(path, ".txt") {
+			rel, err := filepath.Rel("in/docs", path)
+			if err != nil {
+				return err
+			}
+			paths = append(paths, filepath.ToSlash(rel))
+		}
+		return nil
+	})
+	if err != nil {
+		panic(err)
+	}
+	sort.Strings(paths)
+	if err := os.WriteFile("out/result.txt", []byte(strings.Join(paths, "\n")+"\n"), 0o644); err != nil {
+		panic(err)
+	}
+}
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h01.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h01.howl
new file mode 100644
index 0000000..c86b148
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h01.howl
@@ -0,0 +1,5 @@
+(do
+ (let (lines (lines (read_file "in/text.txt")) out (list))
+  (do
+   (for l lines (append out (to_upper l)))
+   (write_file "out/result.txt" (str_join out "\n")))))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h02.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h02.howl
new file mode 100644
index 0000000..02058f9
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h02.howl
@@ -0,0 +1,5 @@
+(do
+ (let (lines (lines (read_file "in/text.txt")) out (list) i (list_len lines))
+  (do
+   (while (> i 0) (do (dec i) (append out (list_get lines i))))
+   (write_file "out/result.txt" (str_join out "\n")))))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h03.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h03.howl
new file mode 100644
index 0000000..e3a4449
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h03.howl
@@ -0,0 +1,5 @@
+(do
+ (let (lines (lines (read_file "in/text.txt")) out (list) n 0)
+  (do
+   (for l lines (do (inc n) (append out (str n ": " l))))
+   (write_file "out/result.txt" (str_join out "\n")))))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h04.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h04.howl
new file mode 100644
index 0000000..a88dfd1
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h04.howl
@@ -0,0 +1,5 @@
+(do
+ (let (lines (lines (read_file "in/text.txt")) out (list))
+  (do
+   (for l lines (let (t (str_trim l)) (unless (blank t) (append out t))))
+   (write_file "out/result.txt" (str_join out "\n")))))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h05.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h05.howl
new file mode 100644
index 0000000..c3a9d50
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h05.howl
@@ -0,0 +1,3 @@
+(do
+ (let (text (read_text "in/text.txt") out (str_replace text "foo" "bar"))
+  (write_file "out/result.txt" out)))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h06.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h06.howl
new file mode 100644
index 0000000..ebefe0e
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h06.howl
@@ -0,0 +1,7 @@
+(do
+ (let (lines (lines (read_file "in/data.csv")) header (str_split (list_get lines 0) ",") col 0 i 0 out (list))
+  (do
+   (for h header (do (when (= h "email") (set col i)) (inc i)))
+   (set i 0)
+   (for row lines (do (when (> i 0) (append out (list_get (str_split row ",") col))) (inc i)))
+   (write_file "out/result.txt" (str_join out "\n")))))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h07.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h07.howl
new file mode 100644
index 0000000..a74e44b
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h07.howl
@@ -0,0 +1,10 @@
+(do
+ (let (lines (lines (read_file "in/data.csv")) header (str_split (list_get lines 0) ",") col 0 i 0 out (list))
+  (do
+   (for h header (do (when (= h "age") (set col i)) (inc i)))
+   (set i 0)
+   (for row lines
+    (do
+     (if (= i 0) (append out row) (when (>= (to_int (list_get (str_split row ",") col)) 30) (append out row)))
+     (inc i)))
+   (write_file "out/result.csv" (str_join out "\n")))))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h08.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h08.howl
new file mode 100644
index 0000000..b7a41bd
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h08.howl
@@ -0,0 +1,8 @@
+(do
+ (let (lines (lines (read_file "in/data.csv")) header (str_split (list_get lines 0) ",") col 0 i 0 total 0)
+  (do
+   (for h header (do (when (= h "amount") (set col i)) (inc i)))
+   (set i 0)
+   (for row lines
+    (do (when (> i 0) (set total (+ total (to_int (list_get (str_split row ",") col))))) (inc i)))
+   (write_file "out/result.txt" (to_string total)))))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h09.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h09.howl
new file mode 100644
index 0000000..6bef366
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h09.howl
@@ -0,0 +1,12 @@
+(do
+ (let (lines (lines (read_file "in/data.csv")) header (str_split (list_get lines 0) ",") rows (list) i 0)
+  (do
+   (for row lines
+    (do
+     (when (> i 0)
+      (let (vals (str_split row ",") rec (dict) j 0)
+       (do
+        (for h header (do (map_set rec h (list_get vals j)) (inc j)))
+        (append rows rec))))
+     (inc i)))
+   (write_file "out/result.json" (encode_json rows)))))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h10.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h10.howl
new file mode 100644
index 0000000..a991378
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h10.howl
@@ -0,0 +1,12 @@
+(do
+ (let (lines (lines (read_file "in/data.csv")) header (str_split (list_get lines 0) ",") col 0 i 0 counts (dict))
+  (do
+   (for h header (do (when (= h "dept") (set col i)) (inc i)))
+   (set i 0)
+   (for row lines
+    (do
+     (when (> i 0)
+      (let (d (list_get (str_split row ",") col))
+       (map_set counts d (+ (if (map_has counts d) (map_get counts d) 0) 1))))
+     (inc i)))
+   (write_file "out/result.json" (encode_json counts)))))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h11.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h11.howl
new file mode 100644
index 0000000..351b37c
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h11.howl
@@ -0,0 +1,11 @@
+(do
+ (let (lines (lines (read_file "in/numbers.txt")) lo 0 hi 0 sum 0 first 1)
+  (do
+   (for l lines
+    (let (n (to_int l))
+     (do
+      (when (or (= first 1) (< n lo)) (set lo n))
+      (when (or (= first 1) (> n hi)) (set hi n))
+      (set sum (+ sum n))
+      (set first 0))))
+   (write_file "out/result.json" (encode_json (dict ("min" lo) ("max" hi) ("sum" sum)))))))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h12.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h12.howl
new file mode 100644
index 0000000..4aaad1f
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h12.howl
@@ -0,0 +1,7 @@
+(do
+ (let (text (read_text "in/words.txt") lines (str_split (to_lower text) "\n") counts (dict))
+  (do
+   (for l lines
+    (for w (str_split l " ")
+     (map_set counts w (+ (if (map_has counts w) (map_get counts w) 0) 1))))
+   (write_file "out/result.json" (encode_json counts)))))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h13.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h13.howl
new file mode 100644
index 0000000..29f42f3
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h13.howl
@@ -0,0 +1,5 @@
+(do
+ (let (scores (parse_json Any (read_text "in/scores.json")) total 0)
+  (do
+   (for s scores (set total (+ total s)))
+   (write_file "out/result.txt" (to_string (to_int (/ total (list_len scores))))))))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h14.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h14.howl
new file mode 100644
index 0000000..322ff92
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h14.howl
@@ -0,0 +1,5 @@
+(do
+ (let (lines (lines (read_file "in/app.log")) n 0)
+  (do
+   (for l lines (when (str_contains l "ERROR") (inc n)))
+   (write_file "out/result.txt" (to_string n)))))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h15.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h15.howl
new file mode 100644
index 0000000..0fedb86
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h15.howl
@@ -0,0 +1,5 @@
+(do
+ (let (lines (lines (read_file "in/text.txt")) best (list_get lines 0))
+  (do
+   (for l lines (when (> (list_len l) (list_len best)) (set best l)))
+   (write_file "out/result.txt" best))))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h16.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h16.howl
new file mode 100644
index 0000000..e35a13c
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h16.howl
@@ -0,0 +1,9 @@
+(do
+ (let (lines (lines (read_file "in/config.kv")) config (dict))
+  (do
+   (for l lines
+    (let (parts (str_split l "=") rest (list) i 0)
+     (do
+      (for p parts (do (when (> i 0) (append rest p)) (inc i)))
+      (map_set config (list_get parts 0) (str_join rest "=")))))
+   (write_file "out/result.json" (encode_json config)))))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h17.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h17.howl
new file mode 100644
index 0000000..317656f
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h17.howl
@@ -0,0 +1,7 @@
+(do
+ (let (lines (lines (read_file "in/events.log")) counts (dict))
+  (do
+   (for l lines
+    (let (level (list_get (str_split l "|") 0))
+     (map_set counts level (+ (if (map_has counts level) (map_get counts level) 0) 1))))
+   (write_file "out/result.json" (encode_json counts)))))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h18.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h18.howl
new file mode 100644
index 0000000..dd5f18b
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h18.howl
@@ -0,0 +1,5 @@
+(do
+ (let (lines (lines (read_file "in/emails.txt")) out (list))
+  (do
+   (for e lines (append out (list_get (str_split e "@") 1)))
+   (write_file "out/result.txt" (str_join out "\n")))))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h19.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h19.howl
new file mode 100644
index 0000000..c5c46d2
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h19.howl
@@ -0,0 +1,7 @@
+(do
+ (let (lines (lines (read_file "in/times.txt")) total 0)
+  (do
+   (for t lines
+    (let (hm (str_split t ":"))
+     (set total (+ total (+ (* (to_int (list_get hm 0)) 60) (to_int (list_get hm 1)))))))
+   (write_file "out/result.txt" (to_string total)))))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h20.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h20.howl
new file mode 100644
index 0000000..f978be5
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h20.howl
@@ -0,0 +1,5 @@
+(do
+ (let (tags (str_split (read_text "in/tags.txt") ";") out (list))
+  (do
+   (for t tags (unless (list_contains out t) (append out t)))
+   (write_file "out/result.json" (encode_json out)))))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h21.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h21.howl
new file mode 100644
index 0000000..07a8323
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h21.howl
@@ -0,0 +1,5 @@
+(do
+ (let (users (parse_json Any (read_text "in/users.json")) out (list))
+  (do
+   (for u users (unless (map_has u "email") (append out (map_get u "id"))))
+   (write_file "out/result.json" (encode_json out)))))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h22.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h22.howl
new file mode 100644
index 0000000..6899f2d
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h22.howl
@@ -0,0 +1,7 @@
+(do
+ (let (people (parse_json Any (read_text "in/people.json")) out (list))
+  (do
+   (for p people
+    (let (age (map_get p "age"))
+     (append out (str (map_get p "id") " " (if (and (>= age 0) (<= age 150)) "valid" "invalid")))))
+   (write_file "out/result.txt" (str_join out "\n")))))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h23.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h23.howl
new file mode 100644
index 0000000..2d99c9a
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h23.howl
@@ -0,0 +1,7 @@
+(do
+ (let (ids (parse_json Any (read_text "in/ids.json")) seen (list) result "ok")
+  (do
+   (for id ids
+    (when (= result "ok")
+     (if (list_contains seen id) (set result (str "duplicate:" id)) (append seen id))))
+   (write_file "out/result.txt" result))))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h24.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h24.howl
new file mode 100644
index 0000000..51f4e2f
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h24.howl
@@ -0,0 +1,7 @@
+(do
+ (let (schema (parse_json Any (read_text "in/schema.json"))
+       config (parse_json Any (read_text "in/config.json"))
+       out (list))
+  (do
+   (for k schema (unless (map_has config k) (append out k)))
+   (write_file "out/result.json" (encode_json out)))))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h25.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h25.howl
new file mode 100644
index 0000000..7a3923b
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h25.howl
@@ -0,0 +1,10 @@
+(do
+ (let (lines (lines (read_file "in/seq.txt")) prev 0 i 0 result "sorted")
+  (do
+   (for l lines
+    (let (n (to_int l))
+     (do
+      (inc i)
+      (when (and (> i 1) (< n prev) (= result "sorted")) (set result (str "unsorted at " i)))
+      (set prev n))))
+   (write_file "out/result.txt" result))))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h26.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h26.howl
new file mode 100644
index 0000000..a465aaf
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h26.howl
@@ -0,0 +1,5 @@
+(do
+ (let (items (parse_json Any (read_text "in/items.json")) out (dict))
+  (do
+   (for it items (map_set out (map_get it "id") (map_get it "name")))
+   (write_file "out/result.json" (encode_json out)))))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h27.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h27.howl
new file mode 100644
index 0000000..feb7b34
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h27.howl
@@ -0,0 +1,3 @@
+(do
+ (let (profile (parse_json Any (read_text "in/profile.json")) user (map_get profile "user") tags (map_get user "tags"))
+  (write_file "out/result.json" (encode_json (dict ("name" (map_get user "name")) ("tag_count" (list_len tags)))))))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h28.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h28.howl
new file mode 100644
index 0000000..7eb9c25
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h28.howl
@@ -0,0 +1,7 @@
+(do
+ (let (members (parse_json Any (read_text "in/members.json")) teams (dict))
+  (do
+   (for m members
+    (let (team (map_get m "team") group (if (map_has teams team) (map_get teams team) (list)))
+     (do (append group (map_get m "member")) (map_set teams team group))))
+   (write_file "out/result.json" (encode_json teams)))))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h29.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h29.howl
new file mode 100644
index 0000000..6f2324f
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h29.howl
@@ -0,0 +1,5 @@
+(do
+ (let (records (parse_json Any (read_text "in/records.json")) out (list))
+  (do
+   (for r records (append out (dict ("id" (map_get r "id")) ("status" (map_get r "status")))))
+   (write_file "out/result.json" (encode_json out)))))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/howl/h30.howl b/experiments/token-efficiency-heldout-v0/solutions/howl/h30.howl
new file mode 100644
index 0000000..c5ce6eb
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/howl/h30.howl
@@ -0,0 +1,5 @@
+(do
+ (let (codes (parse_json Any (read_text "in/codes.json")) out (dict))
+  (do
+   (for k (map_keys codes) (map_set out (map_get codes k) k))
+   (write_file "out/result.json" (encode_json out)))))
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h01.js b/experiments/token-efficiency-heldout-v0/solutions/js/h01.js
new file mode 100644
index 0000000..1a3d67f
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h01.js
@@ -0,0 +1,4 @@
+const fs = require("fs");
+
+const lines = fs.readFileSync("in/text.txt", "utf8").split("\n");
+fs.writeFileSync("out/result.txt", lines.map((line) => line.toUpperCase()).join("\n") + "\n");
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h02.js b/experiments/token-efficiency-heldout-v0/solutions/js/h02.js
new file mode 100644
index 0000000..5cf469b
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h02.js
@@ -0,0 +1,4 @@
+const fs = require("fs");
+
+const lines = fs.readFileSync("in/text.txt", "utf8").split("\n");
+fs.writeFileSync("out/result.txt", lines.reverse().join("\n") + "\n");
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h03.js b/experiments/token-efficiency-heldout-v0/solutions/js/h03.js
new file mode 100644
index 0000000..4149858
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h03.js
@@ -0,0 +1,5 @@
+const fs = require("fs");
+
+const lines = fs.readFileSync("in/text.txt", "utf8").split("\n");
+const numbered = lines.map((line, i) => `${i + 1}: ${line}`);
+fs.writeFileSync("out/result.txt", numbered.join("\n") + "\n");
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h04.js b/experiments/token-efficiency-heldout-v0/solutions/js/h04.js
new file mode 100644
index 0000000..7863b8b
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h04.js
@@ -0,0 +1,5 @@
+const fs = require("fs");
+
+const lines = fs.readFileSync("in/text.txt", "utf8").split("\n");
+const kept = lines.map((line) => line.trim()).filter((line) => line !== "");
+fs.writeFileSync("out/result.txt", kept.join("\n") + "\n");
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h05.js b/experiments/token-efficiency-heldout-v0/solutions/js/h05.js
new file mode 100644
index 0000000..75970e8
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h05.js
@@ -0,0 +1,4 @@
+const fs = require("fs");
+
+const text = fs.readFileSync("in/text.txt", "utf8");
+fs.writeFileSync("out/result.txt", text.replaceAll("foo", "bar") + "\n");
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h06.js b/experiments/token-efficiency-heldout-v0/solutions/js/h06.js
new file mode 100644
index 0000000..ce764c2
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h06.js
@@ -0,0 +1,6 @@
+const fs = require("fs");
+
+const lines = fs.readFileSync("in/data.csv", "utf8").split("\n");
+const col = lines[0].split(",").indexOf("email");
+const emails = lines.slice(1).map((row) => row.split(",")[col]);
+fs.writeFileSync("out/result.txt", emails.join("\n") + "\n");
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h07.js b/experiments/token-efficiency-heldout-v0/solutions/js/h07.js
new file mode 100644
index 0000000..a4fce3a
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h07.js
@@ -0,0 +1,6 @@
+const fs = require("fs");
+
+const lines = fs.readFileSync("in/data.csv", "utf8").split("\n");
+const col = lines[0].split(",").indexOf("age");
+const kept = lines.slice(1).filter((row) => parseInt(row.split(",")[col], 10) >= 30);
+fs.writeFileSync("out/result.csv", [lines[0], ...kept].join("\n") + "\n");
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h08.js b/experiments/token-efficiency-heldout-v0/solutions/js/h08.js
new file mode 100644
index 0000000..dd96f42
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h08.js
@@ -0,0 +1,6 @@
+const fs = require("fs");
+
+const lines = fs.readFileSync("in/data.csv", "utf8").split("\n");
+const col = lines[0].split(",").indexOf("amount");
+const total = lines.slice(1).reduce((sum, row) => sum + parseInt(row.split(",")[col], 10), 0);
+fs.writeFileSync("out/result.txt", `${total}\n`);
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h09.js b/experiments/token-efficiency-heldout-v0/solutions/js/h09.js
new file mode 100644
index 0000000..fb68954
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h09.js
@@ -0,0 +1,9 @@
+const fs = require("fs");
+
+const lines = fs.readFileSync("in/data.csv", "utf8").split("\n");
+const header = lines[0].split(",");
+const rows = lines.slice(1).map((row) => {
+  const values = row.split(",");
+  return Object.fromEntries(header.map((name, i) => [name, values[i]]));
+});
+fs.writeFileSync("out/result.json", JSON.stringify(rows, null, 2) + "\n");
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h10.js b/experiments/token-efficiency-heldout-v0/solutions/js/h10.js
new file mode 100644
index 0000000..e11722f
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h10.js
@@ -0,0 +1,10 @@
+const fs = require("fs");
+
+const lines = fs.readFileSync("in/data.csv", "utf8").split("\n");
+const col = lines[0].split(",").indexOf("dept");
+const counts = {};
+for (const row of lines.slice(1)) {
+  const dept = row.split(",")[col];
+  counts[dept] = (counts[dept] || 0) + 1;
+}
+fs.writeFileSync("out/result.json", JSON.stringify(counts, null, 2) + "\n");
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h11.js b/experiments/token-efficiency-heldout-v0/solutions/js/h11.js
new file mode 100644
index 0000000..f0dcf8b
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h11.js
@@ -0,0 +1,9 @@
+const fs = require("fs");
+
+const numbers = fs.readFileSync("in/numbers.txt", "utf8").split("\n").map((line) => parseInt(line, 10));
+const result = {
+  min: Math.min(...numbers),
+  max: Math.max(...numbers),
+  sum: numbers.reduce((a, b) => a + b, 0),
+};
+fs.writeFileSync("out/result.json", JSON.stringify(result, null, 2) + "\n");
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h12.js b/experiments/token-efficiency-heldout-v0/solutions/js/h12.js
new file mode 100644
index 0000000..7051485
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h12.js
@@ -0,0 +1,10 @@
+const fs = require("fs");
+
+const text = fs.readFileSync("in/words.txt", "utf8").toLowerCase();
+const counts = {};
+for (const line of text.split("\n")) {
+  for (const word of line.split(" ")) {
+    counts[word] = (counts[word] || 0) + 1;
+  }
+}
+fs.writeFileSync("out/result.json", JSON.stringify(counts, null, 2) + "\n");
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h13.js b/experiments/token-efficiency-heldout-v0/solutions/js/h13.js
new file mode 100644
index 0000000..2c89544
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h13.js
@@ -0,0 +1,5 @@
+const fs = require("fs");
+
+const scores = JSON.parse(fs.readFileSync("in/scores.json", "utf8"));
+const total = scores.reduce((a, b) => a + b, 0);
+fs.writeFileSync("out/result.txt", `${Math.floor(total / scores.length)}\n`);
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h14.js b/experiments/token-efficiency-heldout-v0/solutions/js/h14.js
new file mode 100644
index 0000000..ebbd651
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h14.js
@@ -0,0 +1,5 @@
+const fs = require("fs");
+
+const lines = fs.readFileSync("in/app.log", "utf8").split("\n");
+const count = lines.filter((line) => line.includes("ERROR")).length;
+fs.writeFileSync("out/result.txt", `${count}\n`);
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h15.js b/experiments/token-efficiency-heldout-v0/solutions/js/h15.js
new file mode 100644
index 0000000..3337da1
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h15.js
@@ -0,0 +1,5 @@
+const fs = require("fs");
+
+const lines = fs.readFileSync("in/text.txt", "utf8").split("\n");
+const longest = lines.reduce((best, line) => (line.length > best.length ? line : best));
+fs.writeFileSync("out/result.txt", longest + "\n");
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h16.js b/experiments/token-efficiency-heldout-v0/solutions/js/h16.js
new file mode 100644
index 0000000..05dbacd
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h16.js
@@ -0,0 +1,9 @@
+const fs = require("fs");
+
+const lines = fs.readFileSync("in/config.kv", "utf8").split("\n");
+const config = {};
+for (const line of lines) {
+  const i = line.indexOf("=");
+  config[line.slice(0, i)] = line.slice(i + 1);
+}
+fs.writeFileSync("out/result.json", JSON.stringify(config, null, 2) + "\n");
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h17.js b/experiments/token-efficiency-heldout-v0/solutions/js/h17.js
new file mode 100644
index 0000000..543557b
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h17.js
@@ -0,0 +1,9 @@
+const fs = require("fs");
+
+const lines = fs.readFileSync("in/events.log", "utf8").split("\n");
+const counts = {};
+for (const line of lines) {
+  const level = line.split("|")[0];
+  counts[level] = (counts[level] || 0) + 1;
+}
+fs.writeFileSync("out/result.json", JSON.stringify(counts, null, 2) + "\n");
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h18.js b/experiments/token-efficiency-heldout-v0/solutions/js/h18.js
new file mode 100644
index 0000000..af8b3c8
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h18.js
@@ -0,0 +1,5 @@
+const fs = require("fs");
+
+const emails = fs.readFileSync("in/emails.txt", "utf8").split("\n");
+const domains = emails.map((email) => email.split("@")[1]);
+fs.writeFileSync("out/result.txt", domains.join("\n") + "\n");
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h19.js b/experiments/token-efficiency-heldout-v0/solutions/js/h19.js
new file mode 100644
index 0000000..a95f973
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h19.js
@@ -0,0 +1,8 @@
+const fs = require("fs");
+
+let total = 0;
+for (const line of fs.readFileSync("in/times.txt", "utf8").split("\n")) {
+  const [hours, minutes] = line.split(":").map((part) => parseInt(part, 10));
+  total += hours * 60 + minutes;
+}
+fs.writeFileSync("out/result.txt", `${total}\n`);
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h20.js b/experiments/token-efficiency-heldout-v0/solutions/js/h20.js
new file mode 100644
index 0000000..f4cfba2
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h20.js
@@ -0,0 +1,5 @@
+const fs = require("fs");
+
+const tags = fs.readFileSync("in/tags.txt", "utf8").split(";");
+const unique = [...new Set(tags)];
+fs.writeFileSync("out/result.json", JSON.stringify(unique, null, 2) + "\n");
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h21.js b/experiments/token-efficiency-heldout-v0/solutions/js/h21.js
new file mode 100644
index 0000000..7361506
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h21.js
@@ -0,0 +1,5 @@
+const fs = require("fs");
+
+const users = JSON.parse(fs.readFileSync("in/users.json", "utf8"));
+const missing = users.filter((user) => !("email" in user)).map((user) => user.id);
+fs.writeFileSync("out/result.json", JSON.stringify(missing, null, 2) + "\n");
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h22.js b/experiments/token-efficiency-heldout-v0/solutions/js/h22.js
new file mode 100644
index 0000000..da66c2f
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h22.js
@@ -0,0 +1,5 @@
+const fs = require("fs");
+
+const people = JSON.parse(fs.readFileSync("in/people.json", "utf8"));
+const lines = people.map((p) => `${p.id} ${p.age >= 0 && p.age <= 150 ? "valid" : "invalid"}`);
+fs.writeFileSync("out/result.txt", lines.join("\n") + "\n");
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h23.js b/experiments/token-efficiency-heldout-v0/solutions/js/h23.js
new file mode 100644
index 0000000..8cb80c0
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h23.js
@@ -0,0 +1,13 @@
+const fs = require("fs");
+
+const ids = JSON.parse(fs.readFileSync("in/ids.json", "utf8"));
+let result = "ok";
+const seen = new Set();
+for (const id of ids) {
+  if (seen.has(id)) {
+    result = `duplicate:${id}`;
+    break;
+  }
+  seen.add(id);
+}
+fs.writeFileSync("out/result.txt", result + "\n");
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h24.js b/experiments/token-efficiency-heldout-v0/solutions/js/h24.js
new file mode 100644
index 0000000..96e012f
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h24.js
@@ -0,0 +1,6 @@
+const fs = require("fs");
+
+const schema = JSON.parse(fs.readFileSync("in/schema.json", "utf8"));
+const config = JSON.parse(fs.readFileSync("in/config.json", "utf8"));
+const missing = schema.filter((key) => !(key in config));
+fs.writeFileSync("out/result.json", JSON.stringify(missing, null, 2) + "\n");
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h25.js b/experiments/token-efficiency-heldout-v0/solutions/js/h25.js
new file mode 100644
index 0000000..2ba849e
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h25.js
@@ -0,0 +1,11 @@
+const fs = require("fs");
+
+const values = fs.readFileSync("in/seq.txt", "utf8").split("\n").map((line) => parseInt(line, 10));
+let result = "sorted";
+for (let i = 1; i < values.length; i++) {
+  if (values[i] < values[i - 1]) {
+    result = `unsorted at ${i + 1}`;
+    break;
+  }
+}
+fs.writeFileSync("out/result.txt", result + "\n");
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h26.js b/experiments/token-efficiency-heldout-v0/solutions/js/h26.js
new file mode 100644
index 0000000..494e10e
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h26.js
@@ -0,0 +1,5 @@
+const fs = require("fs");
+
+const items = JSON.parse(fs.readFileSync("in/items.json", "utf8"));
+const byId = Object.fromEntries(items.map((item) => [item.id, item.name]));
+fs.writeFileSync("out/result.json", JSON.stringify(byId, null, 2) + "\n");
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h27.js b/experiments/token-efficiency-heldout-v0/solutions/js/h27.js
new file mode 100644
index 0000000..3441499
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h27.js
@@ -0,0 +1,5 @@
+const fs = require("fs");
+
+const { user } = JSON.parse(fs.readFileSync("in/profile.json", "utf8"));
+const result = { name: user.name, tag_count: user.tags.length };
+fs.writeFileSync("out/result.json", JSON.stringify(result, null, 2) + "\n");
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h28.js b/experiments/token-efficiency-heldout-v0/solutions/js/h28.js
new file mode 100644
index 0000000..245f975
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h28.js
@@ -0,0 +1,8 @@
+const fs = require("fs");
+
+const members = JSON.parse(fs.readFileSync("in/members.json", "utf8"));
+const teams = {};
+for (const m of members) {
+  (teams[m.team] ||= []).push(m.member);
+}
+fs.writeFileSync("out/result.json", JSON.stringify(teams, null, 2) + "\n");
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h29.js b/experiments/token-efficiency-heldout-v0/solutions/js/h29.js
new file mode 100644
index 0000000..6174310
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h29.js
@@ -0,0 +1,5 @@
+const fs = require("fs");
+
+const records = JSON.parse(fs.readFileSync("in/records.json", "utf8"));
+const slim = records.map((r) => ({ id: r.id, status: r.status }));
+fs.writeFileSync("out/result.json", JSON.stringify(slim, null, 2) + "\n");
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h30.js b/experiments/token-efficiency-heldout-v0/solutions/js/h30.js
new file mode 100644
index 0000000..18a27e5
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h30.js
@@ -0,0 +1,5 @@
+const fs = require("fs");
+
+const codes = JSON.parse(fs.readFileSync("in/codes.json", "utf8"));
+const inverted = Object.fromEntries(Object.entries(codes).map(([key, value]) => [value, key]));
+fs.writeFileSync("out/result.json", JSON.stringify(inverted, null, 2) + "\n");
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h31.js b/experiments/token-efficiency-heldout-v0/solutions/js/h31.js
new file mode 100644
index 0000000..9b00340
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h31.js
@@ -0,0 +1,5 @@
+const fs = require("fs");
+
+const entries = fs.readdirSync("in/docs", { recursive: true, withFileTypes: true });
+const count = entries.filter((entry) => entry.isFile()).length;
+fs.writeFileSync("out/result.txt", `${count}\n`);
diff --git a/experiments/token-efficiency-heldout-v0/solutions/js/h32.js b/experiments/token-efficiency-heldout-v0/solutions/js/h32.js
new file mode 100644
index 0000000..62f6e1a
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/js/h32.js
@@ -0,0 +1,9 @@
+const fs = require("fs");
+const path = require("path");
+
+const entries = fs.readdirSync("in/docs", { recursive: true, withFileTypes: true });
+const paths = entries
+  .filter((entry) => entry.isFile() && entry.name.endsWith(".txt"))
+  .map((entry) => path.relative("in/docs", path.join(entry.parentPath ?? entry.path, entry.name)).split(path.sep).join("/"))
+  .sort();
+fs.writeFileSync("out/result.txt", paths.join("\n") + "\n");
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h01.py b/experiments/token-efficiency-heldout-v0/solutions/python/h01.py
new file mode 100644
index 0000000..9dc1c00
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h01.py
@@ -0,0 +1,4 @@
+from pathlib import Path
+
+lines = Path("in/text.txt").read_text().split("\n")
+Path("out/result.txt").write_text("\n".join(line.upper() for line in lines) + "\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h02.py b/experiments/token-efficiency-heldout-v0/solutions/python/h02.py
new file mode 100644
index 0000000..f4a5e69
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h02.py
@@ -0,0 +1,4 @@
+from pathlib import Path
+
+lines = Path("in/text.txt").read_text().split("\n")
+Path("out/result.txt").write_text("\n".join(reversed(lines)) + "\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h03.py b/experiments/token-efficiency-heldout-v0/solutions/python/h03.py
new file mode 100644
index 0000000..93ed2e1
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h03.py
@@ -0,0 +1,5 @@
+from pathlib import Path
+
+lines = Path("in/text.txt").read_text().split("\n")
+numbered = [f"{i}: {line}" for i, line in enumerate(lines, start=1)]
+Path("out/result.txt").write_text("\n".join(numbered) + "\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h04.py b/experiments/token-efficiency-heldout-v0/solutions/python/h04.py
new file mode 100644
index 0000000..e473766
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h04.py
@@ -0,0 +1,5 @@
+from pathlib import Path
+
+lines = Path("in/text.txt").read_text().split("\n")
+kept = [line.strip() for line in lines if line.strip()]
+Path("out/result.txt").write_text("\n".join(kept) + "\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h05.py b/experiments/token-efficiency-heldout-v0/solutions/python/h05.py
new file mode 100644
index 0000000..484fbe2
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h05.py
@@ -0,0 +1,4 @@
+from pathlib import Path
+
+text = Path("in/text.txt").read_text()
+Path("out/result.txt").write_text(text.replace("foo", "bar") + "\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h06.py b/experiments/token-efficiency-heldout-v0/solutions/python/h06.py
new file mode 100644
index 0000000..f4b246a
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h06.py
@@ -0,0 +1,7 @@
+from pathlib import Path
+
+lines = Path("in/data.csv").read_text().split("\n")
+header = lines[0].split(",")
+col = header.index("email")
+emails = [row.split(",")[col] for row in lines[1:]]
+Path("out/result.txt").write_text("\n".join(emails) + "\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h07.py b/experiments/token-efficiency-heldout-v0/solutions/python/h07.py
new file mode 100644
index 0000000..5a8b253
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h07.py
@@ -0,0 +1,7 @@
+from pathlib import Path
+
+lines = Path("in/data.csv").read_text().split("\n")
+header = lines[0].split(",")
+col = header.index("age")
+kept = [row for row in lines[1:] if int(row.split(",")[col]) >= 30]
+Path("out/result.csv").write_text("\n".join([lines[0]] + kept) + "\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h08.py b/experiments/token-efficiency-heldout-v0/solutions/python/h08.py
new file mode 100644
index 0000000..d6f8a58
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h08.py
@@ -0,0 +1,6 @@
+from pathlib import Path
+
+lines = Path("in/data.csv").read_text().split("\n")
+col = lines[0].split(",").index("amount")
+total = sum(int(row.split(",")[col]) for row in lines[1:])
+Path("out/result.txt").write_text(f"{total}\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h09.py b/experiments/token-efficiency-heldout-v0/solutions/python/h09.py
new file mode 100644
index 0000000..8bd60b2
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h09.py
@@ -0,0 +1,7 @@
+import json
+from pathlib import Path
+
+lines = Path("in/data.csv").read_text().split("\n")
+header = lines[0].split(",")
+rows = [dict(zip(header, row.split(","))) for row in lines[1:]]
+Path("out/result.json").write_text(json.dumps(rows, indent=2) + "\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h10.py b/experiments/token-efficiency-heldout-v0/solutions/python/h10.py
new file mode 100644
index 0000000..288f0e3
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h10.py
@@ -0,0 +1,8 @@
+import json
+from collections import Counter
+from pathlib import Path
+
+lines = Path("in/data.csv").read_text().split("\n")
+col = lines[0].split(",").index("dept")
+counts = Counter(row.split(",")[col] for row in lines[1:])
+Path("out/result.json").write_text(json.dumps(counts, indent=2) + "\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h11.py b/experiments/token-efficiency-heldout-v0/solutions/python/h11.py
new file mode 100644
index 0000000..8dd5f08
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h11.py
@@ -0,0 +1,6 @@
+import json
+from pathlib import Path
+
+numbers = [int(line) for line in Path("in/numbers.txt").read_text().split("\n")]
+result = {"min": min(numbers), "max": max(numbers), "sum": sum(numbers)}
+Path("out/result.json").write_text(json.dumps(result, indent=2) + "\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h12.py b/experiments/token-efficiency-heldout-v0/solutions/python/h12.py
new file mode 100644
index 0000000..1f6ff2e
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h12.py
@@ -0,0 +1,7 @@
+import json
+from collections import Counter
+from pathlib import Path
+
+text = Path("in/words.txt").read_text().lower()
+counts = Counter(word for line in text.split("\n") for word in line.split(" "))
+Path("out/result.json").write_text(json.dumps(counts, indent=2) + "\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h13.py b/experiments/token-efficiency-heldout-v0/solutions/python/h13.py
new file mode 100644
index 0000000..8d1d8dc
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h13.py
@@ -0,0 +1,5 @@
+import json
+from pathlib import Path
+
+scores = json.loads(Path("in/scores.json").read_text())
+Path("out/result.txt").write_text(f"{sum(scores) // len(scores)}\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h14.py b/experiments/token-efficiency-heldout-v0/solutions/python/h14.py
new file mode 100644
index 0000000..10e4e17
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h14.py
@@ -0,0 +1,5 @@
+from pathlib import Path
+
+lines = Path("in/app.log").read_text().split("\n")
+count = sum(1 for line in lines if "ERROR" in line)
+Path("out/result.txt").write_text(f"{count}\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h15.py b/experiments/token-efficiency-heldout-v0/solutions/python/h15.py
new file mode 100644
index 0000000..0795912
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h15.py
@@ -0,0 +1,4 @@
+from pathlib import Path
+
+lines = Path("in/text.txt").read_text().split("\n")
+Path("out/result.txt").write_text(max(lines, key=len) + "\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h16.py b/experiments/token-efficiency-heldout-v0/solutions/python/h16.py
new file mode 100644
index 0000000..f77d58f
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h16.py
@@ -0,0 +1,6 @@
+import json
+from pathlib import Path
+
+lines = Path("in/config.kv").read_text().split("\n")
+config = dict(line.split("=", 1) for line in lines)
+Path("out/result.json").write_text(json.dumps(config, indent=2) + "\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h17.py b/experiments/token-efficiency-heldout-v0/solutions/python/h17.py
new file mode 100644
index 0000000..ecacea0
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h17.py
@@ -0,0 +1,7 @@
+import json
+from collections import Counter
+from pathlib import Path
+
+lines = Path("in/events.log").read_text().split("\n")
+counts = Counter(line.split("|")[0] for line in lines)
+Path("out/result.json").write_text(json.dumps(counts, indent=2) + "\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h18.py b/experiments/token-efficiency-heldout-v0/solutions/python/h18.py
new file mode 100644
index 0000000..2d55a68
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h18.py
@@ -0,0 +1,5 @@
+from pathlib import Path
+
+emails = Path("in/emails.txt").read_text().split("\n")
+domains = [email.split("@")[1] for email in emails]
+Path("out/result.txt").write_text("\n".join(domains) + "\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h19.py b/experiments/token-efficiency-heldout-v0/solutions/python/h19.py
new file mode 100644
index 0000000..82a7cf9
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h19.py
@@ -0,0 +1,7 @@
+from pathlib import Path
+
+total = 0
+for line in Path("in/times.txt").read_text().split("\n"):
+    hours, minutes = line.split(":")
+    total += int(hours) * 60 + int(minutes)
+Path("out/result.txt").write_text(f"{total}\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h20.py b/experiments/token-efficiency-heldout-v0/solutions/python/h20.py
new file mode 100644
index 0000000..e9f6023
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h20.py
@@ -0,0 +1,6 @@
+import json
+from pathlib import Path
+
+tags = Path("in/tags.txt").read_text().split(";")
+unique = list(dict.fromkeys(tags))
+Path("out/result.json").write_text(json.dumps(unique, indent=2) + "\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h21.py b/experiments/token-efficiency-heldout-v0/solutions/python/h21.py
new file mode 100644
index 0000000..d52d804
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h21.py
@@ -0,0 +1,6 @@
+import json
+from pathlib import Path
+
+users = json.loads(Path("in/users.json").read_text())
+missing = [user["id"] for user in users if "email" not in user]
+Path("out/result.json").write_text(json.dumps(missing, indent=2) + "\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h22.py b/experiments/token-efficiency-heldout-v0/solutions/python/h22.py
new file mode 100644
index 0000000..c2bb29a
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h22.py
@@ -0,0 +1,6 @@
+import json
+from pathlib import Path
+
+people = json.loads(Path("in/people.json").read_text())
+lines = [f"{p['id']} {'valid' if 0 <= p['age'] <= 150 else 'invalid'}" for p in people]
+Path("out/result.txt").write_text("\n".join(lines) + "\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h23.py b/experiments/token-efficiency-heldout-v0/solutions/python/h23.py
new file mode 100644
index 0000000..22aaadd
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h23.py
@@ -0,0 +1,12 @@
+import json
+from pathlib import Path
+
+ids = json.loads(Path("in/ids.json").read_text())
+result = "ok"
+seen = set()
+for item in ids:
+    if item in seen:
+        result = f"duplicate:{item}"
+        break
+    seen.add(item)
+Path("out/result.txt").write_text(result + "\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h24.py b/experiments/token-efficiency-heldout-v0/solutions/python/h24.py
new file mode 100644
index 0000000..6ec7064
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h24.py
@@ -0,0 +1,7 @@
+import json
+from pathlib import Path
+
+schema = json.loads(Path("in/schema.json").read_text())
+config = json.loads(Path("in/config.json").read_text())
+missing = [key for key in schema if key not in config]
+Path("out/result.json").write_text(json.dumps(missing, indent=2) + "\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h25.py b/experiments/token-efficiency-heldout-v0/solutions/python/h25.py
new file mode 100644
index 0000000..9c4ebda
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h25.py
@@ -0,0 +1,9 @@
+from pathlib import Path
+
+values = [int(line) for line in Path("in/seq.txt").read_text().split("\n")]
+result = "sorted"
+for i in range(1, len(values)):
+    if values[i] < values[i - 1]:
+        result = f"unsorted at {i + 1}"
+        break
+Path("out/result.txt").write_text(result + "\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h26.py b/experiments/token-efficiency-heldout-v0/solutions/python/h26.py
new file mode 100644
index 0000000..cb22b33
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h26.py
@@ -0,0 +1,6 @@
+import json
+from pathlib import Path
+
+items = json.loads(Path("in/items.json").read_text())
+by_id = {item["id"]: item["name"] for item in items}
+Path("out/result.json").write_text(json.dumps(by_id, indent=2) + "\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h27.py b/experiments/token-efficiency-heldout-v0/solutions/python/h27.py
new file mode 100644
index 0000000..7bda077
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h27.py
@@ -0,0 +1,6 @@
+import json
+from pathlib import Path
+
+user = json.loads(Path("in/profile.json").read_text())["user"]
+result = {"name": user["name"], "tag_count": len(user["tags"])}
+Path("out/result.json").write_text(json.dumps(result, indent=2) + "\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h28.py b/experiments/token-efficiency-heldout-v0/solutions/python/h28.py
new file mode 100644
index 0000000..2c5c793
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h28.py
@@ -0,0 +1,8 @@
+import json
+from pathlib import Path
+
+members = json.loads(Path("in/members.json").read_text())
+teams = {}
+for m in members:
+    teams.setdefault(m["team"], []).append(m["member"])
+Path("out/result.json").write_text(json.dumps(teams, indent=2) + "\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h29.py b/experiments/token-efficiency-heldout-v0/solutions/python/h29.py
new file mode 100644
index 0000000..09dfbed
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h29.py
@@ -0,0 +1,6 @@
+import json
+from pathlib import Path
+
+records = json.loads(Path("in/records.json").read_text())
+slim = [{"id": r["id"], "status": r["status"]} for r in records]
+Path("out/result.json").write_text(json.dumps(slim, indent=2) + "\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h30.py b/experiments/token-efficiency-heldout-v0/solutions/python/h30.py
new file mode 100644
index 0000000..e4632c5
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h30.py
@@ -0,0 +1,6 @@
+import json
+from pathlib import Path
+
+codes = json.loads(Path("in/codes.json").read_text())
+inverted = {value: key for key, value in codes.items()}
+Path("out/result.json").write_text(json.dumps(inverted, indent=2) + "\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h31.py b/experiments/token-efficiency-heldout-v0/solutions/python/h31.py
new file mode 100644
index 0000000..a6bbbe8
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h31.py
@@ -0,0 +1,4 @@
+from pathlib import Path
+
+count = sum(1 for path in Path("in/docs").rglob("*") if path.is_file())
+Path("out/result.txt").write_text(f"{count}\n")
diff --git a/experiments/token-efficiency-heldout-v0/solutions/python/h32.py b/experiments/token-efficiency-heldout-v0/solutions/python/h32.py
new file mode 100644
index 0000000..c677d03
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/solutions/python/h32.py
@@ -0,0 +1,5 @@
+from pathlib import Path
+
+root = Path("in/docs")
+paths = sorted(p.relative_to(root).as_posix() for p in root.rglob("*.txt") if p.is_file())
+Path("out/result.txt").write_text("\n".join(paths) + "\n")
diff --git a/experiments/token-efficiency-heldout-v0/tasks/README.md b/experiments/token-efficiency-heldout-v0/tasks/README.md
new file mode 100644
index 0000000..4e269aa
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/tasks/README.md
@@ -0,0 +1,32 @@
+# Held-out token-efficiency tasks (heldout-v0): frozen specs
+
+This file and the others in `tasks/` (`manifest.json`, `build_specs.py`) are
+**frozen**. They were committed before any reference solution was written or
+anything was measured.
+
+## Corpus
+
+There are 32 tasks in 7 categories. None of them is one of the three
+falsifiable-v0 release workflows.
+
+| Category | Tasks |
+|---|---|
+| text | h01–h05 |
+| csv | h06–h10 |
+| aggregate | h11–h15 |
+| parse | h16–h20 |
+| validate | h21–h25 |
+| reshape | h26–h30 |
+| walk | h31–h32 |
+
+## Running a reference
+
+- Each reference runs with its working directory set to a fresh sandbox
+  containing only the task's `inputs`.
+- It may read under `in/` and must write exactly the one `output` path.
+- `text` outputs are compared exactly, after trailing newlines are removed
+  from both sides. `json` outputs are compared as decoded values.
+- `expected` values come from the spec-time oracle in `build_specs.py`, which
+  is part of this freeze. The oracle is not a counted or verified reference.
+- A task that a language cannot express with features on main at freeze time
+  is recorded as a gap, not solved by adding features.
diff --git a/experiments/token-efficiency-heldout-v0/tasks/build_specs.py b/experiments/token-efficiency-heldout-v0/tasks/build_specs.py
new file mode 100644
index 0000000..0312f72
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/tasks/build_specs.py
@@ -0,0 +1,138 @@
+"""Held-out task specs (frozen before any reference solution is written).
+
+Each task lists input files, one output path, an output kind and the expected
+output. Expected outputs are computed here by a spec-time oracle from the task
+text; this file is part of the frozen spec commit and is not a reference
+solution (it is never counted or verified as one).
+
+Output comparison (verifier contract):
+  text: exact match after removing trailing newlines from both sides.
+  json: decoded-value equality.
+"""
+import json, re
+from collections import Counter, OrderedDict
+from pathlib import Path
+
+T = []
+def task(id, cat, desc, inputs, out, kind, expect):
+    T.append({"id": id, "category": cat, "description": desc, "inputs": inputs,
+              "output": out, "kind": kind,
+              "expected": expect if kind == "text" else expect})
+
+def lines(s): return s.split("\n")
+
+# ---- text transforms ----
+txt = "alpha beta\nGamma delta\n\nfoo bar foo\n  padded line  \nlast"
+task("h01", "text", "Read in/text.txt. Write out/result.txt with every line upper-cased, same line order.",
+     {"in/text.txt": txt}, "out/result.txt", "text", "\n".join(l.upper() for l in lines(txt)))
+task("h02", "text", "Read in/text.txt. Write out/result.txt with the lines in reverse order.",
+     {"in/text.txt": txt}, "out/result.txt", "text", "\n".join(reversed(lines(txt))))
+task("h03", "text", "Read in/text.txt. Write out/result.txt where each line is prefixed with its 1-based line number and \": \" (e.g. \"1: alpha beta\").",
+     {"in/text.txt": txt}, "out/result.txt", "text", "\n".join(f"{i+1}: {l}" for i, l in enumerate(lines(txt))))
+task("h04", "text", "Read in/text.txt. Trim leading/trailing spaces of every line, drop lines that are then empty, and write the rest to out/result.txt.",
+     {"in/text.txt": txt}, "out/result.txt", "text", "\n".join(l.strip() for l in lines(txt) if l.strip()))
+task("h05", "text", "Read in/text.txt. Replace every occurrence of the substring \"foo\" with \"bar\" and write the result to out/result.txt.",
+     {"in/text.txt": txt}, "out/result.txt", "text", txt.replace("foo", "bar"))
+
+# ---- CSV ----
+csv = "name,email,age,dept,amount\nAda,ada@example.com,36,eng,120\nBob,bob@test.org,25,ops,75\nCy,cy@example.com,30,eng,40\nDee,dee@corp.net,41,sales,300\nEd,ed@test.org,29,ops,15"
+rows = [dict(zip(lines(csv)[0].split(","), r.split(","))) for r in lines(csv)[1:]]
+task("h06", "csv", "Read in/data.csv (header row, comma-separated, no quoting). Write the email column values, one per line in row order, to out/result.txt.",
+     {"in/data.csv": csv}, "out/result.txt", "text", "\n".join(r["email"] for r in rows))
+task("h07", "csv", "Read in/data.csv. Write out/result.csv containing the header row followed by only the rows whose integer age is >= 30, in input order.",
+     {"in/data.csv": csv}, "out/result.csv", "text", "\n".join([lines(csv)[0]] + [l for l, r in zip(lines(csv)[1:], rows) if int(r["age"]) >= 30]))
+task("h08", "csv", "Read in/data.csv. Write the integer sum of the amount column to out/result.txt.",
+     {"in/data.csv": csv}, "out/result.txt", "text", str(sum(int(r["amount"]) for r in rows)))
+task("h09", "csv", "Read in/data.csv. Write out/result.json: a JSON array with one object per row mapping each header name to the row's string value.",
+     {"in/data.csv": csv}, "out/result.json", "json", rows)
+task("h10", "csv", "Read in/data.csv. Write out/result.json: a JSON object mapping each dept value to the integer number of rows with that dept.",
+     {"in/data.csv": csv}, "out/result.json", "json", dict(Counter(r["dept"] for r in rows)))
+
+# ---- aggregation ----
+nums = "12\n-4\n7\n30\n0\n9"
+ns = [int(x) for x in lines(nums)]
+task("h11", "aggregate", "Read in/numbers.txt (one integer per line). Write out/result.json: {\"min\": m, \"max\": M, \"sum\": s} as integers.",
+     {"in/numbers.txt": nums}, "out/result.json", "json", {"min": min(ns), "max": max(ns), "sum": sum(ns)})
+words = "the cat and the hat\nThe bat and THE cat"
+task("h12", "aggregate", "Read in/words.txt. Lower-case it, split every line on single spaces, and write out/result.json mapping each word to its integer count.",
+     {"in/words.txt": words}, "out/result.json", "json", dict(Counter(w for l in lines(words.lower()) for w in l.split(" "))))
+scores = [88, 92, 79, 65, 100, 71]
+task("h13", "aggregate", "Read in/scores.json (a JSON array of integers). Write out/result.txt containing the integer mean rounded down (floor).",
+     {"in/scores.json": json.dumps(scores)}, "out/result.txt", "text", str(sum(scores) // len(scores)))
+log = "INFO start\nERROR disk full\nWARN slow\nERROR timeout\ninfo ERROR-free\nDEBUG x"
+task("h14", "aggregate", "Read in/app.log. Write out/result.txt containing the integer number of lines that contain the substring \"ERROR\" (case-sensitive).",
+     {"in/app.log": log}, "out/result.txt", "text", str(sum("ERROR" in l for l in lines(log))))
+task("h15", "aggregate", "Read in/text.txt. Write the longest line (by character count; the first one on ties) to out/result.txt.",
+     {"in/text.txt": txt}, "out/result.txt", "text", max(lines(txt), key=len))
+
+# ---- string parsing ----
+kv = "host=db.local\nport=5432\nuser=admin\nmode=read=only"
+task("h16", "parse", "Read in/config.kv (key=value per line; split on the first '=' only). Write out/result.json mapping keys to string values.",
+     {"in/config.kv": kv}, "out/result.json", "json", dict(l.split("=", 1) for l in lines(kv)))
+plog = "INFO|started\nWARN|disk 80%\nERROR|crash\nINFO|retry\nINFO|ok\nERROR|again"
+task("h17", "parse", "Read in/events.log (LEVEL|message per line). Write out/result.json mapping each LEVEL to its integer count.",
+     {"in/events.log": plog}, "out/result.json", "json", dict(Counter(l.split("|")[0] for l in lines(plog))))
+emails = "ada@example.com\nbob@mail.test.org\ncy@x.io"
+task("h18", "parse", "Read in/emails.txt (one address per line). Write the domain part (after '@') of each, one per line, to out/result.txt.",
+     {"in/emails.txt": emails}, "out/result.txt", "text", "\n".join(e.split("@")[1] for e in lines(emails)))
+times = "01:30\n00:45\n02:05\n10:00"
+task("h19", "parse", "Read in/times.txt (HH:MM durations, one per line). Write the integer total number of minutes to out/result.txt.",
+     {"in/times.txt": times}, "out/result.txt", "text", str(sum(int(t.split(":")[0]) * 60 + int(t.split(":")[1]) for t in lines(times))))
+tags = "red;green;red;blue;green;yellow"
+task("h20", "parse", "Read in/tags.txt (one line of ';'-separated tags). Write out/result.json: the JSON array of distinct tags in first-seen order.",
+     {"in/tags.txt": tags}, "out/result.json", "json", list(OrderedDict.fromkeys(tags.split(";"))))
+
+# ---- validation ----
+users = [{"id": 1, "email": "a@x"}, {"id": 2}, {"id": 3, "email": "c@x"}, {"id": 4, "name": "d"}]
+task("h21", "validate", "Read in/users.json (array of objects with integer id). Write out/result.json: the JSON array of ids of objects that have no \"email\" key, in input order.",
+     {"in/users.json": json.dumps(users)}, "out/result.json", "json", [u["id"] for u in users if "email" not in u])
+people = [{"id": "p1", "age": 34}, {"id": "p2", "age": -3}, {"id": "p3", "age": 151}, {"id": "p4", "age": 0}, {"id": "p5", "age": 150}]
+task("h22", "validate", "Read in/people.json (array of {id, age} with integer ages). Write out/result.txt with one line per record \"<id> valid\" if 0 <= age <= 150 else \"<id> invalid\".",
+     {"in/people.json": json.dumps(people)}, "out/result.txt", "text", "\n".join(f"{p['id']} {'valid' if 0 <= p['age'] <= 150 else 'invalid'}" for p in people))
+ids = ["a7", "b2", "c9", "b2", "a7"]
+def firstdup(xs):
+    seen = set()
+    for x in xs:
+        if x in seen: return "duplicate:" + x
+        seen.add(x)
+    return "ok"
+task("h23", "validate", "Read in/ids.json (array of strings). Write out/result.txt: \"ok\" if all are distinct, else \"duplicate:<id>\" for the first id that repeats an earlier one.",
+     {"in/ids.json": json.dumps(ids)}, "out/result.txt", "text", firstdup(ids))
+schema = ["host", "port", "user", "timeout"]
+config = {"host": "h", "user": "u", "debug": True}
+task("h24", "validate", "Read in/schema.json (array of required key names) and in/config.json (object). Write out/result.json: the array of required keys missing from config, in schema order.",
+     {"in/schema.json": json.dumps(schema), "in/config.json": json.dumps(config)}, "out/result.json", "json", [k for k in schema if k not in config])
+seq = "1\n3\n3\n8\n5\n9"
+ss = [int(x) for x in lines(seq)]
+task("h25", "validate", "Read in/seq.txt (one integer per line). Write out/result.txt: \"sorted\" if the sequence is non-decreasing, otherwise \"unsorted at <n>\" where n is the 1-based line of the first value smaller than its predecessor.",
+     {"in/seq.txt": seq}, "out/result.txt", "text", next((f"unsorted at {i+1}" for i in range(1, len(ss)) if ss[i] < ss[i-1]), "sorted"))
+
+# ---- JSON reshaping ----
+items = [{"id": "u1", "name": "Ada"}, {"id": "u2", "name": "Bob"}, {"id": "u3", "name": "Cy"}]
+task("h26", "reshape", "Read in/items.json (array of {id, name}). Write out/result.json: an object mapping each id to its name.",
+     {"in/items.json": json.dumps(items)}, "out/result.json", "json", {i["id"]: i["name"] for i in items})
+prof = {"user": {"name": "Ada", "tags": ["x", "y", "z"], "meta": {"active": True}}}
+task("h27", "reshape", "Read in/profile.json ({user: {name, tags, ...}}). Write out/result.json: {\"name\": user.name, \"tag_count\": number of user.tags}.",
+     {"in/profile.json": json.dumps(prof)}, "out/result.json", "json", {"name": "Ada", "tag_count": 3})
+members = [{"team": "red", "member": "a"}, {"team": "blue", "member": "b"}, {"team": "red", "member": "c"}, {"team": "green", "member": "d"}, {"team": "blue", "member": "e"}]
+grp = {}
+for m in members: grp.setdefault(m["team"], []).append(m["member"])
+task("h28", "reshape", "Read in/members.json (array of {team, member}). Write out/result.json: an object mapping each team to the array of its members in input order.",
+     {"in/members.json": json.dumps(members)}, "out/result.json", "json", grp)
+recs = [{"id": 1, "status": "open", "owner": "a", "prio": 2}, {"id": 2, "status": "done", "owner": "b", "prio": 1}]
+task("h29", "reshape", "Read in/records.json (array of objects). Write out/result.json: the same array keeping only the \"id\" and \"status\" fields of each object.",
+     {"in/records.json": json.dumps(recs)}, "out/result.json", "json", [{"id": r["id"], "status": r["status"]} for r in recs])
+codes = {"US": "United States", "FR": "France", "JP": "Japan"}
+task("h30", "reshape", "Read in/codes.json (object of string to string). Write out/result.json: the inverted object mapping each value to its key.",
+     {"in/codes.json": json.dumps(codes)}, "out/result.json", "json", {v: k for k, v in codes.items()})
+
+# ---- file walking ----
+tree = {"in/docs/a.txt": "one", "in/docs/sub/b.txt": "two", "in/docs/sub/c.md": "three", "in/docs/sub/deep/d.txt": "four", "in/docs/e.log": "five"}
+task("h31", "walk", "Recursively walk in/docs. Write out/result.txt containing the integer number of regular files under it.",
+     tree, "out/result.txt", "text", str(len(tree)))
+txts = sorted(p for p in tree if p.endswith(".txt"))
+task("h32", "walk", "Recursively walk in/docs. Write out/result.txt with the paths (relative to in/docs, '/'-separated) of all files ending in .txt, sorted lexicographically, one per line.",
+     tree, "out/result.txt", "text", "\n".join(p[len("in/docs/"):] for p in txts))
+
+Path(__file__).with_name("manifest.json").write_text(json.dumps({"version": "heldout-v0", "tasks": T}, indent=2) + "\n")
+print(len(T), "tasks")
diff --git a/experiments/token-efficiency-heldout-v0/tasks/manifest.json b/experiments/token-efficiency-heldout-v0/tasks/manifest.json
new file mode 100644
index 0000000..21504dd
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/tasks/manifest.json
@@ -0,0 +1,468 @@
+{
+  "version": "heldout-v0",
+  "tasks": [
+    {
+      "id": "h01",
+      "category": "text",
+      "description": "Read in/text.txt. Write out/result.txt with every line upper-cased, same line order.",
+      "inputs": {
+        "in/text.txt": "alpha beta\nGamma delta\n\nfoo bar foo\n  padded line  \nlast"
+      },
+      "output": "out/result.txt",
+      "kind": "text",
+      "expected": "ALPHA BETA\nGAMMA DELTA\n\nFOO BAR FOO\n  PADDED LINE  \nLAST"
+    },
+    {
+      "id": "h02",
+      "category": "text",
+      "description": "Read in/text.txt. Write out/result.txt with the lines in reverse order.",
+      "inputs": {
+        "in/text.txt": "alpha beta\nGamma delta\n\nfoo bar foo\n  padded line  \nlast"
+      },
+      "output": "out/result.txt",
+      "kind": "text",
+      "expected": "last\n  padded line  \nfoo bar foo\n\nGamma delta\nalpha beta"
+    },
+    {
+      "id": "h03",
+      "category": "text",
+      "description": "Read in/text.txt. Write out/result.txt where each line is prefixed with its 1-based line number and \": \" (e.g. \"1: alpha beta\").",
+      "inputs": {
+        "in/text.txt": "alpha beta\nGamma delta\n\nfoo bar foo\n  padded line  \nlast"
+      },
+      "output": "out/result.txt",
+      "kind": "text",
+      "expected": "1: alpha beta\n2: Gamma delta\n3: \n4: foo bar foo\n5:   padded line  \n6: last"
+    },
+    {
+      "id": "h04",
+      "category": "text",
+      "description": "Read in/text.txt. Trim leading/trailing spaces of every line, drop lines that are then empty, and write the rest to out/result.txt.",
+      "inputs": {
+        "in/text.txt": "alpha beta\nGamma delta\n\nfoo bar foo\n  padded line  \nlast"
+      },
+      "output": "out/result.txt",
+      "kind": "text",
+      "expected": "alpha beta\nGamma delta\nfoo bar foo\npadded line\nlast"
+    },
+    {
+      "id": "h05",
+      "category": "text",
+      "description": "Read in/text.txt. Replace every occurrence of the substring \"foo\" with \"bar\" and write the result to out/result.txt.",
+      "inputs": {
+        "in/text.txt": "alpha beta\nGamma delta\n\nfoo bar foo\n  padded line  \nlast"
+      },
+      "output": "out/result.txt",
+      "kind": "text",
+      "expected": "alpha beta\nGamma delta\n\nbar bar bar\n  padded line  \nlast"
+    },
+    {
+      "id": "h06",
+      "category": "csv",
+      "description": "Read in/data.csv (header row, comma-separated, no quoting). Write the email column values, one per line in row order, to out/result.txt.",
+      "inputs": {
+        "in/data.csv": "name,email,age,dept,amount\nAda,ada@example.com,36,eng,120\nBob,bob@test.org,25,ops,75\nCy,cy@example.com,30,eng,40\nDee,dee@corp.net,41,sales,300\nEd,ed@test.org,29,ops,15"
+      },
+      "output": "out/result.txt",
+      "kind": "text",
+      "expected": "ada@example.com\nbob@test.org\ncy@example.com\ndee@corp.net\ned@test.org"
+    },
+    {
+      "id": "h07",
+      "category": "csv",
+      "description": "Read in/data.csv. Write out/result.csv containing the header row followed by only the rows whose integer age is >= 30, in input order.",
+      "inputs": {
+        "in/data.csv": "name,email,age,dept,amount\nAda,ada@example.com,36,eng,120\nBob,bob@test.org,25,ops,75\nCy,cy@example.com,30,eng,40\nDee,dee@corp.net,41,sales,300\nEd,ed@test.org,29,ops,15"
+      },
+      "output": "out/result.csv",
+      "kind": "text",
+      "expected": "name,email,age,dept,amount\nAda,ada@example.com,36,eng,120\nCy,cy@example.com,30,eng,40\nDee,dee@corp.net,41,sales,300"
+    },
+    {
+      "id": "h08",
+      "category": "csv",
+      "description": "Read in/data.csv. Write the integer sum of the amount column to out/result.txt.",
+      "inputs": {
+        "in/data.csv": "name,email,age,dept,amount\nAda,ada@example.com,36,eng,120\nBob,bob@test.org,25,ops,75\nCy,cy@example.com,30,eng,40\nDee,dee@corp.net,41,sales,300\nEd,ed@test.org,29,ops,15"
+      },
+      "output": "out/result.txt",
+      "kind": "text",
+      "expected": "550"
+    },
+    {
+      "id": "h09",
+      "category": "csv",
+      "description": "Read in/data.csv. Write out/result.json: a JSON array with one object per row mapping each header name to the row's string value.",
+      "inputs": {
+        "in/data.csv": "name,email,age,dept,amount\nAda,ada@example.com,36,eng,120\nBob,bob@test.org,25,ops,75\nCy,cy@example.com,30,eng,40\nDee,dee@corp.net,41,sales,300\nEd,ed@test.org,29,ops,15"
+      },
+      "output": "out/result.json",
+      "kind": "json",
+      "expected": [
+        {
+          "name": "Ada",
+          "email": "ada@example.com",
+          "age": "36",
+          "dept": "eng",
+          "amount": "120"
+        },
+        {
+          "name": "Bob",
+          "email": "bob@test.org",
+          "age": "25",
+          "dept": "ops",
+          "amount": "75"
+        },
+        {
+          "name": "Cy",
+          "email": "cy@example.com",
+          "age": "30",
+          "dept": "eng",
+          "amount": "40"
+        },
+        {
+          "name": "Dee",
+          "email": "dee@corp.net",
+          "age": "41",
+          "dept": "sales",
+          "amount": "300"
+        },
+        {
+          "name": "Ed",
+          "email": "ed@test.org",
+          "age": "29",
+          "dept": "ops",
+          "amount": "15"
+        }
+      ]
+    },
+    {
+      "id": "h10",
+      "category": "csv",
+      "description": "Read in/data.csv. Write out/result.json: a JSON object mapping each dept value to the integer number of rows with that dept.",
+      "inputs": {
+        "in/data.csv": "name,email,age,dept,amount\nAda,ada@example.com,36,eng,120\nBob,bob@test.org,25,ops,75\nCy,cy@example.com,30,eng,40\nDee,dee@corp.net,41,sales,300\nEd,ed@test.org,29,ops,15"
+      },
+      "output": "out/result.json",
+      "kind": "json",
+      "expected": {
+        "eng": 2,
+        "ops": 2,
+        "sales": 1
+      }
+    },
+    {
+      "id": "h11",
+      "category": "aggregate",
+      "description": "Read in/numbers.txt (one integer per line). Write out/result.json: {\"min\": m, \"max\": M, \"sum\": s} as integers.",
+      "inputs": {
+        "in/numbers.txt": "12\n-4\n7\n30\n0\n9"
+      },
+      "output": "out/result.json",
+      "kind": "json",
+      "expected": {
+        "min": -4,
+        "max": 30,
+        "sum": 54
+      }
+    },
+    {
+      "id": "h12",
+      "category": "aggregate",
+      "description": "Read in/words.txt. Lower-case it, split every line on single spaces, and write out/result.json mapping each word to its integer count.",
+      "inputs": {
+        "in/words.txt": "the cat and the hat\nThe bat and THE cat"
+      },
+      "output": "out/result.json",
+      "kind": "json",
+      "expected": {
+        "the": 4,
+        "cat": 2,
+        "and": 2,
+        "hat": 1,
+        "bat": 1
+      }
+    },
+    {
+      "id": "h13",
+      "category": "aggregate",
+      "description": "Read in/scores.json (a JSON array of integers). Write out/result.txt containing the integer mean rounded down (floor).",
+      "inputs": {
+        "in/scores.json": "[88, 92, 79, 65, 100, 71]"
+      },
+      "output": "out/result.txt",
+      "kind": "text",
+      "expected": "82"
+    },
+    {
+      "id": "h14",
+      "category": "aggregate",
+      "description": "Read in/app.log. Write out/result.txt containing the integer number of lines that contain the substring \"ERROR\" (case-sensitive).",
+      "inputs": {
+        "in/app.log": "INFO start\nERROR disk full\nWARN slow\nERROR timeout\ninfo ERROR-free\nDEBUG x"
+      },
+      "output": "out/result.txt",
+      "kind": "text",
+      "expected": "3"
+    },
+    {
+      "id": "h15",
+      "category": "aggregate",
+      "description": "Read in/text.txt. Write the longest line (by character count; the first one on ties) to out/result.txt.",
+      "inputs": {
+        "in/text.txt": "alpha beta\nGamma delta\n\nfoo bar foo\n  padded line  \nlast"
+      },
+      "output": "out/result.txt",
+      "kind": "text",
+      "expected": "  padded line  "
+    },
+    {
+      "id": "h16",
+      "category": "parse",
+      "description": "Read in/config.kv (key=value per line; split on the first '=' only). Write out/result.json mapping keys to string values.",
+      "inputs": {
+        "in/config.kv": "host=db.local\nport=5432\nuser=admin\nmode=read=only"
+      },
+      "output": "out/result.json",
+      "kind": "json",
+      "expected": {
+        "host": "db.local",
+        "port": "5432",
+        "user": "admin",
+        "mode": "read=only"
+      }
+    },
+    {
+      "id": "h17",
+      "category": "parse",
+      "description": "Read in/events.log (LEVEL|message per line). Write out/result.json mapping each LEVEL to its integer count.",
+      "inputs": {
+        "in/events.log": "INFO|started\nWARN|disk 80%\nERROR|crash\nINFO|retry\nINFO|ok\nERROR|again"
+      },
+      "output": "out/result.json",
+      "kind": "json",
+      "expected": {
+        "INFO": 3,
+        "WARN": 1,
+        "ERROR": 2
+      }
+    },
+    {
+      "id": "h18",
+      "category": "parse",
+      "description": "Read in/emails.txt (one address per line). Write the domain part (after '@') of each, one per line, to out/result.txt.",
+      "inputs": {
+        "in/emails.txt": "ada@example.com\nbob@mail.test.org\ncy@x.io"
+      },
+      "output": "out/result.txt",
+      "kind": "text",
+      "expected": "example.com\nmail.test.org\nx.io"
+    },
+    {
+      "id": "h19",
+      "category": "parse",
+      "description": "Read in/times.txt (HH:MM durations, one per line). Write the integer total number of minutes to out/result.txt.",
+      "inputs": {
+        "in/times.txt": "01:30\n00:45\n02:05\n10:00"
+      },
+      "output": "out/result.txt",
+      "kind": "text",
+      "expected": "860"
+    },
+    {
+      "id": "h20",
+      "category": "parse",
+      "description": "Read in/tags.txt (one line of ';'-separated tags). Write out/result.json: the JSON array of distinct tags in first-seen order.",
+      "inputs": {
+        "in/tags.txt": "red;green;red;blue;green;yellow"
+      },
+      "output": "out/result.json",
+      "kind": "json",
+      "expected": [
+        "red",
+        "green",
+        "blue",
+        "yellow"
+      ]
+    },
+    {
+      "id": "h21",
+      "category": "validate",
+      "description": "Read in/users.json (array of objects with integer id). Write out/result.json: the JSON array of ids of objects that have no \"email\" key, in input order.",
+      "inputs": {
+        "in/users.json": "[{\"id\": 1, \"email\": \"a@x\"}, {\"id\": 2}, {\"id\": 3, \"email\": \"c@x\"}, {\"id\": 4, \"name\": \"d\"}]"
+      },
+      "output": "out/result.json",
+      "kind": "json",
+      "expected": [
+        2,
+        4
+      ]
+    },
+    {
+      "id": "h22",
+      "category": "validate",
+      "description": "Read in/people.json (array of {id, age} with integer ages). Write out/result.txt with one line per record \"<id> valid\" if 0 <= age <= 150 else \"<id> invalid\".",
+      "inputs": {
+        "in/people.json": "[{\"id\": \"p1\", \"age\": 34}, {\"id\": \"p2\", \"age\": -3}, {\"id\": \"p3\", \"age\": 151}, {\"id\": \"p4\", \"age\": 0}, {\"id\": \"p5\", \"age\": 150}]"
+      },
+      "output": "out/result.txt",
+      "kind": "text",
+      "expected": "p1 valid\np2 invalid\np3 invalid\np4 valid\np5 valid"
+    },
+    {
+      "id": "h23",
+      "category": "validate",
+      "description": "Read in/ids.json (array of strings). Write out/result.txt: \"ok\" if all are distinct, else \"duplicate:<id>\" for the first id that repeats an earlier one.",
+      "inputs": {
+        "in/ids.json": "[\"a7\", \"b2\", \"c9\", \"b2\", \"a7\"]"
+      },
+      "output": "out/result.txt",
+      "kind": "text",
+      "expected": "duplicate:b2"
+    },
+    {
+      "id": "h24",
+      "category": "validate",
+      "description": "Read in/schema.json (array of required key names) and in/config.json (object). Write out/result.json: the array of required keys missing from config, in schema order.",
+      "inputs": {
+        "in/schema.json": "[\"host\", \"port\", \"user\", \"timeout\"]",
+        "in/config.json": "{\"host\": \"h\", \"user\": \"u\", \"debug\": true}"
+      },
+      "output": "out/result.json",
+      "kind": "json",
+      "expected": [
+        "port",
+        "timeout"
+      ]
+    },
+    {
+      "id": "h25",
+      "category": "validate",
+      "description": "Read in/seq.txt (one integer per line). Write out/result.txt: \"sorted\" if the sequence is non-decreasing, otherwise \"unsorted at <n>\" where n is the 1-based line of the first value smaller than its predecessor.",
+      "inputs": {
+        "in/seq.txt": "1\n3\n3\n8\n5\n9"
+      },
+      "output": "out/result.txt",
+      "kind": "text",
+      "expected": "unsorted at 5"
+    },
+    {
+      "id": "h26",
+      "category": "reshape",
+      "description": "Read in/items.json (array of {id, name}). Write out/result.json: an object mapping each id to its name.",
+      "inputs": {
+        "in/items.json": "[{\"id\": \"u1\", \"name\": \"Ada\"}, {\"id\": \"u2\", \"name\": \"Bob\"}, {\"id\": \"u3\", \"name\": \"Cy\"}]"
+      },
+      "output": "out/result.json",
+      "kind": "json",
+      "expected": {
+        "u1": "Ada",
+        "u2": "Bob",
+        "u3": "Cy"
+      }
+    },
+    {
+      "id": "h27",
+      "category": "reshape",
+      "description": "Read in/profile.json ({user: {name, tags, ...}}). Write out/result.json: {\"name\": user.name, \"tag_count\": number of user.tags}.",
+      "inputs": {
+        "in/profile.json": "{\"user\": {\"name\": \"Ada\", \"tags\": [\"x\", \"y\", \"z\"], \"meta\": {\"active\": true}}}"
+      },
+      "output": "out/result.json",
+      "kind": "json",
+      "expected": {
+        "name": "Ada",
+        "tag_count": 3
+      }
+    },
+    {
+      "id": "h28",
+      "category": "reshape",
+      "description": "Read in/members.json (array of {team, member}). Write out/result.json: an object mapping each team to the array of its members in input order.",
+      "inputs": {
+        "in/members.json": "[{\"team\": \"red\", \"member\": \"a\"}, {\"team\": \"blue\", \"member\": \"b\"}, {\"team\": \"red\", \"member\": \"c\"}, {\"team\": \"green\", \"member\": \"d\"}, {\"team\": \"blue\", \"member\": \"e\"}]"
+      },
+      "output": "out/result.json",
+      "kind": "json",
+      "expected": {
+        "red": [
+          "a",
+          "c"
+        ],
+        "blue": [
+          "b",
+          "e"
+        ],
+        "green": [
+          "d"
+        ]
+      }
+    },
+    {
+      "id": "h29",
+      "category": "reshape",
+      "description": "Read in/records.json (array of objects). Write out/result.json: the same array keeping only the \"id\" and \"status\" fields of each object.",
+      "inputs": {
+        "in/records.json": "[{\"id\": 1, \"status\": \"open\", \"owner\": \"a\", \"prio\": 2}, {\"id\": 2, \"status\": \"done\", \"owner\": \"b\", \"prio\": 1}]"
+      },
+      "output": "out/result.json",
+      "kind": "json",
+      "expected": [
+        {
+          "id": 1,
+          "status": "open"
+        },
+        {
+          "id": 2,
+          "status": "done"
+        }
+      ]
+    },
+    {
+      "id": "h30",
+      "category": "reshape",
+      "description": "Read in/codes.json (object of string to string). Write out/result.json: the inverted object mapping each value to its key.",
+      "inputs": {
+        "in/codes.json": "{\"US\": \"United States\", \"FR\": \"France\", \"JP\": \"Japan\"}"
+      },
+      "output": "out/result.json",
+      "kind": "json",
+      "expected": {
+        "United States": "US",
+        "France": "FR",
+        "Japan": "JP"
+      }
+    },
+    {
+      "id": "h31",
+      "category": "walk",
+      "description": "Recursively walk in/docs. Write out/result.txt containing the integer number of regular files under it.",
+      "inputs": {
+        "in/docs/a.txt": "one",
+        "in/docs/sub/b.txt": "two",
+        "in/docs/sub/c.md": "three",
+        "in/docs/sub/deep/d.txt": "four",
+        "in/docs/e.log": "five"
+      },
+      "output": "out/result.txt",
+      "kind": "text",
+      "expected": "5"
+    },
+    {
+      "id": "h32",
+      "category": "walk",
+      "description": "Recursively walk in/docs. Write out/result.txt with the paths (relative to in/docs, '/'-separated) of all files ending in .txt, sorted lexicographically, one per line.",
+      "inputs": {
+        "in/docs/a.txt": "one",
+        "in/docs/sub/b.txt": "two",
+        "in/docs/sub/c.md": "three",
+        "in/docs/sub/deep/d.txt": "four",
+        "in/docs/e.log": "five"
+      },
+      "output": "out/result.txt",
+      "kind": "text",
+      "expected": "a.txt\nsub/b.txt\nsub/deep/d.txt"
+    }
+  ]
+}
diff --git a/experiments/token-efficiency-heldout-v0/verification_test.go b/experiments/token-efficiency-heldout-v0/verification_test.go
new file mode 100644
index 0000000..a2a4eda
--- /dev/null
+++ b/experiments/token-efficiency-heldout-v0/verification_test.go
@@ -0,0 +1,102 @@
+package heldout
+
+import (
+	"crypto/sha256"
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+var exts = map[string]string{"howl": "howl", "python": "py", "js": "js", "go": "go"}
+
+func readJSON(t *testing.T, path string, v any) {
+	t.Helper()
+	data, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := json.Unmarshal(data, v); err != nil {
+		t.Fatal(err)
+	}
+}
+
+// Every task × language is either verified (pass, matching hash) or an
+// explicitly recorded HowlFrame gap with a reason and no solution file.
+func TestHeldoutVerificationCoverage(t *testing.T) {
+	var manifest struct {
+		Tasks []struct {
+			ID string `json:"id"`
+		} `json:"tasks"`
+	}
+	readJSON(t, "tasks/manifest.json", &manifest)
+	var gaps []struct{ Task, Lang, Reason string }
+	readJSON(t, "gaps.json", &gaps)
+	var results []struct {
+		Task, Lang, SHA256 string
+		Pass               bool
+	}
+	readJSON(t, "results/verification.json", &results)
+	if len(manifest.Tasks) < 30 {
+		t.Fatalf("held-out corpus has %d tasks, want >= 30", len(manifest.Tasks))
+	}
+	gapped := map[string]bool{}
+	for _, g := range gaps {
+		if g.Lang != "howl" || g.Reason == "" {
+			t.Errorf("gap %s/%s must be a HowlFrame gap with a reason", g.Task, g.Lang)
+		}
+		gapped[g.Task+"/"+g.Lang] = true
+		if _, err := os.Stat(filepath.Join("solutions", g.Lang, g.Task+"."+exts[g.Lang])); err == nil {
+			t.Errorf("gap %s/%s has a solution file", g.Task, g.Lang)
+		}
+	}
+	seen := map[string]bool{}
+	for _, r := range results {
+		key := r.Task + "/" + r.Lang
+		if seen[key] || gapped[key] {
+			t.Errorf("duplicate or gapped entry %s", key)
+		}
+		seen[key] = true
+		if !r.Pass {
+			t.Errorf("%s did not pass", key)
+		}
+		src, err := os.ReadFile(filepath.Join("solutions", r.Lang, r.Task+"."+exts[r.Lang]))
+		if err != nil {
+			t.Errorf("%s: %v", key, err)
+			continue
+		}
+		if got := fmt.Sprintf("%x", sha256.Sum256(src)); got != r.SHA256 {
+			t.Errorf("%s: stale hash", key)
+		}
+	}
+	for _, task := range manifest.Tasks {
+		for lang := range exts {
+			key := task.ID + "/" + lang
+			if !seen[key] && !gapped[key] {
+				t.Errorf("missing %s", key)
+			}
+		}
+	}
+	if want := len(manifest.Tasks)*len(exts) - len(gaps); len(results) != want {
+		t.Errorf("%d results, want %d", len(results), want)
+	}
+}
+
+// The frozen spec files must match the hashes recorded at freeze time.
+func TestHeldoutSpecsFrozen(t *testing.T) {
+	frozen := map[string]string{
+		"tasks/README.md":      "4441e1e5c4e906a55fe595c382a7d2ef784045e9c1b1230f38887d70bf8318f1",
+		"tasks/build_specs.py": "dcea7d494d8b146ae032456b1f8822364d488c3d8596f7d829aecb21a75148f8",
+		"tasks/manifest.json":  "962fa8198b11ac3ff1066abf35be6dcb87ad8a852574edfc46fee17a0ab9e67e",
+	}
+	for path, want := range frozen {
+		data, err := os.ReadFile(path)
+		if err != nil {
+			t.Fatal(err)
+		}
+		if got := fmt.Sprintf("%x", sha256.Sum256(data)); got != want {
+			t.Errorf("%s changed after freeze: %s", path, got)
+		}
+	}
+}