[SamGu-NRX]

What changed

This PR adds FridgeLuck's rights-approved recipe corpus: data/recipes/corpus.jsonl (401 recipes in the RS v1 schema from data/recipes/SCHEMA.md), its rights manifest, the ingestion and validation tools under data/recipes/tools/, and data/recipes/REPORT.md. All 401 are originals written for FridgeLuck by fl-original-recipes-f1. Ingestion validates every recipe with suite C's rs.py, rights.py and bands.py, joins the accepted substitutions, and runs four gates of its own that the reviews asked for. Recipes that fail a gate wait in data/recipes/quarantine.jsonl with the reason.

The review at 4b6e163 blocked on five findings, and the re-review at 0f39ea8 found four gaps in the first fixes. Each is fixed in the data and enforced by validate_corpus.py, with planted failures in data/recipes/tools/test_gates.py that the 4b6e163 validator accepted.

Doneness endpoints follow the FSIS chart (finding 1; re-review 3)

food_safety.py encodes the USDA FSIS Safe Minimum Internal Temperature Chart, fetched 2026-10-11 (page last updated Apr 14, 2025). The rows it encodes are quoted in FSIS_ROWS, and each class pattern cites its row. The gate refuses a recipe when any food misses its endpoint:

• Casseroles, stuffing, reheated leftovers, precooked meat and ham must reach 165 °F (74 °C).
• Ground poultry is 165 °F, ground meat 160 °F, and fish and shellfish 145 °F.
• Whole cuts of beef, pork, veal and lamb are 145 °F (63 °C), with a 3-minute rest stated in the same step.
• Eggs are 160 °F, with yolk and white firm.
• Foods cooked or mixed together reach the strictest of their minimums. Raw egg folded into fish makes the cakes an egg dish.
• Every reheating step carries its own 165 °F check, whatever was checked before it.
• An animal food with no row in the chart refuses the recipe rather than getting a guessed endpoint. Venison, rabbit and organ meats other than poultry giblets fall here.

fix_doneness.py applies the corrections, each asserted against the exact text it replaces; 89 steps in 74 originals changed. Seven casseroles and bakes went to 165 °F, including the breakfast casserole (fl-original-0008, was 160 °F) and the tuna gratin (0102, was a 145 °F fish check). Twenty whole-cut checks gained a rest, including the citrus pork (0094). The cod cakes (0343) had relied on the 145 °F poach after raw egg went in, and now cook to 160 °F. A flank steak (0023) was pulled at 140 °F on carryover, and the chana masala (0152) served a jammy yolk under a 160 °F reading of the white.

Checks are instructions in the step text (finding 2; re-review 1)

A check that lives only in step.safety never reaches the cook. Every step with a check now tells the cook, in one clause, to read the endpoint on a thermometer in the food, for example "Reheat leftovers until a food thermometer reads 165 °F (74 °C)." Keywords are not enough. "A thermometer reading 165 F is unnecessary" fails because the reading is negated, an oven setting does not count as a reading, and "do not rest 3 minutes" is not a rest. No step judges doneness by look or states an internal temperature without a check. Six checks that sat on browning or garnish steps moved to the step where the food finishes cooking. convert_bundled.py (the reviewer's lines 324-326) now writes the check into the seed step text too.

Rights rest on pinned, affirmative evidence (finding 3; re-review 4)

The 25 bundled seeds cited only absences: no Source line, present since the initial commit, title not in the BBC scrape cache. None of that shows who wrote them or that anyone gave permission, and SPEC 3.5 says these 25 are not automatically clear. They moved to quarantine with their old records.

rights_evidence.py no longer trusts how a citation is spelled. pin_rights_evidence.py writes data/recipes/evidence/rights.json from git: every authoring record at overhaul/original-recipes 607ef99, with its recipe's title and a hash of its ingredient names, plus the terms each external source was checked against, with page hashes and quoted sentences. An original clears only when its record exists at the cited commit and carries the licence, the app's permissions and this recipe's identity. An external source clears only with a pinned "verified" verdict, and a test re-derives the evidence file from git.

Checking the NHS terms changed their verdict. The two NHS Healthier Families recipes link their own Terms of Website Use, which say "All such rights are reserved" and "You must not use any part of the content on the Site for commercial purposes without obtaining a licence". The records cited the general nhs.uk OGL release instead. With the two in conflict the permissions are unknown, so both NHS recipes moved to quarantine.

Claims need known composition (finding 4; re-review 2)

The scanner pinned at 26c7004 skipped ingredient names its alias table did not know, so fl-original-0065 claimed dairy-free with "ghee or vegetable oil" in it. composition.py refuses a claim in four cases:

• An ingredient's composition is unknown to suite C's allergens.json.
• A line offers alternatives that are not all known and allowed.
• A known ingredient varies by product: stock (meat, milk, wheat), cheese (rennet), tamari, curry powder and miso (wheat or barley), pasta (egg), and bread (milk, egg, lard).
• A step names a food the claim forbids. "Peanut butter" is exempt from the dairy-free scan only, so it still fails nut-free. Compound foods such as pesto, mayonnaise and Worcestershire block every claim they contradict.

strip_unproven_claims.py removed 916 of 981 claims and never adds one. Most come back only when alias rows or the FLIngredientFacts package cover the 362 recipes with an unresolved name. (The 0f39ea8 commit message says 342; the measured count is 362.)

The NHS conversion uses the same rules (finding 5)

The NHS conversion rewrote every "no longer pink" to 160 °F, including ground poultry. Any earlier check suppressed the reheating check, the mixed-dish fallback chose egg over poultry, and pork got no rest. It also wrote classes the schema does not have (meat_pork, meat_red, seafood). ingest_nhs.py build_steps now takes each endpoint from the recipe's ingredients through food_safety.py. It checks every reheating step at 165 °F, closes with a check naming any food still unchecked, and skips a page that fails the gate. NHS and NHLBI claims pass the composition gate, and ingest_originals.py stops a recipe that fails either gate.

Why

SPEC 3.1 makes thermometer doneness from FSIS a release blocker, and SPEC 3.5 says unknown rights means quarantine and claims must be checked. Suite C's linter at 26c7004 passed every case above, so the corpus needs its own gates; the operator ruled that nothing under evaluation/ changes. The counts show the cost: 27 fewer recipes and 916 fewer claims, because neither could be proven.

Proof

Run at head 7fc68d48e35f14cce3df1a08c37657f18b738d6a under the long-run heavy lease (lr-lease acquire --kind heavy, lockf -k ~/.long-run/locks/heavy.lock). Suite C was extracted read-only from 26c7004 with git archive 26c7004 evaluation/suites/recipes | tar -x -C /tmp/fl113-suite-26c.

Unit tests (54 planted-failure and corpus tests):

$ python3 -I -m unittest discover -s data/recipes/tools -p 'test_*.py'
Ran 54 tests in 0.253s
OK

Validator (the list of 137 uncovered conversion pairs is omitted):

$ python3 -I data/recipes/tools/validate_corpus.py --suite-dir /tmp/fl113-suite-26c/evaluation/suites/recipes
corpus: 401 recipes, 2829 ingredient lines
rights: 401 rights-clear, 0 quarantined-in-corpus (must be 0)
accepted: 401 recipes in corpus.jsonl; quarantined: 27 in quarantine.jsonl (rights 27)
release gates: 0 violation(s) (rights evidence, original-text overlap, FSIS endpoints, thermometer in step text, claim composition)
claims shipped: 65
linter: 401/401 recipes pass; quantity evidence: 2612/2829 lines resolved (92.33%)
claims: 0 unchecked, 0 stripped by the linter
pass A (suite current semantics): 2612/2829 lines resolved (92.33%)
pass B (corpus read path):        2612/2829 lines resolved (92.33%)
  bands represented: 14/14
  cross-product cells filled: 75/90 (informational)
VALIDATION GREEN: rights-clear corpus lints clean, claims verified, ids unique and in-scheme, no quarantined recipe in corpus.jsonl
measured: recipes=401 resolved_pct=92.33 bands=14/14
exit=0

Every correction tool is a no-op on the committed files, and the pinned evidence re-derives from git:

$ python3 -I data/recipes/tools/fix_doneness.py
96 edits; 0 recipe(s) changed
food-safety rules: 0 recipe(s) still failing
$ python3 -I data/recipes/tools/strip_unproven_claims.py --suite-dir /tmp/fl113-suite-26c/evaluation/suites/recipes
claims: 65 before, 65 kept, 0 removed
$ python3 -I data/recipes/tools/quarantine.py --rights
0 recipe(s) fail the rights gate; 401 stay in the corpus
$ python3 -I data/recipes/tools/pin_rights_evidence.py --check
evidence matches git

Counts before and after, with today's gates applied to both trees. The script reads each tree's files with git show and runs the current tools; a recipe can fail several gates.

ref 4b6e163: accepted 428 in corpus.jsonl, quarantined 0 in quarantine.jsonl (none)
  claims shipped: 1050
  recipes the current gates refuse: 416
    claims: composition not proven: 400 recipe(s)
    rights: no affirmative evidence: 27 recipe(s)
    safety: FSIS endpoint: 36 recipe(s)
    safety: check not in step text: 70 recipe(s)
ref HEAD: accepted 401 in corpus.jsonl, quarantined 27 in quarantine.jsonl ({'rights': 27})
  claims shipped: 65
  recipes the current gates refuse: 0

scripts/recipe_scraper/ has no tests and this PR does not change it; python3 -I -m py_compile scripts/recipe_scraper/scrape_recipes.py exits 0, and that is the only check run on it. The repository's CI runs the iOS build and tests, which do not exercise these Python gates. Nothing here needs Xcode, and none was run.

Risk

• Rights bar margin. 401 rights-clear recipes is one over the 400 bar. Two more quarantines drop below it.
• Resolution bar. Quantity evidence still fails at 2612/2829 = 92.33%. The bar needs 2688, so the corpus is short 76 lines; it was short 68 at 4b6e163, and the quarantined recipes took 159 resolved lines with them. The 217 unresolved lines state no amount anywhere in their records (REPORT section 9). Closing the gap is the operator's open decision.
• The gate is a floor. It classifies by ingredient name and reads which foods a step names. Chorizo is always treated as raw (160 °F). Bacon is a whole pork cut at 145 °F with rest, because FSIS lists no bacon row. Stuffed vegetables and traybakes are not casseroles.
• Suite C at 9e04af4. The newer suite fails 51 recipes on its R3 rule (measured at e311653, before the NHS quarantine), and none is a missed endpoint. Its keywords classify chicken stock, canned tuna, fish sauce and burger buns as raw proteins. It also wants a check naming each class where this gate accepts a stricter check on the same food. REPORT section 10 lists the 55 violations by cause, and the validator still pins 26c7004.
• Upstream copy. The 89 safety-step corrections live in the corpus. data/recipes/originals/ belongs to fl-original-recipes-f1, so a re-ingest from there must run fix_doneness.py until the same edits land upstream.

Handoff

• Done. Finding 1 in fed1a05, rests moved into the check step in e311653, soft yolks in eb93e03, and chart-row classes in 976ab51. Finding 2 in 1e10b85, with affirmative readings in 7e946b2 and its test fix in 37a4e90. Finding 3 in 7733b10, with pinned evidence and the NHS quarantine in 1dfe8bb. Finding 4 in 0f39ea8, with per-claim exemptions in c3ce34d. Finding 5 in 79b36ee. REPORT in 68ad837 and 7fc68d4.
• Not done. The 95% resolution bar is the operator's decision. The 25 seeds stay quarantined until someone produces an authorship or permission record. The 2 NHS recipes stay quarantined until someone with authority settles which NHS terms govern Healthier Families pages. Claims return only as composition data grows. The first review's lesser findings are open: NaN, null and duplicate rows in the conversions loader, the 95% bar printed but not enforced, and the substitution context checks.
• Head: 7fc68d48e35f14cce3df1a08c37657f18b738d6a

Checklist

• The scope is focused; related changes belong together
• I explained what changed and why
• N/A before/after images: data and tooling only, no visible UI change
• N/A video: no motion or interaction change

🔗 Obvious Project · 🧵 Obvious Thread · 🔀 Obvious PR Artifact

Writer ruling (2026-10-11, operator). fl-original-recipes-f1 now owns PR #107, overhaul/original-recipes, and data/recipes/originals/; this programme has stopped writing that branch and directory. This programme's commits on overhaul/original-recipes — all authored Sam Gu 127461594+SamGu-NRX@users.noreply.github.com with the obvious-autobuild[bot] co-author trailer, diffs confined to data/recipes/originals/, append-only — were: Part 5 (thread th_MzvPCFVg) 654afe2, 47e9274, 9b69c34, 0b7a08b; Part 5b continuation M4-M11 (thread th_lm6J9eSh) 1855545, 0fc9b63, f484ebb, 667b759, 9bf12e5, dc3e6da, f04f508, 607ef99. Branch head at handoff: 607ef99, still this programme's last commit — the new owner has not yet appended. The operator's 19:10Z instruction to fold #111 into #107 is voided by the same ruling and was never acted on here. The food-safety rulebook (#97) belongs to fl-root-food-safety-f1. This programme's scope stands: the rights audit, PR #43, licensed ingestion into corpus.jsonl (ingesting data/recipes/originals through the same validation as any source), and substitution table (#114). Open with the operator: the 68-line quantity-gap decision (upstream re-authoring vs accept 92.74% as shipped); any data/recipes/originals/ re-authoring routes through fl-original-recipes-f1.

🤖 Generated with Claude Code

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

---

_{Comment @coderabbitai help to get the list of available commands.}

[SamGu-NRX]

Verdict: BLOCK

Reviewed head 27613a1 against SPEC section 3 and task fl-recipe-platform-f1. The latest join contains 428 recipes, not the 27 still described in the Proof and REPORT.

Blockers:

• data/recipes/corpus.jsonl:35,121,129, joined fl-original-0008, fl-original-0094 and fl-original-0102. The breakfast casserole ends at 160 F instead of the casserole requirement of 165 F; the whole-pork recipe gives 145 F without the required three-minute rest; the tuna gratin has a 145 F fish check attached to a bubbling-cheese instruction instead of a 165 F casserole check. ingest_originals.py:158-176 accepts these through the external suite. Quarantine or correct them and add controls for food class, mixed dishes and rest duration. Source: https://www.foodsafety.gov/food-safety-charts/safe-minimum-internal-temperatures
• data/recipes/tools/convert_bundled.py:386-401. All 25 seeds receive original-work permissions because they lack a Source line, existed in an initial commit and are absent from the BBC cache. Those observations do not establish authorship or permission to redistribute. Obtain affirmative provenance or quarantine them. This is an evidence gap, not a finding that they were copied.
• data/recipes/tools/ingest_nhs.py:671-732. Generic meat gets 160 F even if it is ground poultry. Any earlier safety annotation suppresses all fallback checks, including reheating; when no annotation exists, eggs take priority over chicken or seafood. A chicken-and-egg dish can therefore end at 160 F, and leftovers can lose their 165 F check. Check every applicable food and process, then require the strongest applicable safe endpoint and rest.
• data/recipes/tools/convert_bundled.py:324-326 and corpus.jsonl:1,9. Safety metadata does not rewrite the instruction the cook reads. Fried rice still says to scramble eggs and stir-fry rice for 2-3 minutes; chicken rice still says simmer 18-20 minutes. Require explicit thermometer wording in the rendered instruction or a verified consumer that presents the check before completion.

Lesser findings:

• apply_substitutions.py:113-128 labels swaps verified for any supported role without resolving the ingredient's role in this recipe. This is table association, not a recipe-context validation or an eater-allergen check. Keep that distinction in the schema and handoff.
• validate_corpus.py:123-153 prints the quantity percentage but never fails below 95%. The original join can now bring in unknown quantities. Enforce the acceptance threshold and bind the external suite and table to exact source SHAs.
• convert_bundled.py:291 flattens optional and required ingredients into the same bill. Preserve optionality so fit and nutrition do not claim a complete required bill incorrectly.

Criteria: rights evidence partly met; deterministic structured ingestion met; no invented mass partly met; nutrition/USDA identity missing; safe atomic steps partly met; context-specific substitutions partly met; dedupe partly met; report and stratified audit missing; scope/no workflow edits met. 2/9 met.

CI: the old head's iOS run 38154577398 passed. The exact reviewed head's run 38164707676 is queued. Those app checks do not establish the Python pipeline's safety or rights claims. No local tests, Xcode or simulator run. The latest source was inspected before reading the PR Proof; the Proof remains stale at 27 recipes.

[SamGu-NRX]

Verdict: BLOCK

Reviewed b1e6cdc for fl-recipe-platform-f1 against plan #86. This supersedes my review at 27613a1. The delta changes only REPORT.md; none of the previous source blockers was fixed. The updated counts and quantity shortfall are useful, but REPORT.md:42-48,374-386 and the Handoff incorrectly claim safety/rights clearance and only a resolution decision left.

Blockers:

• data/recipes/corpus.jsonl:35,121,129: the breakfast casserole still specifies 160 F instead of 165 F; Citrus-Braised Pork still specifies 145 F without a three-minute rest; the tuna pasta gratin still attaches a 145 F fish check to bubbling-cheese text instead of a 165 F casserole check. tools/ingest_originals.py:158-176 accepts these through the external linter. Correct or quarantine them and require dish/process-specific checks. See the federal temperature chart.
• data/recipes/tools/convert_bundled.py:386-401: all 25 seeds still receive original-work permissions based on no Source line, initial-commit presence and absence from a BBC cache. None establishes authorship or redistribution permission. Require affirmative provenance or quarantine. This is an evidence gap, not an allegation of copying; REPORT.md:21,42,54-57 must not call these cleared.
• data/recipes/corpus.jsonl:92: fl-original-0065 offers “ghee or vegetable oil” while claiming dairy-free unconditionally. The pinned scanner at 26c7004, evaluation/suites/recipes/rs.py:259-285, ignores unknown ingredient names and accepts an empty conflicts list. validate_corpus.py:114-118 and ingest_originals.py:158-168 trust that result, so this ambiguous milk-based choice passes. Resolve every included ingredient and alternative; unknown composition must block restriction claims. REPORT.md:44 cannot substantiate zero unchecked claims with this scanner.
• data/recipes/tools/ingest_nhs.py:671-732: generic meat still gets 160 F even for ground poultry; any earlier annotation suppresses fallback checks, including leftovers; eggs take priority over poultry in the fallback. Whole pork also lacks its required rest. Check every applicable food and process, including mixed dishes and reheating, before accepting a recipe.
• data/recipes/tools/convert_bundled.py:324-326 and corpus.jsonl:1,9: attaching thermometer metadata still leaves the cook-facing instructions as “Stir-fry rice for 2-3 minutes” and “simmer 18-20 min.” Put the required measurement in the instruction, or prove a consumer presents it before completion.

Lesser findings:

• REPORT.md:379-386: adding conversion-table references alone cannot resolve the 217 lines still marked unknown. The pinned rs.py:212-222 counts only resolved quantity kinds. Rebuild affected corpus lines from actual measure evidence after adding reviewed conversions; preserve unknown where evidence is absent. validate_corpus.py:123-153 still prints the percentage without enforcing the 95% bar.
• tools/apply_substitutions.py:113-128 still marks all supported roles verified without identifying the ingredient's role in this recipe or checking eaters. REPORT.md:232-245 describes table association, not recipe-context or restriction clearance.
• convert_bundled.py:291 still flattens required and optional ingredients. The expanded report still lacks the required stratified 60-recipe audit and byte-for-byte regeneration command. The corpus has no equipment, nutrition or untested fields, so the Handoff's “nothing” left is unsupported.

Criteria: rights evidence partly met; deterministic structured ingestion met; quantity evidence partly met; nutrition/USDA identity missing; safe atomic steps partly met; context-specific substitutions partly met; dedupe partly met; report/audit partly met; scope/no workflow edits met. 2/9 met.

I independently parsed 428 recipes, 2,988 ingredient lines, the 416/2,355/217 quantity-kind split, 1,050 claims and 2,087 substitution annotations on 405 recipes. CI run 38164866629 and all three checks passed at this head; the workflow/project targets cover the existing iOS app, not these Python/data gates. The PR's component-lint output does not clear the counterexamples above. No local tests, Xcode or simulator run; no exhaustive authorship or culinary clearance claimed.

[SamGu-NRX]

Verdict: BLOCK

Reviewed 1d51ce63d999fbbfe527639e9ca0e306dca8e1ae for fl-recipe-platform-f1. The delta from my reviewed b1e6cdc changes only REPORT.md. The new quantity-evidence handoff correctly assigns data fixes to this branch; it does not fix the source blockers.

Blockers:

• data/recipes/corpus.jsonl:35,121,129: the breakfast casserole still uses 160 F; whole-cut pork still uses 145 F without the required three-minute rest; tuna pasta gratin still attaches a 145 F fish check to a visual bubbling-cheese endpoint. Correct or quarantine these. The federal temperature chart specifies 165 F for casseroles and the pork rest. The linter's green result does not establish safe dish/process checks.
• data/recipes/tools/convert_bundled.py:386-401: the 25 seeds still receive original-work permissions from absence of a Source line, initial-commit presence and absence from a BBC cache. Those facts do not establish authorship or permission to redistribute. Require affirmative provenance or quarantine. This is an evidence gap, not an allegation of copying.
• data/recipes/corpus.jsonl:92: fl-original-0065 still offers ghee or vegetable oil with an unconditional dairy-free claim. The pinned scanner at 26c7004, evaluation/suites/recipes/rs.py:259-285, ignores unrecognized ingredient names and accepts matching empty conflicts. validate_corpus.py:114-118 and ingest_originals.py:158-168 trust that result. Unknown composition and alternatives must block restriction claims, not establish dairy-free clearance.
• data/recipes/tools/ingest_nhs.py:671-732: generic meat gets 160 F even for ground poultry; any earlier annotation suppresses later fallback checks, including reheating; an egg classification outranks poultry in mixed dishes. Whole pork lacks its rest. Validate every applicable food/process before accepting the recipe.
• data/recipes/tools/convert_bundled.py:324-326, corpus.jsonl:1,9: thermometer metadata is still absent from the corresponding cook-facing rice reheating and chicken instructions. Put the check in the instruction or prove the consumer presents it before completion.

Lesser findings:

• REPORT.md:375-382 still says nothing else blocks and reports rights, safety and claim clearance. The PR body still says only the resolution bar remains and that this task has nothing left. The corrected conversion ownership does not resolve the rights, safety and restriction failures above.
• validate_corpus.py:123-153 prints resolution coverage but does not enforce the 95% bar. The corpus still has 217 unknown lines and needs at least 68 more evidenced resolutions. Adding conversion rows alone cannot resolve lines whose quantity kind remains unknown.
• tools/apply_substitutions.py:113-128 still associates supported roles without identifying the role in this recipe or checking eaters. convert_bundled.py:291 still flattens optional and required ingredients. Nutrition/USDA identity, equipment, untested flags, the stratified 60-recipe audit and a byte-for-byte REPORT regeneration command remain unfinished.

Criteria, 2/9 met: rights evidence partly met; deterministic structured ingestion met; quantity evidence partly met; nutrition/USDA identity missing; safe atomic steps partly met; context-specific substitutions partly met; dedupe partly met; report/audit partly met; scope/no workflow edits met.

I independently re-parsed 428 recipes, 2,988 ingredient lines, quantity kinds 416 source_grams / 2,355 household / 217 unknown, 1,050 claims and 2,087 substitution annotations. The preceding 32fbfef passed all three checks. At this reviewed head, CodeRabbit passed and iOS build-and-test was still running when inspected. This workflow tests the existing app rather than these Python/data gates. No local tests, Xcode or simulator were run.

[SamGu-NRX]

Verdict: BLOCK

Reviewed 4b6e163f4aed3e3bfa77b70197cc9322881e9135 for fl-recipe-platform-f1. Since my review at 1d51ce6, only REPORT.md and three quantity-audit/validation tools changed. The corpus, rights records and ingestion remain unchanged. Preserving unknown quantities is correct; it does not resolve the release blockers.

Blockers:

• data/recipes/corpus.jsonl:35,121,129: the breakfast casserole uses 160 F; whole-cut pork uses 145 F without a three-minute rest; tuna pasta gratin relies on bubbling cheese with a 145 F fish annotation. Correct or quarantine these records. The federal temperature chart requires 165 F for casseroles and the pork rest.
• data/recipes/tools/convert_bundled.py:386-401: the 25 seeds receive original-work permissions because they lack Source lines, existed in the initial commit and are absent from the BBC scrape cache. None establishes authorship or redistribution permission. Require affirmative provenance or quarantine. This is an evidence gap, not an allegation of copying.
• data/recipes/corpus.jsonl:92: fl-original-0065 offers ghee or vegetable oil with an unconditional dairy-free claim. The pinned scanner at 26c7004, evaluation/suites/recipes/rs.py:259-285, ignores unrecognized ingredient names and accepts matching empty conflicts. data/recipes/tools/validate_corpus.py:223-228 trusts that result. Unknown composition and alternatives must prevent restriction clearance.
• data/recipes/tools/ingest_nhs.py:671-732: the generic no-pink rewrite assigns 160 F even for ground poultry; any earlier safety annotation suppresses the later reheating check; the mixed-dish fallback chooses egg before poultry. The pork rewrite also omits rest. Check each applicable food and process, including poultry and reheating at 165 F.
• data/recipes/tools/convert_bundled.py:324-326, data/recipes/corpus.jsonl:1,9: rice reheating and chicken checks remain metadata absent from the corresponding cooking instructions. Put them in the instruction or prove the consumer presents them before completion.

Lesser findings:

• New data/recipes/tools/validate_corpus.py:85-100,283-291 accepts NaN/Infinity conversion masses and null name/url/quote fields because it stringifies them. Duplicate list IDs overwrite earlier rows. The downstream R2 checks at pinned rs.py:120-134 verify reference presence but never bind the row's ingredient/unit to the recipe line or verify grams = value * grams_per_unit. A valid citation for a different ingredient or measure can therefore count an unsupported mass as resolved. Reject malformed and duplicate rows, require finite values and validate each line against its conversion. No corpus conversion table ships yet; these are defects in the new acceptance path.
• data/recipes/tools/validate_corpus.py:292-323 prints coverage without enforcing the 95% minimum. The unchanged corpus has 2,771/2,988 resolved lines, 92.74%, with 217 unknown lines and a shortfall of 68. A green exit cannot prove the release bar.
• data/recipes/REPORT.md:464-468 and the PR body still treat rights, safety and restriction clearance as established. The narrower quantity handoff does not discharge the blockers above. The 60-recipe field audit, byte-for-byte REPORT regeneration, USDA identity/nutrition, equipment and untested flags remain unfinished. Context-specific substitution checks also remain incomplete at data/recipes/tools/apply_substitutions.py:113-128.

Criteria, 2/9 met: rights evidence partly met; deterministic structured ingestion met; quantity evidence partly met; nutrition/USDA identity missing; safe atomic steps partly met; context-specific substitutions partly met; dedupe partly met; report/audit partly met; scope/no workflow edits met.

I independently re-parsed 428 recipes and 2,988 ingredient lines, with 416 source_grams, 2,355 household and 217 unknown. All three checks passed at this exact head, including iOS build-and-test. The workflow covers the existing app, not these Python/data gates. No local tests, Xcode or simulator were run.

[SamGu-NRX]

The food-safety canon linter (fl-safety-lint, canon 299ff32, umbrella PR #126) was run over a verbatim extraction of this branch's corpus.jsonl (428 recipe records @ 4b6e163): 360 findings. Full list: docs/audits/food-safety/113-recipe-corpus-v2.md. Please apply the required changes on this branch.

[SamGu-NRX]

Verdict: BLOCK

Reviewed 0f39ea8 for fl-recipe-platform-f1, including both new commits since 4b6e163. The 25 unsupported seeds are removed from the corpus and manifest into rights quarantine. Previously cited casserole/pork instructions are corrected; the ghee alternative no longer claims dairy-free. The final delta calls both text and composition gates, so the intermediate metadata-only validator finding is resolved.

Blockers:

• data/recipes/tools/food_safety.py:220-241,302-305,365-373 treats keywords as affirmative cooking checks. With chicken ingredients and poultry/165 F metadata, an instruction saying "A thermometer reading 165 F is unnecessary; cook until no longer pink" satisfies the thermometer and temperature searches. Checked steps then skip the appearance test. Likewise, "Do not rest 3 minutes" satisfies whole-cut rest. Require an affirmative internal-food measurement and rest instruction; reject negated checks and oven temperatures. The federal chart requires food temperature measurement, not keyword presence.
• data/recipes/tools/composition.py:71-74,132-156 removes peanut butter and almond milk from step text before every claim scan. A known lentil/oil recipe with a nut-free claim and a step-only peanut-butter serving suggestion retains that claim: no listed ingredient conflicts, and the only step evidence is erased. Exempt these phrases only from claims they actually do not contradict; resolve compound foods before checking every affected allergen. The sour-cream test does not cover this branch.
• data/recipes/tools/food_safety.py:77-87,109 classifies venison as a 145 F whole cut. A venison roast with matching thermometer metadata and a three-minute rest therefore passes this gate below the federal chart's 160 F venison minimum. Venison is not currently in the 403 records, but this is an unsafe acceptance rule in the new corpus tooling. Use the canonical rulebook or refuse unsupported classes instead of maintaining another incomplete temperature table.
• data/recipes/tools/rights_evidence.py:42-65 checks the spelling of an authoring-record citation or the presence of two HTTPS URLs, without resolving either to supporting evidence. A nonexistent original record with the matching id/SHA-shaped string passes; an OGL-labelled government record passes with two unrelated URLs. Verify the referenced record, recipe identity and source terms against pinned evidence before accepting. String format alone does not establish permission.

Lesser findings:

• data/recipes/tools/validate_corpus.py:337-345 still prints quantity coverage without enforcing 95%. Static parsing at this head finds 2,618/2,835 lines with source_grams or household quantity kinds and 217 unknown. The earlier conversion-integrity and context-specific substitution findings are unchanged.
• data/recipes/REPORT.md:464-468 and the PR's Verification/Handoff still describe 428 recipes, zero quarantine and older proof. Current files contain 403 corpus records, 25 rights-quarantined records and 65 claims, down from 981. Update the report and prove the new gates. The rights gate's 12-word overlap cutoff is a heuristic, not clearance of source rights.

Task criteria, 2/9 met:

• Rights evidence: partly met. Seeds quarantined; evidence resolution remains missing.
• Deterministic structured ingestion: met in the inspected tooling.
• Quantity evidence: partly met. Unknowns preserved; coverage and conversion validation incomplete.
• Nutrition/USDA identity: missing.
• Safe atomic steps: partly met. Many records corrected; unsafe gate inputs remain.
• Context-specific substitutions: partly met, unchanged.
• Dedupe: partly met, unchanged.
• Report/60-recipe audit: partly met. Current totals and regenerated proof missing.
• Scope/no workflow edits: met. Own changes remain under data/recipes.

Current-head app build-and-test is in progress in run 38193861233. It does not execute these Python corpus gates; the preceding head's checks were cancelled. I inspected source, data differences and tests without running local tests. The counterexamples above are source-derived, not claimed execution results.

[SamGu-NRX]

Verdict: BLOCK

Reviewed 7fc68d4 for fl-recipe-platform-f1. The new gate, ingestion and test changes were inspected, along with all changed corpus steps. The four prior examples are addressed: negated checks now refuse, peanut butter survives the nut-free scan, venison refuses, and rights citations resolve against pinned evidence. Static comparison confirms all 401 pinned original records match their source manifest/title/ingredient-name hash at 607ef99. The corpus now contains 401 originals and 27 quarantines. Quarantining the two NHS records agrees with the Healthier Families terms, which require permission for commercial content use.

Blockers:

• data/recipes/tools/food_safety.py:305-316,384-388,475-481 still does not interpret the temperature comparison. Chicken plus poultry/165 metadata and the step "Cook the chicken until a food thermometer in its centre reads below 165 °F" satisfies affirmative_check: thermometer, reads and 165 are present, with neither recognized negation nor oven setting. text_temps sees only 165, so no lower numeric temperature is rejected. Require a reading at or above the minimum, and reject below/less-than instructions. The federal chart specifies a minimum internal temperature. This is a source-derived acceptance path, not an executed test.
• data/recipes/tools/composition.py:139-150,168-180 still resolves only listed ingredients, then uses finite word lists for foods introduced in steps. A known lentil/oil recipe claiming dairy-free with "Fold in mascarpone before serving" has no listed conflict and no step-pattern match: mascarpone appears in neither STEP_CONFLICTS nor STEP_COMPOUNDS. Unlisted step ingredients need composition evidence too. Refuse or remove claims until those foods are reconciled with the ingredient list and canonical facts. Adding pesto closes that example but does not make unknown step composition fail closed.

Lesser findings:

• data/recipes/tools/ingest_nhs.py:725-727,755-760 can put UNCHARTED into _endpoint, which indexes MIN_F without that key. An uncharted meat with a matching doneness clause crashes ingestion instead of returning a refused page with a reason.
• data/recipes/tools/validate_corpus.py still prints the 95% quantity bar without enforcing it. Static counts are 2,612 source_grams/household lines out of 2,829, with 217 unknown, matching the updated report's 92.33%. Earlier conversion-integrity and context-specific substitution findings remain.
• REPORT.md now gives current totals and limits, but the PR Verification/Handoff still claims 428 rights-clear recipes, zero quarantine and older proof. Update the body and prove these gates at this head. The overlap cutoff remains a heuristic, not evidence of permission.

Task criteria, 3/9 met:

• Rights evidence: met for the shipped corpus. Unsupported seeds/NHS are quarantined; original records resolve to pinned source evidence.
• Deterministic structured ingestion: met, with the uncharted-input error-handling issue above.
• Quantity evidence: partly met. Unknowns remain explicit; coverage and conversion validation remain incomplete.
• Nutrition/USDA identity: missing, unchanged.
• Safe atomic steps: partly met. Corrected corpus instructions and stronger tests; unsafe acceptance paths remain.
• Context-specific substitutions: partly met, unchanged.
• Dedupe: partly met, unchanged.
• Report/60-recipe audit: partly met. Current report improved; PR proof and the remaining acceptance bars are unfinished.
• Scope/no workflow edits: met. Delta stays under data/recipes.

Current-head build-and-test was queued in run 38197238709 at inspection. The app workflow does not execute these Python gates. New tests cover the repaired examples and pinned evidence reproduction; no local tests ran.