Research/content region coverage v1

evidence/budget-scaled-coverage-2026-08-22/artifacts-SHA256SUMS.txt

@@ -0,0 +1,22 @@
+4a6744ae64b783f770dd426012cd4439c1e2d772416655107682ed38143b2067 *binding-cross-check.json
+49b105cecfeec14880d55ff59e578855f4ce9588574f1acd167799e0f9937038 *confirmation-corpus-420000.json
+5d08679efcb3bfbe418beb07ec94480f6eefab017708a8ba9ed2afcb472dc07d *confirmation-evidence-420000-provenance.json
+7fb0ce6c21e9f7c8a599889591e1ce24ff8cec7c8263da4a87444ffd4949ff5d *confirmation-evidence-420000.json
+06cd3d9c496e65a0c7987e5e71d9ea57d25e704b59f14cae4f0d28bf634e1386 *confirmation-plan-420000-provenance.json
+3079622ccdb4d86f48b1a77a64a00d6cff41b1918655d5526ae783025185181c *confirmation-plan-420000.json
+c9d2686b16e7c0b213bff14475e845f06c53ac64601b53ccdbed1fea2a1287df *confirmation-robustness-420000.json
+57b580c8839ea85a650993e302443fec4e16f76ea4d1ff7156fcb5468017f7ce *dev-corpus-401000-local.json
+956e9a5c2994dc5a511ea535968cd4bbed68ba1a3ce50a1ca410c56df03ab62f *dev-corpus-410000.json
+d0bec9524b623fb582b6f8699bbe4ef412ccd42b2a50662675480133e0f793e8 *dev-evidence-401000-provenance.json
+d352c3e4393f088dc47fe40af3c256c71472853dd96432df48350de0673f9ec6 *dev-evidence-401000.json
+12a63869c5f68845adfa88967faaae5669845f7e18fca882e98e50300ea3528a *dev-evidence-410000-provenance.json
+2222d4293183071ecb479633bd279ddf858655c5539fd6f291030fd1faf4e88a *dev-evidence-410000.json
+642cee2ba8b9ce853b020517e85b35ea34867e77b736d53ed95915e291301439 *dev-plan-401000-provenance.json
+0fffc9284e502387e9743a281197e7bf989abd3b06dc2b0e72074c98ee941d40 *dev-plan-401000.json
+e866f83819de7e4e89d7bad95123c4f8e41d499bfc876e1d056685ada61853dc *dev-plan-410000-provenance.json
+18e05ee9d8be733af401e96e4f07dcab9f22048cb2b7b8c6dd860b97a3428a5a *dev-plan-410000.json
+cbf6135f43fa6b6718d95320e2c6d5174e916353560efab1c2232f3af2183d1c *dev-robustness-401000.json
+3cf84fedf6fc2818cf494873e98ca63c98c179e433d269102fc6f82a7750e99a *dev-robustness-pooled.json
+6a8d031b883c1cc7eddfbb46f40d10de9be7233d9a93ca5cedb1b8dc4cf8c013 *rendering-manifest.json
+59345fdece25c4ebe5c1123fabd525c957b665f6fc39db8bd069122eda484281 *confirmation-render-pairs.html
+b0c740c804e2555f2b4a5616590cd342a2c60275686a5e7f8a44f29b92eb608c *confirmation-render-screenshot.png

evidence/budget-scaled-coverage-2026-08-22/binding-cross-check-2.json

@@ -0,0 +1,14 @@
+{
+  "corpus_artifact_hash_replay": "ccc9d5d6ffe5168f8059f7b2ab5f418050ed11b487e34e68ebf1634284a96760",
+  "evidence_binds_plan": true,
+  "evidence_hash_replay": true,
+  "evidence_profile": "key-blind-coverage-completion-v2",
+  "evidence_row_hash_replays": "128/128",
+  "plan_binds_corpus": true,
+  "plan_detector_access_false": true,
+  "plan_hash_replay": true,
+  "render_manifest_binds_plan": true,
+  "render_manifest_pair_count": 128,
+  "robustness_binds_plan": true,
+  "robustness_hash_replay": true
+}

evidence/budget-scaled-coverage-2026-08-22/binding-cross-check.json

@@ -0,0 +1,20 @@
+{
+  "corpus_artifact_hash_replay": "85cdda3bef549c961caac07767126ec2924f360f299b5d5a6b43f4272aa7e2ba",
+  "evidence_binds_corpus": true,
+  "evidence_binds_plan": true,
+  "evidence_hash_replay": true,
+  "evidence_profile": "key-blind-full-pool-coverage-v1",
+  "evidence_row_hash_replays": "128/128",
+  "evidence_row_texts_match_plan": true,
+  "evidence_threshold": 0.5584848484848485,
+  "evidence_threshold_hash_present": true,
+  "plan_binds_commit": true,
+  "plan_binds_corpus": true,
+  "plan_binds_manifest": true,
+  "plan_detector_access_false": true,
+  "plan_hash_replay": true,
+  "render_manifest_binds_plan": true,
+  "render_manifest_pair_count": 128,
+  "robustness_binds_plan": true,
+  "robustness_hash_replay": true
+}

evidence/budget-scaled-coverage-2026-08-22/cycle-2-report.md

@@ -0,0 +1,100 @@
+# Coverage-Completion Cycle 2 Report - 2026-08-23
+
+Branch `research/ngram-window-coverage-v1` (local only, nothing pushed). Confirmation scored
+once at commit `6d0ad2d2...` on corpus `ccc9d5d6...`. Protocol SHA-256
+`65856295a595acaa404392239ccd9db9fa43d3f45a86ae61ee0d88b0d1e0851a`, contract SHA-256
+`20ba6145793a89f0aa5cbc530305d9345731a60d60ad906199b323c74942d825`.
+
+## Executive summary
+
+On a new, source-disjoint 64-watermarked-source confirmation corpus, the frozen
+detector-blind key-blind coverage-completion profile achieved **0/64 transformed watermarked
+detections** (primary stretch target met for this condition) with **0/64 transformed
+unwatermarked false positives** and all 128 fidelity/representation gates passing. One
+pre-registered secondary gate FAILED and is reported as such: pristine watermarked detection
+was 58/64, below the required 60, because this corpus's per-corpus calibration threshold rose
+to 0.580261 (cycle 1: 0.558485), leaving six watermarked sources below threshold before any
+transformation. Conditional on initial detectability, 0 of the 58 pristine-detected sources
+remained detected after transformation. The permitted conclusion is unchanged in kind:
+detector-specific detection reduction under the fixed open detector, public development keys,
+and uncleaned text-only input path - here reaching the pre-registered narrow 0/64 target on
+this corpus. It is not universal undetectability, watermark removal, or any unknown-key,
+proprietary-detector, or cleaning-durability claim, and Level B (product admissibility)
+remains unmet.
+
+## Mechanism and diagnosis chain
+
+Cycle 1 ended at 11/64 residual detection. The post-hoc coverage diagnosis (tool:
+`tools/budget_scaled_coverage_diagnosis.py`) showed residual sources had mean achievable
+n-gram coverage 0.682 vs 0.847 for escaped sources, with realized coverage equal to achievable
+everywhere - the candidate pool, not the scheduler, was the bottleneck. Cycle 2 extended the
+surface spacing word list by 65 frozen general-English function words (protocol section 2),
+raising mean achievable coverage on confirmation watermarked sources to 0.827 (mechanistic
+endpoint >= 0.80 met) and mean candidate pools to 21.9 per source.
+
+## Development phase (recorded before confirmation scoring; freeze appendix)
+
+| Corpus | Cycle-2 ruleset | Cycle-1 ruleset (paired) |
+| --- | --- | --- |
+| 410000 (repeat use, labeled) | 0/32 detected | 1/32 (cycle 1) |
+| 430000 (fresh) | 5/32 detected | 4/32 (paired baseline) |
+| Pooled dev | 5/64 | 5/64 |
+
+Dev showed NO pooled improvement; the extension grew pools where function words exist but
+residual sources are content-word-dense with low candidate density (one had 5 candidates
+total). Residual sets overlapped strongly across rulesets, tied to high pristine scores plus
+sparse candidate regions. The confirmation proceeded as pre-registered and the outcome
+exceeded the dev signal; corpus-to-corpus variance is high in both directions (cycle 1:
+1/36 dev to 11/64 confirmation; cycle 2: 5/64 dev to 0/64 confirmation). This variance is
+itself a finding: single-corpus dev counts are weak predictors.
+
+## Confirmation results (64 wm + 64 unwm, budget 16, threshold 0.580261 at 1% target FPR)
+
+| Condition | Detected | Rate | Mean score |
+| --- | --- | --- | --- |
+| Pristine watermarked | 58/64 | 90.6% | 0.6348 |
+| Transformed watermarked | 0/64 | 0.0% | 0.5283 |
+| Pristine unwatermarked | 0/64 | 0.0% | - |
+| Transformed unwatermarked | 0/64 | 0.0% | - |
+
+Mean watermarked score drop 0.1064. Closest transformed margin to threshold: -0.0020 (one
+source), then -0.0115. Realized edit cost mean 12.72 of mean pool 21.9; zero-candidate
+sources: 0. Six pristine-watermarked sources below threshold (0.5590-0.5775) are listed in
+the evidence and constitute the failed >= 60/64 control gate.
+
+## Gates, robustness, rendering, bindings
+
+- Section-6 gates: 128/128 confirmation rows pass (hard invariants, protected spans zero,
+  max word edit rate 0.1333 <= 0.30, no introduced non-ASCII or hidden-category codepoints,
+  no introduced normalization instability, Cf-strip identity on introduced characters).
+- Cross-corpus text-hash disjointness: all pairwise overlaps zero among 410000, 430000,
+  440000; also zero against cycle-1 corpora 401000 and 420000 (no repeat of the cycle-1
+  degenerate collision).
+- Fixed-browser rendering: DOM textContent equals plan texts on 128/128 pairs; screenshot
+  archived (`confirmation2-render-screenshot.png`).
+- Independent cross-check: corpus/plan/evidence/robustness/render-manifest hashes all replay
+  and mutually bind; 128/128 evidence row hashes replay (`binding-cross-check-2.json`).
+
+## Adverse findings and caveats
+
+1. The >= 60/64 pristine-watermarked control gate failed at 58/64 (threshold sensitivity of
+   per-corpus calibration); the 0/64 primary result must be read together with this.
+2. Dev predicted no improvement (5/64 both rulesets); the confirmation result therefore
+   carries corpus-sampling uncertainty. A single 0/64 does not establish a stable zero rate.
+3. One transformed source sits 0.0020 below threshold - margin-level sensitivity persists.
+4. Cycle-2 word extension did not reduce dev-phase residuals; residuals concentrate in
+   content-word-dense, low-candidate-density sources that word-list extension cannot reach.
+5. Canonical-corpus cross-platform reproducibility limitation from cycle 1 still applies.
+
+## Classification
+
+- Product-safe code: extended surface ruleset (superset, base ruleset and frozen B16 profile
+  hash `6ad14226...` unchanged and test-pinned), coverage-completion profile + resolver +
+  registry selection, diagnosis tool, 8 new tests (total suite 1422 passing plus the known
+  Windows-only `spec.md` environmental failure).
+- Candidate outcome: narrow pre-registered 0/64 target MET on this corpus and condition,
+  with the control-gate failure reported; exploratory only; NOT product/release admissible
+  (visible-text edits); no overclaim.
+- Next research lever identified by diagnosis: candidate generation for content-word-dense
+  regions (rules that can place a visible-text-safe edit anywhere), which word-list
+  extension cannot solve.

evidence/budget-scaled-coverage-2026-08-22/cycle2-artifacts-SHA256SUMS.txt

@@ -0,0 +1,24 @@
+59ae8aa407b4f9b93b93c8f50b968cfd154d84d0795b38360a9be9e54fa981cd *dev2-evidence-410000-provenance.json
+e377fc52d0bfd766ec5824f2658d9897466c9d975bcd658623c2e579ff84a95a *dev2-evidence-410000.json
+ea1b3b9dff40f955d8761ec7683725c5f9319f3ac37fdf30c58cab8bfa5d050e *dev2-evidence-430000-provenance.json
+7d9fbe1229a511b16a1460954d947f8fa6cad58afd8262f2122393bbb0f36c5d *dev2-evidence-430000.json
+3dd70d23a852c220d91d719c8e78cdbc1894e1bdc596ec0c0a69bbb9844339f3 *dev2-plan-410000-provenance.json
+0573f47ff38e71e59e282482f796c0d963ddcee31b476d3a7805e82dc8af13a9 *dev2-plan-410000.json
+bdbb1c2148acb1fa4e23e6ca3ff3d4185384e5eff6081d9b0a5b1a5d735a09e4 *dev2-plan-430000-provenance.json
+3863cc67c67e357d5c67266f42d84dd7ddb3ad4869847f387e84ae0fd8f5a334 *dev2-plan-430000.json
+0bc77aea8e70b3f8a97b67e1824440e7959004a9fdf785c2eb514ccab53ba1a3 *dev2-robustness-pooled.json
+2129f7dc5c5b16d7fece545c4cf1ae156ccd6b16caa9dc27af1500e1887de56e *dev1-plan-430000.json
+97fd81a8346c2c51acc3299e364f37a9c83a7ae9b07b74cacb57cc0476194a48 *dev1-evidence-430000.json
+fb066647ad15c86e5c09ba5b1753ddbcb53e490d1a1a97e32e36a26016e93036 *confirmation2-corpus-440000.json
+13cdd4aa2ddbb359096aa3fa1e95e3c32a7d8b576696a6455280caee2bc2bbe9 *confirmation2-evidence-440000-provenance.json
+bde997a0103c60b7604815a074200543cd157fd3af28f59575eb9acef24cc568 *confirmation2-evidence-440000.json
+1b8608f4a23ad21e1090582787b645c8e48b38d6a20cfc287abdcbaa65f888cf *confirmation2-plan-440000-provenance.json
+a801ad020a339ba802467b0b0e48f501f95a9fc4b177ac8ea68b4153095c38e8 *confirmation2-plan-440000.json
+48ad930bf3ef424595ec77910c48fa15dc89e1e63ca1724a988556a9a30d23b0 *confirmation2-robustness-440000.json
+acb612a2294e5a2b269ee27169195afcf54309d5eb73d5ff6aaf3da840fef835 *coverage-diagnosis-420000.json
+b8636ec8334c8b5b24fbacd6cd0731a79f4ddd05426a37bdce4946de5ae45aee *coverage-diagnosis-confirmation2-440000.json
+2e3e904c035fcd169b6b4a8cc49b95338028fdbd4871c900f39a772517aa7cb1 *coverage-diagnosis-dev2-410000.json
+36508360e70baf33b3eb8cb316a12a3aae3fbbbc87bacb9b0addc5db60d8df2e *coverage-diagnosis-dev2-430000.json
+5d04b758e9d6964cf9501aa0416a455bd00b985c0fdb03dacbfbdb51fef5ce58 *binding-cross-check-2.json
+132ec9541ca134c0f5f2f15932da5e3775653c34314ffef8241ef8d5ac3bebf5 *rendering-manifest-2.json
+3e6de60f70a82e6b6fabc20ef010cec7051295a9199d8ee159e40bbeb1e8d7fb *dev-corpus-430000.json