[codex] test source-conditioned ETR-1 frontiers

[TheGreenCedar]

Context

The valid Phase 1A experiment on draft PR #2120 established that authenticated match-centred hydration alone cannot repair packet relevance. The next and only authorized hypothesis is ETR-1: test whether a second semantic retrieval conditioned on each authenticated first-pass source fragment creates a better bounded evidence frontier than the same number of unconditioned searches.

Refs #2116 and #2119. This issue is a benchmark-only stacked child of PR #2120 and does not alter or replace its failed receipt.

Outcome

Produce one immutable, independently verified ETR-1 receipt over the burned 24-case development corpus. Compare equal-budget raw-question and question-plus-seed-source retrieval over the same frozen fragments and production CodeRankEmbed model.

Fixed inputs

• Parent head: c9c935d87129a79f326b650bbf23d73191df8b4f
• Fragment diagnostic: ca185ed13c635bbb4b64cc6760c5025799700359ebcc4dd3bcc53e34f8cf9194
• Fragment build: 2201780e1a752db4bfcceb047bf5cd0b5a854733c4050330ef87575b960f3baf
• Lexical membership freeze: e6867b5c79706160021ec5edf60792273345ca97cae8377e69934d5e2c9992ee
• Questions: 8e7219a59c973c02f8ea93120bb680da46a75b8272153986c76e55bfb73ca3b6
• Annotations: 52b0cc223292bc70f1e4fa3f52b67bf42a91e4d4b9ed997aa12c648c068e9ade
• Model contract: cb0e3c00290f1eb21ecdcd873521d03331069b1efa766fcd1e493e6d4299b4b7
• Model: 666db8df27c88570cdc07adca28646260038b8ca65354911d57b936ebf56efaa
• Tokenizer: 7465b93c945b7a266481e6785aa13e505c625562c1c046c4b762bb4da4d46082
• Lexical policy source: 43b2478d75abd3d5689d05e08c072e4148fd21ab29bcc55533d30a494edf986b

Acceptance

• Both arms share BM25 top-sixteen seeds, frozen fragments, model, search count, exclusion policy, source budget, and feasibility constraints.
• The control repeats the raw question; the candidate uses the raw question followed by two newlines and the exact authenticated seed source.
• Graph, symbol documents, BGE, host queries, production packet routing, public schemas, and annotations are excluded from retrieval.
• All outputs freeze before the evaluator can read annotations.
• An independent verifier reconstructs identities, queries, exclusions, pools, costs, optima, aggregates, and latency equations at the exact head.
• The frozen four-stage decision table determines stop, control frontier, conditioned frontier, or one byte-identical latency repair.

Non-goals

No selector, production integration, release work, new corpus, model change, parameter sweep, graph layer, public API, plugin guidance, or version bump.

[TheGreenCedar]

ETR-1 execution-boundary recovery is authorized before another corpus run.

The user accepted the revised evidence-engine plan on 2026-09-04. PR #2126 remains a benchmark-only stacked child of draft PR #2120, based on c9c935d. The current owner is the existing codex/etr1-frontier worktree; one implementer and the existing adversarial verifier remain the entire team.

The repair covers one causal class: the canary must use real preparation, native encoding, the Rust paired runner, authoritative validation and evaluation; native/server/request identities replace the incorrectly assumed command sequence; exclusions have one canonical ordering; generator/runner inputs and outputs receive independent pre-dispatch and terminal bindings; clean analysis tools are bound; cancellation and a fixed 1,800-second paired-run deadline are enforced; enclosing request wall time includes explicit unaccounted work.

ETR-1 questions, fragments, model, tokenizer, searches, source budgets, exclusions, optimizer, quality thresholds and 1.25-second latency gate remain unchanged. If only conditioning is sufficient but does not satisfy its material-value gate, no frontier is selected. Results name unconditioned_frontier_selected, conditioned_frontier_selected, no_frontier_selected, or invalid/not_evaluated. No result establishes a working packet.

A valid ETR-1 failure may lead to exactly one separately preregistered STR-1 experiment using certain source-witnessed one-hop repository relationships. This is a prospective authorization, not an alteration of the ETR-1 result or any earlier stop. A frontier pass authorizes preparation of the bounded SEL-1 contract only. Model execution remains local; no new model is selected. Production behavior, public schemas, plugin guidance, confirmation/builder/sealed corpora and release work remain unopened.

Current focused repair checks: Rust ETR-1 tests 14 passed; JavaScript evidence tests 15 passed; diff whitespace check passed. These are development checks, not exact-head acceptance or a real canary. Next: finish the consolidated execution-boundary tests, freeze one clean child head, build and run the actual synthetic pipeline, then request the existing verifier's complete mutation matrix before the 72-row corpus run.

[TheGreenCedar]

ETR-1 execution freeze

The execution-boundary repairs are committed and pushed on draft #2126. No production, model, representation, threshold, or public-surface change is included. No further source or workflow changes are planned for this run.

• Source: 8c7ba1e02f68b3cdbafd107ea21deb53affd6c3c
• Tree: 4e16debd285a667b7a79a476e05ccdf734e79cc8
• Runner SHA-256: 03b2ea0aea77e1a6c198a708cf0fb5401a0826d88364a7689d2c13f25b875861
• Native document diagnostic SHA-256: fa70fc6b1b47f646b8310e2d15b6d90103dff614af101725bc5d4a90fa3f005b
• Method freeze: 2b8aa08096d031af02180698e724ea81ed0bf145b73db278d5abd9734a405a7f
• Preparation: a8a970c7f71f708bef8096d747ac87cb30833b39bb57eff16cf6f046cc3e0713
• Fragment document input: aeb1ba80f56d43c1deedb356559708f39a75031b64e0f3d6e769b1b44b31891f
• Passing actual native canary receipt: 59ff73cfde8dab09b4239909c5fbec0c93006abd722bb0f1f9bdf62e974fce4c

The real canary exercised preparation, native document encoding, Rust paired retrieval, authoritative validation, and the exact evaluator. It covers sixteen seeds, natural underfill to one, empty seeds, multiple query batches, and cumulative exclusions. The prior failed canary is preserved without quality authority.

Corrections bind omitted native optional fields correctly, native completion identities, globally sorted exclusion unions, enclosing request time and unaccounted work, cancellation and the 30-minute paired deadline, exact child environment/cwd, and a mandatory same-head/binary canary before corpus launch. Analysis source and runtime identity are bound independently of the generated vectors.

Focused ETR JavaScript tests: 17 passed; witness evidence tests: 7 passed. Previously completed unchanged Rust lanes remain passing: ETR 14, witness 4, evidence-address 2, addressed-hydration 2, witness-seam 1, native diagnostic 6. Both generalization checks and whitespace checks pass.

The existing independent verifier owns the exact-head hostile matrix. Corpus model execution waits for that acceptance. Thereafter document vectors are regenerated, and all 72 wording rows run from zero. No partial query reuse. Annotation access remains forbidden until both arms and execution receipts are frozen and validation passes. Prepared-state timing remains the frozen ETR ordering; no result implies a working packet.

External evidence: /Users/albert/.codex/evidence/codestory-etr1-recovery.t1acxt. Parent draft #2120 and prior valid failures remain unchanged.

[TheGreenCedar]

ETR-1 corrected execution freeze

Draft #2126 is clean and pushed at f54b8e6c0f503f2f4a61eed35b303f9d72004e0d, tree 69597539574c74e9acd251ae2c0687e4c5b3d0f4. No further source or workflow changes are planned. Parent #2120 is unchanged. This replaces the pre-execution source freeze at 8c7ba1e; no corpus model rows were launched there.

• Runner: ea9c035c87c56ea076fb78d3c8fff2d4f892ef3cfd13780b63caf1fca367ead4
• Native document encoder: a87a2f41da33cc7d739e8400fa1dd385b5f2db85f1c4b3eb645fb3555df0893f
• Preparation: ebed3d07f1c034c0e2e6c58da8f824e93a5324abc318c6cb556f6d68cd4233e4
• Exact fragment document input: aeb1ba80f56d43c1deedb356559708f39a75031b64e0f3d6e769b1b44b31891f
• Actual passing native canary: 2ce93358192238ddbac9b66057c13491e706b147747ff7a233e4545cdb0a2a94
• Predeclared document job: 59770c9ab27193614c4237e31797c88438325e783157959ce5808849b49c9f22

The last correction preserves macOS's captured process-locale variable in the recorded child environment and reconciles every native document completion with the pinned encoder's consecutive sixteen-record batch progress. It does not change representation, model, retrieval, scoring, exclusions, thresholds, or arm ordering.

Focused ETR JavaScript tests: 18 passed. The complete native canary passes preparation, document encoding, Rust paired execution, authoritative validation, and synthetic evaluation. Corpus preparation authenticates 10,369 fragments and 72 wordings without annotation access. Previously passing unchanged Rust proof is reused; no broad release proof is opened.

The existing verifier is completing the named hostile matrix. Corpus encoding remains gated on acceptance and the same-head/binary passing canary. The fresh output root is /Users/albert/.codex/evidence/codestory-etr1-f54.8JaKOz. Document construction is separate from the paired run's 30-minute deadline. Queries and pools will be regenerated from zero, frozen before validation, and annotations opened only after validity passes.

[TheGreenCedar]

ETR-1 execution completed; validation rejected before scoring

At frozen head f54b8e6c0f503f2f4a61eed35b303f9d72004e0d, document construction and all 72 paired rows completed. Validation rejected the run before annotation access with public row cost changed: 694 != 323.

• Run manifest: be8e6299f100d30d883a1ae43f74ff26e421c49bc099de4f830918ac01d30053
• Independent paired execution: d27aeee4a34409d2d5d339d25b0ac9886940bb96a5cc142ebe5674f13dc8a5e4
• Invalid validation receipt: efc3a7400aacb7ce4f7f7f38fe5f9b001f40d0802bd0ab4bc404916831651f16
• Complete row-cost audit: 68e6e9ccd6bf28594c1f4ced017107a1f6b74d48f8e86f2230df3ee377b6705e

The validator's JavaScript dot-regex discarded source before CRLF boundaries. The existing Rust producer and retained oracle costs agree. An audit of all 10,369 fragments finds exactly 25 affected fragments; reconstructing the existing LF-only line rule resolves every difference without changing source, rows, costs, or budgets.

The correction is limited to that generic validator rule, generic line-ending tests, and mixed line endings in the actual synthetic canary. The existing verifier is checking remaining execution stages before the next freeze. No annotation, feasible recall, or conditioning result has been opened or inferred. This receipt remains invalid/not_evaluated; it cannot select a frontier or authorize STR-1 as a valid quality failure. A corrected candidate must rerun from zero.

[TheGreenCedar]

ETR-1 complete-line correction freeze

Draft #2126 is clean and pushed at eb968e4f68e8999ce575ba25bee9d1c13b56cfdf, tree 8109976c4b3bede4a2c02961b71594be2fa89760. No further source or workflow edits are planned. The source-line contract is now shared by public-cost validation and query reconstruction; both match the existing Rust LF-inclusive representation. Generic tests and the actual native canary include CRLF and Unicode separators. Frozen corpus source, costs, queries, retrieval policy, model, limits, and thresholds are unchanged.

• Runner: 828b41d3cd5856997fe0eeb9876ed72f6724ab77175afd1ac18477161fc52691
• Native document encoder: d92a4eec56bc71f4855a066ce3842425e6a408ddf466244b183c9977e5f229f6
• Preparation: 30b84d4d848f96bd4fe799f2e0f28b9114971da0e47bf98ebe54fe36242199fd
• Unchanged fragment input: aeb1ba80f56d43c1deedb356559708f39a75031b64e0f3d6e769b1b44b31891f
• Unchanged method: 2b8aa08096d031af02180698e724ea81ed0bf145b73db278d5abd9734a405a7f
• Passing actual mixed-line-ending canary: df338060ef832a36a306566fa1db51f4640f9be4f8a3066caf7c29b013366c6f
• Document job frozen before launch: 6f34ed001a2fc8dfc88eedee0ebd2b3f5df6e8427693f92a24d7a1e2e3537adb

The prior f54 run remains invalid/not-evaluated. Its binaries and receipts are preserved. Independent diagnostic reconstruction after correcting both splitter sites found no additional mismatch in all72rows,288querybatches,2304queries,649documentbatches and their source/score/pool/timing bindings. That diagnostic does not repair or qualify the prior run, and no annotations were opened.

Current focused ETR JavaScript tests:19passed. The sole verifier is executing the retained correction matrix at this exact head; no new review lane. After acceptance, the entire document and paired-query pipeline reruns fromzero in /Users/albert/.codex/evidence/codestory-etr1-final.8ttvNp, including the fixed first-native-batch repeatability sample. Thirty-minute paired deadline remains. Annotations and quality remain gated on fresh frozen outputs and successful authoritative validation.

[TheGreenCedar]

ETR-1: no frontier selected

The complete paired run and independent reconstruction are valid at eb968e4f68e8999ce575ba25bee9d1c13b56cfdf (tree 8109976c4b3bede4a2c02961b71594be2fa89760). Neither frontier meets the frozen source-evidence sufficiency gates.

Measure	Unconditioned	Source-conditioned	Required
Mean feasible source recall	71.19%	72.52%	85%
Complete acceptable-set rate	40.28%	41.67%	75%
Cross-file relationship recall	52.50%	49.60%	70% per group
Debugging/impact/change recall	63.95%	66.94%	70% per group
Known-location recall	88.23%	88.89%	70% per group
Unknown-name concept recall	80.09%	84.66%	70% per group

Conditioning adds 1.33 recall points and 1.39 complete-set points, below the ten-point material-value requirement. It loses 2.91 cross-file recall points and adds an atom in 6 of 13 eligible incomplete cases. Source/address authentication is complete. Latency selection is not_evaluated because quality selects no frontier. This is an answer-aware frontier result, not a packet result.

Bound receipts

• Run: c14da697d03707c0096f5f2fd7a97bff2ab5b4a6f9326c4a9a03da2066d545f2
• Validation: dd15c5842b68857c8c684b680cf3d1e307cd4b1a874736441c594490723e7145
• Evaluation: 545a7d2ba1731d32bda7a4f1daf346d11b81f32bbbb732a3bb91673fad11c982
• Independent reconstruction: 2fc235d14d3477b8c7985d0a5078f2e4ab2e4a1c56c668a2d79b7d198a3e51b3
• Evidence root: /Users/albert/.codex/evidence/codestory-etr1-final.8ttvNp
• Independent root: /Users/albert/.codex/evidence/codestory-etr1-final-verifier.t3DikR

The verifier reconstructed all 72 wordings, 144 arm outputs, 2,304 queries, 1,990,848 score values, exclusions, successors, legal pools, 180 exact acceptable-set optima, and 24 case aggregates. The accepted retained oracle reproduced 24/24 cases. Original artifacts remained unchanged.

Focused source checks and the real native canary passed. ETR-1 remains benchmark-only; no production behavior, model, public schema, plugin guidance, version, calibration, or release changed. Earlier invalid runs, Phase 1A failure, and the old compiler stop remain preserved.

Next authority

The latest accepted plan prospectively permits exactly one separate STR-1 structural-frontier experiment after this valid failure. It must be preregistered before candidate output and retain ETR-1's frozen unconditioned control, questions, fragment universe, source-only oracle, model, and absolute/material/latency gates. No ETR retune, selector implementation, production integration, or release work is authorized by this result.

PR #2126 remains draft as the ETR implementation and receipt record; PR #2120 remains unchanged. Closes #2125 only on eventual merge; Refs #2116 and #2119.