schemaVersion: 1 parentHarness: H04 generation: H04 generationBundleId: H04-FORMAL-20260812-R1 analysisRunId: h04-postformal-20260812-019ff55f-a provenance: agent-derived userConfirmation: pending userProposalSelection: pending implementationAuthorized: false experimentAuthorized: false authorityBinding: workArenaCommit: 66e2b00153039b78cb0f630be1d9658e4dfdd073 workArenaTree: c94509c6a1c7b5b11f267e6b639ad40a18533919 matrixSpecCommit: 6a5a17e9a137d5e253b7f427002bbd03bf193546 matrixSpecTree: af7251e7b965458a5f61b95a612f6c2acdf91d45 contractSha256: 320159f59abb5e0420396bfaf69ea1057644f3f99b2e8c5d3f16f635255fffc5 workflowSha256: bdd094b9dedc4479744ac670cfe90ac0b357f0cf5cfba93ff9d80ad45301d5e4 resultSha256: 3ccddb992ef2cbc7691a81440c7024aba42c15fdbdce2f1840bcf8ae6ab8c76c coverage: 24/24 selected valid_behavior cells; 313/313 actor/session segments exclusions: targeted D1/D5 and D5 replacement are outside the denominator comparatorBoundary: H03 is descriptive only; neither aggregate deltas nor false-terminal=0 establish H04 causality weaknesses: - weaknessId: H04/h04-postformal-20260812-019ff55f-a/W-01 shortName: conflicting implementation terminal contracts decisionReadiness: decision-ready evolutionDiscussionAlias: H052 routedToMain: true routedAt: "2026-08-12" userConfirmation: pending userProposalSelection: pending implementationAuthorized: false experimentAuthorized: false patternRefs: - pattern-evidence/raw/C008.json - pattern-evidence/raw/C009.json firstDivergence: >- In C008 and C009, the implementation prompt combined the H04 readiness contract with frozen task text that instructed the Candidate to end with the legacy IMPLEMENTATION_COMPLETE sentinel. Each Candidate returned only IMPLEMENTATION_COMPLETE; the H04 readiness parser then deterministically returned missing-ready-readiness before implementation confirmation, focused-receipt binding, Review, or finalization. hypothesis: >- H04's implementation information flow exposes two mutually incompatible terminal contracts in one actor context, so a product-complete Candidate can follow one visible instruction and still be rejected by the controller. attribution: - Information Flow - Knowledge - Agent Responsibility support: - C008 / D1r2: focused pass, Judge 99 mean, workflow failed with missing-ready-readiness. - C009 / D1r3: focused pass, Judge 100 mean, workflow failed with the same error. - Both frozen Candidate final messages contain only IMPLEMENTATION_COMPLETE. alternatives: - Candidate exact-marker choice is the proximate behavior and remains actor responsibility. - Model stochasticity can determine which of the two visible instructions wins. - Product difficulty is disfavored for the terminal failure because both focused results were green. counterevidence: - C007 / D1r1 completed under the same task and H04 generation. - Other H04 cells emitted the H04 readiness markers and progressed. scope: >- Confirmed in two D1 replicates. It applies only where a composed implementation context contains conflicting terminal instructions; it does not explain D5/D8 product failures or C022's stale receipt. counterfactualPrediction: >- With one authoritative implementation terminal contract visible to the Candidate, a product-ready D1 trajectory should no longer end with only IMPLEMENTATION_COMPLETE and then fail missing-ready-readiness. Continued occurrence of the same mismatch would falsify the proposed mechanism. confidence: high rawRefs: - campaigns/cli-item-list-fields-matrixspec-harness-evolution-h04-formal-20260812-r1/runs/h04-formal-20260812-r1-cli-item-list-fields-r2-attempt-01/state.json#L18 - campaigns/cli-item-list-fields-matrixspec-harness-evolution-h04-formal-20260812-r1/runs/h04-formal-20260812-r1-cli-item-list-fields-r2-attempt-01/candidate-main-turn-12.final.txt#L1 - campaigns/cli-item-list-fields-matrixspec-harness-evolution-h04-formal-20260812-r1/runs/h04-formal-20260812-r1-cli-item-list-fields-r3-attempt-01/state.json#L18 - campaigns/cli-item-list-fields-matrixspec-harness-evolution-h04-formal-20260812-r1/runs/h04-formal-20260812-r1-cli-item-list-fields-r3-attempt-01/candidate-main-turn-12.final.txt#L1 - engine/drivers/matrixspec-evolution/harness.ps1#L776 - weaknessId: H04/h04-postformal-20260812-019ff55f-a/W-02 shortName: verification side effects invalidate controller product projection userConfirmation: pending userProposalSelection: pending implementationAuthorized: false experimentAuthorized: false patternRefs: - pattern-evidence/raw/C022.json firstDivergence: >- C022 entered Review with an accepted implementation and a passed, controller-anchored focused receipt. The Reviewer reran the hidden hook and pytest; those commands generated bytecode under tests/, a directory recursively included in productDiffPaths. The Review gate recomputed a different product projection and rejected the approved Review as stale. hypothesis: >- H04's verification boundary is not observationally pure relative to its product-projection boundary: verifier-generated transients can be treated as product drift and invalidate otherwise current evidence. attribution: - Verification - Artifact - Information Flow support: - C022 implementation was accepted with passed=true and product projection 468f... at revision epoch 0. - Review independently passed and returned Approved plus MATSPEC_STAGE_READY. - The frozen terminal error is controller-anchored receipt stale for current product projection. - The verification hook copies a test into tests/, runs pytest, and removes only the source .py while generated cache remains in the recursively hashed path. alternatives: - Reviewer chose to rerun verification, but independent verification is expected behavior rather than a product mutation. - A different product mutation could also produce projection drift; the frozen file set and timing identify generated test cache here. - This is not a Candidate product defect: focused and Judge results were green and Review approved. counterevidence: - C023 and C024 completed the same D6 task. - Other task projections did not necessarily include verifier-generated files or did not produce the same side effect. scope: >- Directly confirmed in C022. The reachable class is any task whose product projection recursively includes paths where verification creates or refreshes non-product artifacts. counterfactualPrediction: >- If verification is observationally pure with respect to the product projection, rerunning the same green verification during Review without a source change will preserve the anchored projection and the positive gate will accept the receipt. A projection change from verifier-only transients would falsify the mechanism change. confidence: high rawRefs: - campaigns/sqlmodel-field-constraints-matrixspec-harness-evolution-h04-formal-20260812-r1/runs/h04-formal-20260812-r1-sqlmodel-field-constraints-r1-attempt-01/state.json#L18 - campaigns/sqlmodel-field-constraints-matrixspec-harness-evolution-h04-formal-20260812-r1/runs/h04-formal-20260812-r1-sqlmodel-field-constraints-r1-attempt-01/stage-events.jsonl#L5 - campaigns/sqlmodel-field-constraints-matrixspec-harness-evolution-h04-formal-20260812-r1/runs/h04-formal-20260812-r1-sqlmodel-field-constraints-r1-attempt-01/process-artifacts/matspec/changes/AR20260808-sqlmodel-field-constraints-h04-formal-20260812-r1-sqlmodel-field-constraints-r1-attempt-01/focused-verification.json#L14 - campaigns/sqlmodel-field-constraints-matrixspec-harness-evolution-h04-formal-20260812-r1/runs/h04-formal-20260812-r1-sqlmodel-field-constraints-r1-attempt-01/candidate-review-turn-01.jsonl#L15 - packs/tasks/sqlmodel-field-constraints/task-dense-spec-v1.json#L24 - packs/tasks/sqlmodel-field-constraints/hooks/verify.ps1#L8 - engine/drivers/matrixspec-evolution/harness.ps1#L970 - engine/drivers/matrixspec-evolution/harness.ps1#L1375 - weaknessId: H04/h04-postformal-20260812-019ff55f-a/W-03 shortName: stable no-progress recovery consumes generic backtrack budget userConfirmation: pending userProposalSelection: pending implementationAuthorized: false experimentAuthorized: false patternRefs: - pattern-evidence/raw/C001.json - pattern-evidence/raw/C005.json - pattern-evidence/raw/C006.json - pattern-evidence/raw/C013.json - pattern-evidence/raw/C014.json - pattern-evidence/raw/C017.json firstDivergence: >- After a negative Review had established a stable blocker, the controller consumed review-to-implementation backtracks and scheduled another Implementation/verification/Review cycle even when the adjacent recovery epoch had the same controller-bound product projection. The trajectories then exhausted the generic four-backtrack ceiling. hypothesis: >- H04 Recovery lacks an explicit no-progress disposition for recurrence of the same blocker on the same product projection, so unchanged-state work can be rescheduled until the generic backtrack limit rather than being terminated or escalated with the stable blocker preserved. attribution: - Recovery - Workflow - Information Flow support: - C014 contains the strongest sequence: the same projection across epochs 0-3 with repeated blocker review. - C001, C005, C006, C013, and C017 each contain at least one adjacent recovery epoch with the same product projection and a recurring blocker. - The controller, not an autonomous Reviewer loop, consumed distinct review-to-implementation backtracks and scheduled each new cycle. - These runs terminate at maxBacktracks=4. alternatives: - The original product defect remains Candidate-owned and may be genuinely difficult. - A Review may add useful evidence even when the product projection is unchanged. - Reviewer blocker phrasing or repair-target precision remains Reviewer responsibility. - Some cycles include real product changes; C015 is excluded from unchanged-state support because every epoch changed. counterevidence: - C011, C012, C016, C023, and C024 converted findings into repair and positive closure. - Not every repeated Review is zero-information or controller waste. - Same-task H04 successes are absent for D5 and D8, so task difficulty remains a strong explanation for their ultimate product failures. scope: >- Only stable blocker plus unchanged controller-bound product projection across recovery epochs. It does not attribute the original defect to H04, does not include changed-projection repair attempts, and does not claim that raising the backtrack limit would improve quality. counterfactualPrediction: >- When the same blocker recurs on an unchanged product projection, a no-progress-aware recovery policy should produce an explicit non-looping disposition before scheduling another equivalent Candidate/Reviewer cycle. Another equivalent cycle ending only at the generic limit would falsify the mechanism change. confidence: medium rawRefs: - campaigns/axum-custom-executor-matrixspec-harness-evolution-h04-formal-20260812-r1/runs/h04-formal-20260812-r1-axum-custom-executor-r1-attempt-01/stage-events.jsonl#L5 - campaigns/bat-sanitize-matrixspec-harness-evolution-h04-formal-20260812-r1/runs/h04-formal-20260812-r1-bat-sanitize-r2-attempt-01/stage-events.jsonl#L5 - campaigns/prometheus-utf8-negotiation-matrixspec-harness-evolution-h04-formal-20260812-r1/runs/h04-formal-20260812-r1-prometheus-utf8-negotiation-r2-attempt-01/stage-events.jsonl#L5 - experiments/matrixspec-harness-evolution-h04-formal/tasks/bat-sanitize/experiment.json#L19 - weaknessId: H04/h04-postformal-20260812-019ff55f-a/W-04 shortName: cross-stage authority conflict is returned to a downstream artifact userConfirmation: confirmed userConfirmedAt: "2026-08-12" userConfirmedInterpretation: >- The empty product patch, absent --sanitize option, five focused failures at CLI argument parsing, and Judge scores 3/0 are downstream consequences. The first divergence is the earlier delta-artifact versus tasks/Operator cat-mode authority conflict and its repeated unresolved return to tasks. userProposalSelection: pending implementationAuthorized: false experimentAuthorized: false patternRefs: - pattern-evidence/raw/C004.json firstDivergence: >- C004's implementation readiness identified a conflict between confirmed upstream delta artifacts and the tasks/Operator instruction for cat-mode behavior. Recovery returned to tasks four times. The first three accepted the same task artifact; the final task artifact recorded an override, but the contradictory confirmed delta-spec and delta-design remained intact. hypothesis: >- H04 Recovery can route an authority conflict to a downstream stage that cannot reconcile the still-confirmed upstream source, without an explicit path to establish precedence or repair the earliest conflicting artifact. attribution: - Recovery - Artifact - Knowledge - Agent Responsibility support: - C004 has four implementation-to-tasks backtracks with the same blocker fingerprint. - The first three reaccepted task digests were identical. - Operator consistently required interactive-only sanitization and cat byte-through; this is not an Operator defect finding. - The final tasks override coexisted with contradictory confirmed upstream deltas and the run terminated at max-backtracks. - The final product-code.diff is effectively empty, so no --sanitize implementation was produced. - All five focused tests failed at CLI parsing because --sanitize did not exist; the two Judges scored the result 3 and 0. alternatives: - Formal precedence among user-confirmed artifacts and later Operator correction is itself ambiguous. - A different Candidate could have requested a farther back target or interpreted the correction differently. - This is a single-cell mechanism and does not explain the D5 product defects in C005/C006. counterevidence: - C012 eventually completed despite a related parser/backtrack loop. - Other H04 cells resolved artifact or verification findings and completed. scope: >- One user-confirmed C004 pre-implementation authority-conflict trajectory. The within-cell responsibility chain is directly evidenced, but there is no cross-cell replication and no claim that this is a general H04 causal effect. counterfactualPrediction: >- For another cross-stage contradiction, recovery should resolve or explicitly prioritize the earliest conflicting source before downstream re-entry. Recurrence of the same blocker while upstream artifact digests remain authoritative and unchanged would falsify the mechanism change. confidence: medium rawRefs: - campaigns/bat-sanitize-matrixspec-harness-evolution-h04-formal-20260812-r1/runs/h04-formal-20260812-r1-bat-sanitize-r1-attempt-01/stage-events.jsonl#L4 - campaigns/bat-sanitize-matrixspec-harness-evolution-h04-formal-20260812-r1/runs/h04-formal-20260812-r1-bat-sanitize-r1-attempt-01/terminal-receipt.json#L9 - campaigns/bat-sanitize-matrixspec-harness-evolution-h04-formal-20260812-r1/runs/h04-formal-20260812-r1-bat-sanitize-r1-attempt-01/product-code.diff#L1 - campaigns/bat-sanitize-matrixspec-harness-evolution-h04-formal-20260812-r1/runs/h04-formal-20260812-r1-bat-sanitize-r1-attempt-01/tests.log#L5 - campaigns/bat-sanitize-matrixspec-harness-evolution-h04-formal-20260812-r1/runs/h04-formal-20260812-r1-bat-sanitize-r1-attempt-01/tests.log#L160 - experiments/matrixspec-harness-evolution-h04-formal/reports/generation-result.json#L250 notHarnessAttributable: - >- Candidate product defects in C001, C005, C006, C013, C014, C015, and C017 are not H04 SDD evolution effects merely because they persisted or occurred in concentrated D5/D8 failures. - >- Reviewer overchecking, vague blocker wording, or failure to communicate a repair target remains Reviewer behavior unless a controller scheduling or handoff failure is independently evidenced. - >- D5 and D8 have no same-task H04 success; task-specific difficulty and public-versus-hidden contract gaps remain major alternatives. - >- The H04 batch satisfied false-terminal=0, but H03 also observed 0; this is a safety result, not evidence that H04 caused an uplift. coverageGaps: - Some D5/D8 hidden-hook expectations cannot be classified as publicly grounded Candidate omissions from the frozen Pattern layer alone. - H04 formal evidence can support the mechanisms above but cannot validate the effect of any Proposal without a new controlled intervention.