schemaVersion: 1 parentHarness: H04 analysisRunId: h04-postformal-20260812-019ff55f-a provenance: agent-derived userConfirmation: pending userProposalSelection: pending implementationAuthorized: false experimentAuthorized: false proposals: - proposalId: H04/h04-postformal-20260812-019ff55f-a/P-01 sourceWeakness: H04/h04-postformal-20260812-019ff55f-a/W-01 decisionReadiness: decision-ready evolutionDiscussionAlias: H052 routedToMain: true routedAt: "2026-08-12" userConfirmation: pending userProposalSelection: pending implementationAuthorized: false experimentAuthorized: false evidence: summary: >- C008 and C009 completed the D1 product work and passed focused tests but returned the legacy IMPLEMENTATION_COMPLETE sentinel exposed in their frozen task context. The H04 parser required a different three-part readiness contract and rejected both before Review. affectedTasks: [cli-item-list-fields] affectedRuns: [C008, C009] observedImpact: quality: Product evidence was green, so the workflow failure did not track product quality. completion: Two D1 replicates stopped before implementation acceptance and Review. cost: Candidate implementation work and focused verification did not reach formal closure. stability: The exact failure repeated in two of three D1 replicates; C007 is a success counterexample. rootCause: hypothesis: One actor context exposes incompatible implementation terminal contracts. confidence: high strategyChange: currentStrategy: >- Controller readiness instructions and frozen task completion text can each define a different terminal marker, while the parser accepts only the H04 readiness form. proposedStrategy: >- Make the controller-owned implementation readiness contract the single authoritative terminal contract presented at execution time; inherited task content may describe completion obligations but must not introduce a competing terminal signal. rationale: >- The single lever is terminal-contract ownership, not Candidate prompting style, test content, product implementation, or parser leniency. expectedEffect: behavioralChange: Candidate reports READY or BACK through one visible, unambiguous controller contract. targetedPatternChange: The legacy-sentinel-versus-H04-parser mismatch disappears. expectedMetrics: quality: No uplift is assumed; watch for product-quality regression separately. completion: Product-ready cells should no longer fail solely on missing-ready-readiness caused by a competing sentinel. cost: Avoids discarding completed implementation/verification work at the readiness boundary. stability: Same execution context should not expose two terminal contracts. rejectionConditions: - Product-ready runs still return only IMPLEMENTATION_COMPLETE and fail missing-ready-readiness. - The old sentinel disappears but another competing marker or prose contract remains reachable. - The controller accepts readiness without the required obligation-to-evidence account, creating a false terminal. - A completion or quality regression is attributable to the new terminal-contract ownership. regressionRisks: - Over-normalizing terminal output could accept incomplete self-reports. - Removing task text without preserving human-readable completion obligations could weaken execution guidance. - proposalId: H04/h04-postformal-20260812-019ff55f-a/P-02 sourceWeakness: H04/h04-postformal-20260812-019ff55f-a/W-02 userConfirmation: pending userProposalSelection: pending implementationAuthorized: false experimentAuthorized: false evidence: summary: >- C022 had a passed controller-anchored receipt and Approved Review, but Review verification generated files within the recursively hashed tests projection. The positive gate then rejected the receipt as stale. affectedTasks: [sqlmodel-field-constraints] affectedRuns: [C022] observedImpact: quality: Focused tests and both Judges were green; no product defect explains the terminal failure. completion: Review approval could not be accepted. cost: Independent Review verification converted valid evidence into a stale receipt. stability: Directly observed once; same-task C023/C024 completed. rootCause: hypothesis: Verifier-generated transients are included in the controller's product identity. confidence: high strategyChange: currentStrategy: >- Product projection recursively includes configured product paths even when verification creates transient files under those paths. proposedStrategy: >- Define controller product identity over intended product artifacts only, so running an authorized verifier is observationally pure with respect to the projection unless it actually changes an intended product artifact. rationale: >- The single lever is product-projection semantics; it does not weaken the requirement that a real post-receipt product change invalidates evidence. expectedEffect: behavioralChange: Rerunning verification without a product-source change preserves the anchored product identity. targetedPatternChange: Reviewer-created cache/transient files no longer trigger a stale-receipt failure. expectedMetrics: quality: Real product drift must remain blocked; no score uplift is assumed. completion: A green approved Review can close when only verifier transients changed. cost: Avoids rerunning otherwise current implementation evidence. stability: Product identity remains stable across repeated observational verification. rejectionConditions: - Verification-only transients still change the product projection. - The policy excludes an intended product artifact and accepts genuinely stale evidence. - The stale error is replaced by another equivalent projection mismatch. - A completion or quality regression is attributable to weakened real-drift detection. regressionRisks: - An over-broad exclusion could hide real generated artifacts that are part of the delivered product. - Task-specific product boundaries may be incomplete or ambiguous. - proposalId: H04/h04-postformal-20260812-019ff55f-a/P-03 sourceWeakness: H04/h04-postformal-20260812-019ff55f-a/W-03 userConfirmation: pending userProposalSelection: pending implementationAuthorized: false experimentAuthorized: false evidence: summary: >- Across C001, C005, C006, C013, C014, and C017, controller-scheduled recovery sometimes repeated a blocker on an unchanged bound product projection and continued until the generic four-backtrack ceiling. affectedTasks: [axum-custom-executor, bat-sanitize, prometheus-utf8-negotiation, pytest-addini-type-expressions] affectedRuns: [C001, C005, C006, C013, C014, C017] observedImpact: quality: Original defects remained Candidate-owned; no automatic quality uplift is inferred. completion: Stable no-progress cycles consumed the remaining recovery budget and ended at max-backtracks. cost: Equivalent Candidate verification and Review work was rescheduled on unchanged product state. stability: The reachable pattern appears across four tasks, with successful repair counterexamples. rootCause: hypothesis: Recovery has no explicit disposition for same blocker plus same product projection. confidence: medium strategyChange: currentStrategy: >- Every negative Review can consume another generic backtrack even when the blocker and controller-bound product state have not changed. proposedStrategy: >- Treat recurrence of the same blocker on the same product projection as an explicit no-progress outcome that must terminate or escalate with preserved blocker evidence rather than schedule an equivalent cycle. rationale: >- The single lever is no-progress disposition. It does not raise the backtrack limit, auto-approve a defect, or transfer product ownership. expectedEffect: behavioralChange: Stable blocker recurrence produces a non-looping disposition before another equivalent implementation/review cycle. targetedPatternChange: Adjacent recovery epochs no longer repeat unchanged product state and blocker until max-backtracks. expectedMetrics: quality: Defects remain blocked; watch for premature terminal classification. completion: No generic uplift is predicted because some tasks are genuinely unsolved. cost: Equivalent unchanged-state recovery cycles should disappear. stability: The disposition should behave consistently across task types without converting failures into false terminals. rejectionConditions: - Same blocker and same projection still schedule an equivalent cycle. - The loop is merely renamed or shifted to another stage. - Legitimate state-changing repairs are prematurely terminated. - A false terminal or attributable quality/completion regression appears. regressionRisks: - Blocker identity may be too coarse and collapse materially different findings. - A no-progress stop may reduce useful independent corroboration. - proposalId: H04/h04-postformal-20260812-019ff55f-a/P-04 sourceWeakness: H04/h04-postformal-20260812-019ff55f-a/W-04 sourceWeaknessConfirmation: confirmed generationStatus: generated_after_user_weakness_confirmation userConfirmation: pending userProposalSelection: pending implementationAuthorized: false experimentAuthorized: false evidence: summary: >- C004 repeatedly returned an upstream-versus-downstream cat-mode contract conflict to tasks. The final tasks recorded an override while the contradictory confirmed delta artifacts remained authoritative. After four backtracks, the run ended with an effectively empty product patch: --sanitize did not exist, all five focused tests failed during CLI argument parsing, and the two Judges scored the result 3 and 0. affectedTasks: [bat-sanitize] affectedRuns: [C004] observedImpact: quality: No product implementation was produced; Judge scores were 3 and 0. completion: Four unresolved implementation-to-tasks backtracks ended at max-backtracks before implementation. cost: Task-stage work repeated without eliminating the conflicting upstream authority. stability: Single-cell evidence only. rootCause: hypothesis: Recovery can target a downstream artifact that cannot resolve the earliest authoritative conflict. confidence: medium strategyChange: currentStrategy: >- A cross-stage conflict may be routed to the requested downstream stage without resolving source precedence or invalidating the conflicting upstream artifact. proposedStrategy: >- Treat cross-stage authority conflict as a distinct recovery outcome: route repair to the earliest conflicting authoritative source and allow downstream re-entry only after the conflict is removed or an explicit precedence decision is recorded. rationale: >- The single lever is authority-conflict resolution, not a fixed preference for Operator, tasks, delta-spec, or delta-design in every case. expectedEffect: behavioralChange: >- A Candidate that reports conflicting confirmed artifacts receives one authority-resolving recovery target, rather than repeated downstream tasks rewrites. targetedPatternChange: >- The same IMPLEMENTATION_READINESS blocker fingerprint does not recur while the upstream conflicting artifacts remain authoritative and unchanged. expectedMetrics: quality: >- No numeric uplift is predicted; the next implementation must at least produce the public --sanitize surface before focused behavioral quality is judged. completion: The authority conflict should either close before implementation or terminate explicitly without four equivalent backtracks. cost: Avoids repeated downstream rewrites that cannot settle upstream authority. stability: Cross-stage precedence remains auditable across recovery. rejectionConditions: - The same conflict recurs after downstream re-entry with unchanged upstream authority. - The strategy hard-codes an authority order that contradicts a current user-confirmed decision. - The conflict is hidden rather than resolved. - The workflow advances but still produces an empty patch or lacks the required --sanitize CLI surface. - A false terminal or attributable regression appears. regressionRisks: - Escalating farther back may increase work for conflicts a downstream clarification could resolve. - Incorrect precedence could overwrite a later valid user correction.