schemaVersion: 1
miner:
  model: gpt-5.6-sol
  reasoningEffort: high
coverage:
  expected: 24
  included: 24
  excluded: 0
  cells_considered:
  - C001
  - C002
  - C003
  - C004
  - C005
  - C006
  - C007
  - C008
  - C009
  - C010
  - C011
  - C012
  - C013
  - C014
  - C015
  - C016
  - C017
  - C018
  - C019
  - C020
  - C021
  - C022
  - C023
  - C024
batch_assessment:
  outcome_summary: The frozen denominator contains 24 valid_behavior cells and 48 valid Blind Judge outputs. Seventeen workflows
    completed and seven failed; 19 cells passed focused tests and five did not. Judge scores range from 60 to 100 with a mean
    of 94.48. All seven workflow failures reached revisionEpoch 4, while 13 runs used back/recheck and six still completed
    after backtracking, so recovery is often effective and a divergence is not necessarily terminal.
  pattern_summary: PT-01 has 18 independent occurrences in 11 cells spanning seven tasks; six of those 11 workflows failed,
    but five completed, and several retained passing focused tests or perfect Judge scores. PT-02 has two independent occurrences
    in C001 and C004, both workflow failures. PT-03 contains five duplicate runtime emissions across three cells but explicitly
    has independentBehaviorOccurrenceCount=0. PT-04 occurs once in C013. Thirteen cells have no PT-01, 22 have no PT-02, and
    23 have no PT-04. Frequency is therefore treated as support for recurrence, not as causal proof.
  causal_boundary: This is observational H02-only attribution. The known infrastructure-invalid attempts and non-formal smoke
    are excluded, and no prior H0/H1 Pattern or Weakness is used as an H02 fact. The normalized evidence reports 192 normalized
    and nine dropped references; none of the First Divergence anchors below relies on a dropped reference. The evidence can
    support falsifiable Harness-strategy hypotheses but cannot distinguish a Harness cause from task difficulty or model stochasticity
    without a frozen intervention.
weaknesses:
- weakness_id: W-01
  pattern_refs:
  - PT-01
  first_divergence: 'C005/S005 is the earliest divergence in that trajectory: the confirmed obligation explicitly included
    U+FEFF at raw-sessions/candidate/2026/08/11/rollout-2026-08-11T15-29-12-019fefb9-b509-7e00-b2ae-e0ca0413cd02.jsonl#L58,
    but S005 reported all implementation tasks and verification complete at raw-sessions/candidate/2026/08/11/rollout-2026-08-11T15-59-12-019fefd5-2a8d-7950-9684-0778361d5f5b.jsonl#L254
    while the leading-U+FEFF case remained wrong; the omission became explicit only in review at raw-sessions/reviewer/2026/08/11/rollout-2026-08-11T16-33-41-019feff4-bb33-7a11-a980-324189aafbee.jsonl#L76.'
  hypothesis: The H02 implementation-readiness strategy can treat aggregate task-complete and passing-suite signals as sufficient
    even when a confirmed obligation lacks direct behavioral or artifact evidence. This shifts discovery of already-specified
    gaps into review and recovery. It plausibly increases backtracking and recovery exhaustion, but it does not imply that
    every such omission causes terminal failure.
  attribution:
  - Workflow
  - Information Flow
  - Verification
  support:
  - PT-01 recurs in 11 of 24 cells across seven of eight tasks, with 18 independent occurrences.
  - 'The same sequence appears with explicit upstream obligations: C004/S005 missed valid UTF-8 C1 behavior, C005/S005 missed
    the leading-U+FEFF case, C014/S006 missed the confirmed U__ encoding, and C016/C018 reported readiness before required
    typing or focused-matrix boundaries were closed.'
  - 'Six PT-01 cells failed workflow completion: C001, C004, C005, C008, C014, and C016. Only one of the 13 non-PT-01 cells,
    C017, failed workflow completion.'
  - 'The association is strongest for late review discovery and recovery cost, not final artifact quality: C001, C004, C005,
    and C008 failed workflow completion despite passing focused tests, and C001 and C008 received 100/100 Judge scores.'
  alternatives:
  - 'handoff_loss: C001 contains a genuine contract mismatch between confirmed documents and a downstream consumer, so that
    occurrence may be a handoff or authority-conflict problem; the same explanation does not cover the explicit obligations
    missed in C004, C005, C014, C016, and C018.'
  - 'upstream_artifact_quality: imprecise or inconsistent upstream artifacts remain plausible for C001 and the advisory artifact
    warning in C024, but several other occurrences cite precise accepted behavior or test obligations before implementation.'
  - 'downstream_instruction_quality: incomplete implementation instructions could explain isolated misses, but multiple affected
    cells had task-level statements naming the missing behavior or verification matrix.'
  - 'task_specific_difficulty: the affected tasks contain edge-heavy Unicode, protocol, CLI, and typing requirements, and
    the entire pytest-plugin-entry-points task has no PT-01 occurrence; this materially limits generalization.'
  - 'model_stochasticity: same-task counterexamples such as C002/C003 versus C001, C006 versus C004/C005, and C007/C009 versus
    C008 show that occurrence is stochastic rather than inevitable.'
  - 'not_harness_attributable: the underlying coding or testing mistakes are Candidate behavior. The Harness-attributed hypothesis
    is limited to the strategy that recognizes implementation readiness despite missing obligation-level evidence.'
  counterevidence:
  - 'PT-01 is not necessarily terminal: C011, C012, C013, C018, and C024 completed their workflows.'
  - C011 and C012 completed with focused tests passing and 100/100 Judge scores after later correction or review, demonstrating
    successful recovery.
  - C018 completed with focused tests passing and Judge scores 88/97 despite several PT-01 occurrences.
  - C024 completed with focused tests passing and 100/100 Judge scores; its PT-01 occurrence concerns advisory workflow-check
    wording and supplies little evidence of outcome harm.
  - C002, C003, C006, C007, C009, C010, C019, C020, C021, C022, and C023 completed without PT-01, showing that the same Harness
    often achieves clean readiness.
  - C015 completed without PT-01 but still failed focused tests, while C017 failed workflow completion without PT-01, so PT-01
    is neither necessary for poor outcomes nor sufficient for failure.
  scope: Supported only for H02 implementation-readiness transitions where an obligation was already explicit and review later
    established its absence. The strongest support spans bat-sanitize, prometheus-utf8-negotiation, and pytest-addini-type-expressions;
    the weaker artifact-only C024 occurrence is not used as causal outcome support. The hypothesis does not attribute the
    underlying product defect to the Harness and does not claim that readiness should imply final correctness.
  counterfactual_prediction: If the related readiness strategy changes, the first observable difference should be fewer implementation-ready
    emissions followed immediately by a first-review discovery of an already-explicit obligation. The missing evidence should
    instead be surfaced before readiness or readiness should remain pending; Judge scores or final focused-test outcomes need
    not change first.
  confidence: medium
  raw_refs:
  - raw-sessions/candidate/2026/08/11/rollout-2026-08-11T15-29-12-019fefb9-b509-7e00-b2ae-e0ca0413cd02.jsonl#L58
  - raw-sessions/candidate/2026/08/11/rollout-2026-08-11T15-59-12-019fefd5-2a8d-7950-9684-0778361d5f5b.jsonl#L254
  - raw-sessions/reviewer/2026/08/11/rollout-2026-08-11T16-33-41-019feff4-bb33-7a11-a980-324189aafbee.jsonl#L76
  - raw-sessions/candidate/2026/08/11/rollout-2026-08-11T16-05-23-019fefda-d413-7723-8474-fea3b945f44e.jsonl#L139
  - raw-sessions/reviewer/2026/08/11/rollout-2026-08-11T16-20-54-019fefe9-09fa-7112-8b65-6ddee349f31d.jsonl#L40
  - raw-sessions/candidate/2026/08/11/rollout-2026-08-11T12-29-12-019fef14-e600-72c0-91c4-99caac45036e.jsonl#L118
  - raw-sessions/candidate/2026/08/11/rollout-2026-08-11T12-50-54-019fef28-c6b5-7412-ac23-67139ff80e93.jsonl#L218
  - raw-sessions/reviewer/2026/08/11/rollout-2026-08-11T13-00-19-019fef31-6544-7fa3-a083-7ddd65af1e12.jsonl#L59
  - raw-sessions/candidate/2026/08/11/rollout-2026-08-11T11-57-43-019feef8-14bf-7a83-838d-627b65602480.jsonl#L101
  - raw-sessions/reviewer/2026/08/11/rollout-2026-08-11T12-00-33-019feefa-adb9-7ec3-82f9-91d46fe694f6.jsonl#L95
- weakness_id: W-02
  pattern_refs:
  - PT-02
  first_divergence: 'C001/S017 is the first repeated-state divergence: S015 had already established the execute-versus-spawn
    blocker and 11/11 focused result at raw-sessions/reviewer/2026/08/11/rollout-2026-08-11T16-03-44-019fefd9-4f08-7b01-a250-6d86935c04d2.jsonl#L45;
    without an intervening material implementation change, S017 reported the same blocker again at raw-sessions/reviewer/2026/08/11/rollout-2026-08-11T16-09-12-019fefde-5115-7f22-9a11-92c8912468eb.jsonl#L74.
    S019 repeated it again at raw-sessions/reviewer/2026/08/11/rollout-2026-08-11T16-15-20-019fefe3-f10d-7c23-a2ed-285e2cf1e2e5.jsonl#L69.'
  hypothesis: The H02 recovery strategy can re-enter review or blocker investigation before a material repair or new evidence
    state is visible. In a bounded recovery workflow, unchanged-state review consumes convergence headroom and can contribute
    to terminal workflow failure even when focused tests and final judged quality remain strong.
  attribution:
  - Recovery
  - Information Flow
  - Agent Responsibility
  support:
  - 'PT-02 has two independent occurrences across two tasks: C001 and C004.'
  - Both PT-02 cells failed workflow completion at revisionEpoch 4 after four backtracks.
  - C001 repeated the same contract blocker across S015, S017, and S019 before a later review added a distinct compatibility
    finding.
  - C004 repeated the same valid UTF-8 C1 investigation across S010 and S011 before later implementation and review states
    changed.
  - The audit records 41 total backtracks and a bounded maximum revision epoch of four, so an unchanged-state review has observable
    opportunity cost.
  alternatives:
  - 'handoff_loss: C001''s conflicting operation names make handoff loss or authority conflict a strong alternative for that
    cell; C004''s explicit C1 requirement does not share the same ambiguity.'
  - 'upstream_artifact_quality: the C001 contract conflict may originate upstream, but C004''s repeated investigation followed
    a concrete focused reproduction rather than an unclear artifact.'
  - 'downstream_instruction_quality: repair or review instructions may have failed to communicate what state change was required;
    the normalized evidence establishes repetition but not the exact instruction content received at every controller boundary.'
  - 'task_specific_difficulty: both tasks involve difficult protocol or byte-level behavior, which can justify independent
    revalidation and limits the inference from two cells.'
  - 'model_stochasticity: PT-02 is absent from 22 cells and from other repetitions of both affected tasks, making stochastic
    reviewer behavior a substantial explanation.'
  - 'not_harness_attributable: the repeated inspection may be an Agent choice. Harness attribution is limited to the recovery
    sequencing that allowed another review state without visible repair evidence.'
  counterevidence:
  - PT-02 is absent in 22 of 24 cells, including many cells with review or recovery work.
  - Five workflow failures—C005, C008, C014, C016, and C017—occurred without PT-02, so it is not necessary for terminal failure.
  - The audit reports 13 back/recheck runs and six completions after backtracking, demonstrating that recovery frequently
    succeeds.
  - C007, C009, C011, C012, and C018 completed after one or more backtracks without PT-02.
  - C001 passed focused tests and received 100/100 Judge scores; C004 also passed focused tests and received 94/89. Repetition
    therefore did not imply artifact collapse.
  - C004 later repaired the initial UTF-8 C1 issue and encountered a distinct route-specific gap, so the repeated investigation
    cannot be treated as the sole terminal cause.
  scope: Supported only as a low-confidence H02 recovery-efficiency weakness in two cells and two tasks. It concerns unchanged-state
    review sequencing, not the correctness of independent revalidation in general. It does not establish that additional recovery
    would necessarily complete either workflow.
  counterfactual_prediction: If the related recovery strategy changes, the first observable difference should be the disappearance
    of consecutive review verdicts that restate the same blocker and test result without an intervening material state change.
    A repair, new evidence state, or distinct recovery decision should appear before the next review; workflow-completion
    rates would be a later outcome.
  confidence: low
  raw_refs:
  - raw-sessions/reviewer/2026/08/11/rollout-2026-08-11T16-03-44-019fefd9-4f08-7b01-a250-6d86935c04d2.jsonl#L45
  - raw-sessions/reviewer/2026/08/11/rollout-2026-08-11T16-09-12-019fefde-5115-7f22-9a11-92c8912468eb.jsonl#L74
  - raw-sessions/reviewer/2026/08/11/rollout-2026-08-11T16-15-20-019fefe3-f10d-7c23-a2ed-285e2cf1e2e5.jsonl#L69
  - raw-sessions/reviewer/2026/08/11/rollout-2026-08-11T16-20-54-019fefe9-09fa-7112-8b65-6ddee349f31d.jsonl#L40
  - raw-sessions/candidate/2026/08/11/rollout-2026-08-11T16-23-18-019fefeb-3af3-7652-b7cd-6632c22f3a9d.jsonl#L49
  - raw-sessions/reviewer/2026/08/11/rollout-2026-08-11T16-26-09-019fefed-d534-7492-a9a8-0d82e0eb37a7.jsonl#L40
unattributed_patterns:
- pattern_refs:
  - PT-03
  classification: not_harness_attributable
  reason: PT-03 is explicitly normalized as duplicate runtime emission with five raw log candidates but independentBehaviorOccurrenceCount=0.
    It is not counted as Agent behavior cost and has mixed outcomes across C003, C014, and C018, so it supplies no behavioral
    Harness weakness.
  support:
  - The aggregate provenance records rawCandidateCount=5 and independentBehaviorOccurrenceCount=0.
  - The candidates are byte-identical or adjacent duplicate message emissions without an independent state transition.
  counterevidence:
  - C003 and C018 completed their workflows and passed focused tests despite PT-03.
  - C014 failed, but its substantive PT-01 obligation miss independently explains the observed review failure.
  - Twenty-one cells contain no PT-03, and the three occurrence cells have mixed workflow, focused-test, and Judge outcomes.
  raw_refs:
  - raw-sessions/candidate/2026/08/11/rollout-2026-08-11T15-29-46-019fefba-372b-77b3-9490-8ef301524b64.jsonl#L12
  - raw-sessions/candidate/2026/08/11/rollout-2026-08-11T15-29-46-019fefba-372b-77b3-9490-8ef301524b64.jsonl#L13
  - raw-sessions/candidate/2026/08/11/rollout-2026-08-11T16-06-14-019fefdb-99f6-7c92-8df8-d76a1435fc98.jsonl#L12
  - raw-sessions/candidate/2026/08/11/rollout-2026-08-11T16-06-14-019fefdb-99f6-7c92-8df8-d76a1435fc98.jsonl#L13
  - raw-sessions/reviewer/2026/08/11/rollout-2026-08-11T16-13-37-019fefe2-5ee2-7b81-8a70-03943f22ee1e.jsonl#L12
  - raw-sessions/reviewer/2026/08/11/rollout-2026-08-11T16-13-37-019fefe2-5ee2-7b81-8a70-03943f22ee1e.jsonl#L13
  - raw-sessions/candidate/2026/08/11/rollout-2026-08-11T12-50-54-019fef28-c6b5-7412-ac23-67139ff80e93.jsonl#L240
  - raw-sessions/candidate/2026/08/11/rollout-2026-08-11T12-50-54-019fef28-c6b5-7412-ac23-67139ff80e93.jsonl#L241
  - raw-sessions/candidate/2026/08/11/rollout-2026-08-11T11-03-35-019feec6-84a8-7351-939e-13e53e7919be.jsonl#L12
  - raw-sessions/candidate/2026/08/11/rollout-2026-08-11T11-03-35-019feec6-84a8-7351-939e-13e53e7919be.jsonl#L13
- pattern_refs:
  - PT-04
  classification: unresolved
  reason: The authorization-only child segment in C013 produced no local execution or evidence, but it is a single occurrence,
    the parent workflow later made productive progress and completed, and the batch does not establish that this segment caused
    the failed focused outcome or lower Judge scores.
  support:
  - C013/S006 contains startup and authorization events but no execution, artifact change, evidence generation, or task-state
    closure.
  - C013 is the only occurrence among 24 cells.
  counterevidence:
  - C013 ultimately completed its workflow.
  - Later C013 implementation and review segments produced repairs, 97 passing tests, and an approved review.
  - The substantive C013 PT-01 cross-adapter coverage omission is a more direct explanation for its recovery activity and
    weak focused/Judge outcomes.
  - Twenty-three cells do not contain PT-04, including both strong and weak outcomes.
  raw_refs:
  - raw-sessions/candidate/2026/08/11/rollout-2026-08-11T12-52-11-019fef29-f134-7501-bb20-674d7492a50d.jsonl#L2
  - raw-sessions/candidate/2026/08/11/rollout-2026-08-11T12-52-11-019fef29-f134-7501-bb20-674d7492a50d.jsonl#L10
  - raw-sessions/candidate/2026/08/11/rollout-2026-08-11T12-52-11-019fef29-f0f9-7803-9999-8ca38f9fe478.jsonl#L35
  - raw-sessions/candidate/2026/08/11/rollout-2026-08-11T12-52-11-019fef29-f0f9-7803-9999-8ca38f9fe478.jsonl#L52