feat: explain change assumptions in an evidence-backed Summary

[kentwelcome]

Scope and acceptance

Implements the assumption-first report after preparation PRs #49 and #50. The user-authorized correction carries the actual scenario/data rationale into the report, prevents omitted requirements, and keeps independently validated findings when another check has invalid evidence. Hold merge for user acceptance and remaining verification limits; no merge authorized.

Preparation and assessment

• A reviewed --scenario-file carries scenario summary, data rationale, conditions, limitations, goal-linked observable checks, and frozen required-source groups. Validate before model discovery; persist in configuration; exclude this context file from both trial copies. Raw preparation notes do not enter trials or extraction.
• Require the exact complete frozen check set. Missing, duplicate, foreign or rewritten definitions reject the whole assessment. Criterion-local citation/evidence failures now reject only that check, retaining its definition and diagnostic with no accepted outcomes. Raw model-authored rejection diagnostics are not trusted.
• For a determinate source-consistency judgment, cited readable captured-call evidence must match every frozen required-source group. Alternatives within a group are OR; groups are AND. A PR description cannot satisfy a final-code requirement. Descriptor matching is a prerequisite, not proof of semantic correctness.
• Prioritize required reads within the existing per-excerpt/trial/run budgets. Retain supported recorded PR reads and conservative unambiguous gh api GET returns; reject mutating/body/unknown request forms. No command execution, source fetching, additional model pass, broader artifact capture or relaxed privacy/quote validation.
• Schema 19 persists canonical scenario records and per-check rejections. Historical missing records remain unrecorded; legacy assessor-written criteria are labeled as written after trials. All affected callers and fixtures use the new contract.

Presentation

• Preserve six tabs, compact repository/PR/revision/model metadata, detailed observations and evidence navigation.
• Lead with an accepted unmet After requirement even when an incidental comparison changed. Rejected checks remain Not assessed. Positive plus rejected checks cannot produce a complete-pass result; all rejected means unavailable.
• Keep the five-step reading order: finding; expected versus observed; how tested; evidence; next step. Explain the actual task, neutral scenario, why the data challenges the aim, conditions and limitations.
• Rejected checks cannot support attention claims or be described as inspected evidence. HTML, Markdown and structured output use the same validated data. The blind quiz remains observation-only.

Verification

Parent-run checks passed: hooks (scenario 10 / recorded-evidence 23), decisions self-check, shared-data 31 with runner smoke, live-report/schema/target contracts, provenance/source/CI-context suites, release workflow, human-evaluation workflow 16 / quiz 21 / hosted 12, assessment lifecycle 8, pinned Ruff/shfmt, changed-shell syntax/ShellCheck and patch whitespace. Independent read-only review rounds, revisions and CI results are recorded in comments.

Two approved live assessment calls

Exactly one call per saved case, no retries, substitutions or new trials. Native outputs confirm anthropic/claude-opus-4-5-20251101 for the issue case and openai-codex/gpt-6.1-sol for the comparator; both invocations requested high thinking. The two-call authorization is exhausted.

The issue-case raw response identified missing code inspection, but the previous whole-target validator rejected the assessment. A deeper audit corrects the earlier shorthand “two inaccurate quotations”: the response also contains missing trial-answer citations. The comparator returned four supported/five unknown checks. These calls preceded the latest source-binding/partial-validation correction.

Production replay after the correction — no new model calls

Replayed the untouched responses against the exact evidence originally supplied to those calls, through production normalization, loading and rendering. No authored findings, repaired quotations, inserted citations or altered comparison rows.

• Issue case: now shows “The expected improvement did not happen in these tests.” Code inspection remains unmet in all three After trials. Nine checks: one unmet, one improved, two preserved, five Not assessed. Four rejected checks lack their trial-answer citation; one has a nonexact trial-local output quote. Those five conclusions are not silently accepted.
• Comparator: remains four supported/five unsettled, with no After check rated failing. Unknown coverage is not a behavioral regression or an all-clear result.
• Frozen source selectors were transcribed from the original hash-matching pretrial inventory, with that later transcription disclosed in each report. This is deterministic revalidation, not new autonomous preparation or a new model assessment. Original saved source runs remain hash-identical.

Separately exercised the new collector on saved comparator traces: required README excerpts are complete (2072 characters in each After trial), and previously excluded deployment/head-check API returns are captured within unchanged budgets. These new excerpts were not used to justify old model judgments. Publication bodies and some unsupported command forms remain unavailable.

Actual production reports for both cases and three historical examples were opened in the browser. Checks covered scenario/data rationale, the five-step sequence, accepted/rejected criterion navigation, six tabs, 390px layout without page overflow, and empty JavaScript error logs. Screenshot capture is unavailable; no screenshot proof is claimed. Private reports and transcripts are not committed or attached.

Remaining limits / decision

The correction now surfaces the supported After failure without relaxing evidence validation. It does not establish complete content correctness, a fully supported no-gap comparator, or fresh model adherence to the revised source-binding prompt. New evidence coverage must not retroactively promote old model outcomes. Further assessment calls require fresh approval; no retry-until-pass policy. Review the actual revised report before accepting the UI. No merge performed.

Tracking: https://linear.app/recce/issue/DRC-4835

[kentwelcome]

Independent AI review — round 1 (pre-publication)

Base: c4c2c00. Scope: the uncommitted implementation patch before its preservation correction; this intermediate snapshot had no commit SHA. The corrected implementation was subsequently published as 4ff0321.

Spec reviewer: REVISE. Standards review was still in progress at this point.

Material finding (P2): preservation returned a positive verdict based only on healthy After results, before checking whether Before evidence was complete and met the criterion. Uncertain, unassessable, interrupted, or failing Before evidence could therefore be mislabeled as preserved.

Disposition: corrected in 4ff0321. Before health is checked before preservation; all-met Before/After yields preserved, incomplete Before yields baseline_incomplete, and a supported Before failure can yield improved. Regression coverage includes failed, uncertain, unassessable and interrupted baselines.

Decisions required: none. Optional suggestions: none. Scope audit: no unauthorized additions; assessment-derived Summary, native completion and quiz projection are within the approved scope.

This records an independent read-only AI subagent review, not a human or GitHub approval. The reviewer ran no checks. The next round records the corrected snapshot and parent-run verification.

[kentwelcome]

Independent AI review — round 2

Reviewed implementation: 4ff0321 (pre-commit patch subsequently published unchanged), base c4c2c00.

Standards: APPROVE. Spec: APPROVE. Both independent read-only AI subagents inspected the preservation correction and escaped blocked-reason rendering. No material findings, decisions required, optional suggestions, or unauthorized scope additions remain in this round. No dependencies, private evidence publication, or live models were introduced. Source-consistency checks still require readable returned content plus an exact answer excerpt; captured calls prove only the action.

Parent-run checks passed: hooks, decisions --check, live-report-contract, report-schema, provenance/source/context checks, release-workflow, shared-data (one conditional skip), human-evaluation workflow/quiz/hosted suites, recorded-evidence and progress-assessment, pinned Ruff/shfmt, and diff whitespace checks. Browser smoke covered the shipped renderer, mobile metadata wrapping, tabs, evidence links and native blocked reasons. Reviewers ran no checks themselves.

Evidence limit: all seven saved runs rendered without a valid target assessment. The approved historical-example finding was an authored audit conclusion; this production renderer does not splice that conclusion into missing or rejected model assessments. New extraction with separately approved model cost is required to recover an assessment, and is not guaranteed to reproduce that conclusion. Older records without a saved rejection reason remain generically unavailable. Native interruption accounting also rejects old comparison rows whose membership still includes an unfinished trial, rather than silently altering model-authored counts; re-extraction is required for a valid replacement. Raw trial evidence remains available.

No merge performed. This records AI review results, not a human or GitHub approval.

[kentwelcome]

Independent AI review — round 3 / published correction

Reviewed and published head: 5ebc982, base c4c2c00. Both independent AI reviewers re-read the complete scope and the latest formatted two-file correction: Standards APPROVE; Spec APPROVE.

Post-publication findings resolved: the plan status table still described the old deferred design/unmerged preparation stage; it now matches the approved design and publication. The sanitizer assertion had been testing discarded headline content instead of retained observations; it now injects hostile script/link/style/onclick into an existing summary-why paragraph, verifies safe version emphasis survives, and verifies active content is removed.

No material findings, decisions required or optional suggestions remain. Scope audit: documentation and an existing synthetic regression test only; no runtime or dependency changes. Parent-run quiz suite passed all 20 tests, followed by pinned formatting and whitespace checks. Reviewers ran no checks. Initial malformed fixture shape was corrected before approval; the strict shape validator was not relaxed.

The prior head 4ff0321 passed GitHub Format and Unit CI. The follow-up head starts a new CI run. All previously stated live-model and historical-assessment limitations remain. No merge performed.

[kentwelcome]

Publication verification: head 5ebc982 passed both Unit and Format in CI run https://github.com/spacedock-dev/behavior-diff/actions/runs/39364095294 . This is the reviewed documentation/test-only follow-up to the implementation. No merge performed.

[kentwelcome]

Post-approval real-case verification — head 5ebc982

User approved two assessment-only model calls on private copies of saved real-case trials. Executed one call per case, no retries, no new trials; original artifacts hash-verified unchanged. Both extraction commands and shipped renders completed. Browser inspection confirmed the actual summaries. No private excerpts or artifacts are attached.

Known-issue case: target assessment rejected because output excerpts combined noncontiguous answer text rather than exact trial-local substrings. The current report correctly exposes rejection, but the extractor also missed the required code-change check; its narrative claims the verification improvement succeeded. Raw-evidence audit still supports the known gap. Some necessary source returns were absent from extraction inputs, so the false-content claim cannot be established by extraction from those inputs alone.

Regular-behavior comparator: valid assessment with eight checks: two improvements, three preserved requirements, three unassessable requirements. Overall result is inconclusive, not a clean no-gap outcome. The expected unchanged verdict is correctly preserved; excluded/truncated source material and missing publication evidence prevent broader conclusions.

Models followed each saved configured extractor: OMP with Claude Opus 4.5 for the issue case and OMP with GPT-6.1 Sol/high for the comparator. Issue-case native response confirms the model; comparator exposes only configured selector, not provider-returned identity. The two models/harnesses differ, so this is not a controlled model comparison.

Disposition: this does NOT establish the desired end-to-end issue-found/no-gap acceptance pair. Validation fail-closed behavior works, but assessment reliability and evidence coverage remain unresolved. These observations at this head do not establish that the PR introduced the extraction defects. Recommend holding merge for acceptance investigation rather than treating earlier deterministic/reviewer approvals as live effectiveness proof. No code changed and no merge performed.

[kentwelcome]

Post-approval presentation finding: user reported the verification reports differ substantially from the approved preview. Direct browser comparison confirmed design drift in headline hierarchy, Expected/Observed panels, method explanation and evidence density, separate from legitimate inconclusive assessment states. Disposition: pending production correction; hold merge. The current local refinement is explicitly an authored presentation preview using fresh verification evidence, not a production renderer fix or new model result. Earlier approvals do not establish design fidelity. No private preview or evidence is published here.

[kentwelcome]

Design checkpoint: the user accepted the revised content-preserving preview ("This one looks better"). Carry this direction into the production correction: concise finding and Expected/Observed framing above the original detailed Before/After observations, counts, primary result, explanations, cautions, attention findings and assessment checks. Preserve original evidence wording and navigation; expose full task/setup/limits through disclosures. Flag conflicting model interpretations rather than silently dropping or endorsing them. This acceptance is of the presentation preview, not evidence that the production renderer has been corrected or that the real-case assessment gaps are resolved. Production correction and verification remain pending; no merge authorized.

[kentwelcome]

Independent AI review — round 4 (unpublished production correction)

Scope: head 5ebc982 plus local presentation correction; base c4c2c00. Standards: APPROVE. Spec: REVISE (P2).

Spec finding: preservation checks with an unmet After outcome say "no longer met" even when Before already failed or was unknown. That wording implies an unsupported regression. Remedy: neutral "did not meet" wording, with both-sides-unmet and unknown-baseline coverage. Fix pending this round; no other material findings, scope expansion or new decisions identified.

Parent checks passed before this finding: report-schema, quiz21, hosted12, live-report-contract, report-ci-context/report-source, pinned Ruff140 and whitespace checks. Actual production rerenders preserve both saved cases structured data and original observation text; original artifacts remain unchanged. Blind projections, browser navigation and mobile layout checks passed. Screenshot capture timed out and is not claimed. User visual acceptance remains a separate pre-merge gate. No model calls or merge.

[kentwelcome]

Independent AI review — round 5

Published reviewed head: 8103673; base c4c2c00. Standards APPROVE; Spec APPROVE. Both reviewers inspected the final neutral preservation-meaning correction and regression coverage; no material findings, decisions required or optional suggestions remain. Scope audit: bounded presentation accuracy and synthetic coverage; no new models, evidence relaxation, authored case-specific findings or schema changes in this correction.

The regression failed before the fix and passed afterward. Parent final report-schema, pinned Ruff140 and diff whitespace checks passed. Other relevant deterministic suites and immutable actual-report rerenders passed as recorded in the updated body. Screenshot capture remained unavailable; browser interactions/mobile metrics were exercised. Reviews were read-only and did not run checks themselves.

The accepted content-preserving layout is now implemented in the real renderer, not a preview wrapper. User visual acceptance of the generated reports is still required before recommending merge. Known assessment failures are separately tracked; this PR does not claim they are resolved. No merge performed.

[kentwelcome]

Published head 8103673 passed Unit and Format CI: https://github.com/spacedock-dev/behavior-diff/actions/runs/39379254135 . Independent review is approved. Remaining gate: user visual acceptance of actual production-rendered reports; no merge performed.

[kentwelcome]

Post-approval visual finding at head 8103673: the user could not identify the main focus in the production first screen. Generic assessment-status wording and repeated validation caveats dominate the specific behavioral question. Disposition: pending presentation correction; hold merge. The user authorized continuing with a concrete observed-change focus, validated result badge and explicit unanswered requirement, with technical failure details disclosed separately. Original detailed observations must remain. This does not authorize new model calls, invented findings, or changes to assessment states. Prior CI/static approval does not replace the failed visual acceptance gate.

[kentwelcome]

Independent AI review — round 6 (unpublished goal-focus correction)

Scope: head 8103673 plus local goal-focus correction; base c4c2c00. Spec APPROVE; Standards REVISE (P2 stale navigation contract).

New production presentation names the extracted comparison topic/status, shows the recorded aim and counted observations, and lists unmet/unresolved requirements. Canonical unmet conclusions remain prominent, other canonical results remain visible, and the original detailed observation subtree is unchanged. No assessment/schema/extraction/model changes.

Standards finding: tests/presentation_contract.py still locates Instruction changes in Step2 and expects Step1 to have no links. Actual Step1 now includes aim, comparison and unresolved-check links. Migrate expected links while retaining target-existence assertions. Fix in progress; no other material findings or scope decisions. User visual acceptance remains a separate gate. Verification is not yet complete for this local correction; no merge performed.

[kentwelcome]

Independent AI review — round 7

Published reviewed head: cac7425, base c4c2c00. Standards APPROVE; Spec APPROVE. The stale navigation contract is migrated with exact link order and existing-target validation retained. Optional blank topics fall back to the recorded decision label. No material findings, decisions required or optional suggestions remain; user visual acceptance remains a separate gate.

Parent verification passed: report-schema, quiz21, workflow16, hosted12, live-report-contract, report-ci-context/report-source, pinned Ruff140 and whitespace checks. Five saved examples were rerendered without new models; both real-case report-data and original observation text are unchanged, originals hash-verified unchanged, blind projections pass. Browser checks cover diagnostic expansion, unresolved-criterion links, mobile overflow/visible steps and first-screen heading/change visibility. Screenshots remain unavailable, including on old unchanged reports.

This corrects presentation focus only. It does not fix the known assessment/evidence defects or manufacture an automatic issue-found result. No merge performed.

[kentwelcome]

Published head cac7425 passed both Unit and Format CI: https://github.com/spacedock-dev/behavior-diff/actions/runs/39382953340 . Independent reviews are approved. Awaiting user visual acceptance of the actual focused production reports; no merge performed.

[kentwelcome]

Independent AI review — round 8 (unpublished scenario/assessment correction)

Reviewed published head: cac7425399436a82d3399c4cf97b83cd2fcb5233; base: c4c2c001d491ef4599b385965d46ebaf16d446cd. Review includes the complete branch and local correction, including three new files. Local patch snapshot: 2521871c8d40892369f73041af878443320c256f7a0b80d601c447a47f452992 (tracked binary diff plus ordered new-file names/contents).

Independent AI Standards: APPROVE. Independent AI Spec: APPROVE. Both read-only reviewers report no material findings, decisions required, optional suggestions or unauthorized scope additions. This is code-review approval, not a human/GitHub approval, model-verification result or merge authorization. Reviewers ran no checks or model calls.

Latest user instruction authorizes the producer/assessment correction beyond the earlier presentation-only boundary. Added mechanisms: reviewed scenario CLI/validator; trial isolation and canonical persistence; complete immutable goal-linked check coverage; schema-18 report integration; bounded recorded PR-read return recognition; synthetic regressions. Replaced unrestricted assessor-written criteria only when frozen checks exist, and replaced unconditional missing-rationale presentation when a valid record exists. No source fetching, artifact oracle, new provider, automatic outcome or additional model call is introduced.

Parent verification passed: hooks including scenario/recorded-return contracts; decisions self-check; shared-data 31 with runner smoke; live-report/report-schema; provenance/source/CI-context; release workflow; human-evaluation workflow 16, quiz 21, hosted 12; assessment lifecycle 8; pinned Ruff143/shfmt; shell syntax/ShellCheck and whitespace. Initial new-test type misuse and incidental wording assertions were corrected/removed before the reviewed snapshot; production validators were not weakened.

Actual offline smoke: shipped validator and prompt emitter accept the recovered frozen setup; twelve previously omitted PR-read returns reach the emitted issue-case prompt; all declared check IDs are present; raw private preparation notes remain excluded. Two cases and three standard saved examples render. Original case artifacts are hash-identical. Original configurations retain original observation text. Transcribed scenario records are explicitly labeled as later transcriptions from hash-matching original pretrial notes, not new autonomous preparation; incompatible old assessments remain rejected rather than receiving authored replacements.

Browser interaction verified scenario/rationale visibility, criterion navigation and narrow layouts without page overflow. Screenshot capture remains unavailable and is not claimed. No private evidence is attached.

Remaining gate: fresh model-assessment approval, observed automatic real-case results and user acceptance. The correction has not yet demonstrated automatic detection of the expected After gap. No model call, retry, new trial or merge was performed in this correction.

[kentwelcome]

Published reviewed correction: ed9aaef (base c4c2c00), matching round 8's approved local snapshot. Both independent AI reviewers approved. Unit and Format CI passed: https://github.com/spacedock-dev/behavior-diff/actions/runs/39388384990 . The scenario/assessment handoff is now in the PR; no new model calls or trials were performed. Automatic real-case finding verification still requires fresh model-cost approval, followed by user acceptance. Hold merge; no merge performed.

[kentwelcome]

Fresh verification authorization received after ed9aaef: exactly one assessment call per saved case, using each saved pinned extractor (anthropic/claude-opus-4-5-20251101:high and openai-codex/gpt-6.1-sol:high). Existing trial evidence only; no new trials, retries, model substitution or outcome rewriting. Verification is in progress on isolated private copies. This authorization does not authorize merge; results and limits will be appended after inspection.

[kentwelcome]

Fresh bounded model verification — hold merge

Verified commit ed9aaef00435f498fc2943971e3c442dc3751108 using the newly approved two assessment calls: exactly one per saved case, no retries, no substitutions, and no new trials. Native provider outputs confirm the pinned models; both invocations requested high thinking.

Case	Model	Actual automatic result
Issue case	anthropic/claude-opus-4-5-20251101:high	Raw assessor identifies missing implementation inspection in all 3 After trials. Two inaccurate output quotations reject the entire target assessment; Summary remains unavailable.
Regular comparator	openai-codex/gpt-6.1-sol:high	Assessment accepted: 4 checks supported, 5 unsettled, no After check rated failing. This is not an all-checks-pass result.

Audit findings

• The issue-case missing-code-inspection finding agrees with the recorded calls. The assessor now receives the PR-body returns and evaluates the explicit frozen implementation-inspection check, rather than silently omitting it.
• The rejected quotations are in separate workflow/checkpoint criteria. Current validation rejects the complete target, hiding even independently supported findings.
• Separate grounding defect: some issue-case source-consistency checks are marked met using PR descriptions, despite frozen requirements for implementation content. Merely accepting every structurally valid remaining check would not establish semantic correctness.
• Comparator unknowns trace to excluded command returns, fair-share truncation of relevant source text, and publication bodies absent from the bounded assessment input. Manual inspection of full saved traces is not substituted for automatic assessment.

Both actual new reports were rendered and opened in the browser. They show the recorded scenario and data rationale and preserve six tabs. DOM inspection confirms the unavailable/partly-settled states. Original saved source runs remain hash-identical. No private run artifacts or transcripts are committed or attached, and no manual outcome edits were made.

Decision: do not merge yet. Independent code review and CI passed, but the issue-found/no-gap product gate remains unmet. The next correction must address failure isolation, required-source grounding and evidence coverage; exact quotation checks must remain strict, and unknowns must not become passes. The two-call approval is exhausted; further model calls need fresh approval.

[kentwelcome]

Independent AI review — round 9 (unpublished correction)

Reviewed published head ed9aaef00435f498fc2943971e3c442dc3751108, base c4c2c001d491ef4599b385965d46ebaf16d446cd, plus local snapshot 346062533f2814a3ab98759ecc2feb2943919109f0c52a16e57143a3b93165eb (tracked binary diff plus ordered untracked path/NUL/bytes).

• Standards: APPROVE. No material findings, decisions required, or optional suggestions. Reviewed privacy/consent, captured GET filtering, unchanged evidence budgets, required-source provenance, per-check failure isolation, persistence/load parity, attention gates and regression coverage.
• Spec: APPROVE. No material findings or decisions required. Reviewed complete frozen definitions/source groups, partial/all-rejected states, unchanged After misses taking precedence, scenario rationale, HTML/Markdown/schema-19 consistency and no invented outcomes. One optional suggestion: align a stale assessor-prompt sentence with the deterministic unmet-finding headline precedence.

Both independent AI reviewers inspected the complete branch and local correction read-only. Neither ran tests, models, formatters, or accessed private verification data. Their scope audits found no unauthorized permanent mechanisms. These are source-review verdicts, not human GitHub approvals or merge authorization.

The optional prompt wording correction is now applied locally and will receive final review before publication. No model call is authorized or needed for that deterministic review.

Parent verification

Passed hooks (including scenario 10 and recorded-evidence 23 contracts), decisions self-check, live-report/schema/target contracts, shared-data 31 with runner smoke, assessment lifecycle 8, provenance/source/CI-context suites, release workflow, human-evaluation workflow 16 / quiz 21 / hosted 12, pinned shfmt/Ruff, changed-shell syntax/ShellCheck, and whitespace checks. Removed an incidental old schema-version literal in the CI-header test rather than repinning it; actual roundtrip/context checks remain.

Production replay of untouched saved responses against their original model-input evidence now retains the issue-case missing-code-inspection finding and leaves five invalid checks Not assessed. The comparator remains four supported/five unknown. No new model calls and no manual outcomes. Separate collector smoke recovers relevant previously omitted/truncated sources without increasing budgets; those new excerpts were not used to promote old model outcomes.

Browser checks exercise accepted/rejected criterion navigation, six tabs, 390px layout without page overflow, and no JavaScript errors. Original saved source runs remain unchanged. Fresh revised-prompt adherence is untested; current comparator coverage is not an all-pass result. No merge performed or authorized.

[kentwelcome]

Independent AI review — round 10 / published correction

Published reviewed head: 8fd14adc716c66e4e1b069e4fc9c0b1b268a165c; base: c4c2c001d491ef4599b385965d46ebaf16d446cd. Exact precommit snapshot 64ef22de131007ec34dde77faa29ba492359d359b2abf13aef1ba958a08cebeb was checked unchanged before staging and DCO-signed publication. Remote head/base were unchanged from the review preconditions.

Independent AI Standards: APPROVE. Independent AI Spec: APPROVE. Both reviewed the complete latest branch plus correction. No material findings, decisions required, optional suggestions or unauthorized mechanisms remain. Round 9’s optional stale prompt sentence is corrected: an independently validated unmet After requirement takes heading precedence. Reviewers ran no checks/models and accessed no private artifacts. This is source-review approval, not human approval or merge authorization.

Added/replaced mechanisms are within the user-approved correction: frozen source groups, captured required-read priority under existing budgets, conservative recorded GET support, per-check rejection with whole-target identity/completeness enforcement, schema 19 and matching consumers. Exact quotations, source provenance, privacy, attention gates and unknown states remain strict. No extra assessment pass or source fetching.

All full deterministic gates passed as recorded in round 9. After the final prompt-only edit, decisions self-check, pinned formatting and whitespace checks passed again. The actual five-step reading sequence, scenario/rationale, accepted/rejected links, six tabs, mobile widths and empty JS-error logs were exercised. Screenshot capture remains unavailable; no screenshot proof claimed.

Production replay uses untouched saved responses and their original assessment input evidence. The issue case now shows missing code inspection in 3/3 After trials, while five checks stay Not assessed for invalid citations/quotations. The comparator stays four supported/five unknown. A separate collector smoke improves available relevant sources without retroactively promoting old outcomes. No new model calls or trials were made for this correction; original saved sources remain unchanged.

The PR description now corrects the earlier shorthand about “two inaccurate quotations”: missing answer citations are also present. Fresh adherence to the new required-source prompt remains untested, and the comparator is not a complete no-gap result. User acceptance of the actual report remains required. Unit and Format CI passed for this published head: https://github.com/spacedock-dev/behavior-diff/actions/runs/39396917868 . No paid dogfood run requested. No merge performed.