Measure declared region bounds, keep raw and effective verdicts distinct, and use active blockers for repair feedback. Revalidate plate receipts against current asset, region, and comp hashes; require an explicit quoted comp-authority downgrade for force overrides. Add regressions for missing assets, copied comp pixels, repeated failures, generic delegation, and stale reports. Preserve fidelity thresholds. AI assistance: implemented and validated with OpenAI Codex.
3.3 KiB
Comp gate evidence and integrity
Comp fidelity still has two independent requirements: the numeric score and the absence of blocking findings. A high overall score, a repeated attempt, or an existing plate file does not satisfy a missing-region check.
Measurements and decisions
comp-diff measures the declared region bounds. It does not enlarge a narrow
region to include neighbouring elements. Coverage and region-kind checks remain
the responsibility of comp-spec; this change does not authorize omitting
regions or shrinking their bounds to exclude required work.
A standalone comp-diff report contains raw verdicts. The hero gate can interpret
those measurements using current plate validation and rendered presence. For
hero evidence:
raw-report.jsonretains uninterpreted metrics and verdicts.report.jsonretains the same scores, addsrawVerdictto each region, and publishes the effectiveverdictused by the gate and paired image label.gate.okandgate.reasonsdescribe whether advancement is allowed and all unresolved blockers.driftalone does not mean a region is nonblocking.- Paired images label their gate status as
BLOCKING,NONBLOCKING, orCHECK GATE, using the same report fields. - Region-specific blockers are recorded as
blockingReasons.blocking: nullmeans an unscoped blocking finding prevents attributing a clean bill of health to that region. Unscoped findings remain ingate.unscopedReasons; they are not hidden or waived.
When capture, spec, or plate validation fails before measurement, the current
report has measurementsAvailable: false, no region results, and the failed
gate reasons. It does not present the preceding capture as current evidence.
Stall feedback follows repeated blocking reasons. It never chooses an asset to regenerate solely because that asset has the lowest raw score. The feedback is additional context; it does not clear a finding or advance the phase.
Plate validation
Successful plate receipts include SHA-256 fingerprints of the asset bytes,
measured region, and comp. A changed or deleted file, a changed region, or a
changed comp invalidates the receipt. Legacy score-only receipts are revalidated.
An invalid plate cannot receive an ok receipt merely because its PNG decoded.
Rendered presence is still checked after asset validation, so a file hidden in
the page does not count as placed.
Overrides
A --force --reason must contain a direct quoted downgrade of comp authority
attributed to the user. Generic delegation, the builder's surrounding claim that
it may proceed, or a gate exception does not establish that authorization.
The quote parser is deliberately conservative. It cannot authenticate a quote:
the calling harness must retain the actual user answer and assess provenance.
Neither local receipts nor caller-written state are a security boundary against
an agent with arbitrary write access to all files and engine code.
Validation
The comp-verbs regressions cover narrow-region isolation, changed/deleted plate receipts, comp-crop reuse, hidden rendered assets, repeated failed attempts, ambiguous overrides, spec coverage and kind refusals, and raw/effective report agreement. They do not certify semantic equivalence of arbitrary artwork or justify relaxing a fidelity threshold.