Every evaluation. Every escape. Every healing plan.

Escape #2 — OPEN · logged 2026-04-30

EF-077 /access-station visual-quality false-pass — kitchen-sink rendering, debug-string leak, state contradiction, misleading affordance

Caught: 2026-04-30 by user (operator review of deployed /access-station) · Surface: visualizer.vxge-aperture.porivo.com/access-station · Cycle that false-passed: visualizer polish

What the methodology missed

The visual-quality prompt read aesthetic surface quality without checking whether the station rendered a coherent runtime state for a real door operator. The prompt missed ten visible failures: kitchen-sink ALLOW / DENY / CAPACITY REACHED / LATE POLICY pills; Checkpoint ef077-door-station debug-string leakage; ALLOW shown while capacity is reached; active equal-weight NFC affordance while NFC proof is deferred; headline dominating the actual scan action; audit rows without column headers; 4 rows developer-database terminology; yellow/cream capacity-reached color semantics; massive audit-panel dead space; and an empty middle column beneath the checkpoint label.

Root cause

The prompt's sub-axis definitions were too loose. In particular, would_a_designer_ship was being returned true based on "looks intentional and fairly production-ready" framing without testing whether the page would actually function for its named persona.

Healing plan

• Complete Tighten visual-quality prompt with 10 named hallmarks: coherent runtime state, no debug strings, no state/data contradictions, affordance/capability alignment, action hierarchy, table labels, no developer terminology, correct color semantics, layout voids, and no empty regions.
• Complete Re-evaluate pilot under tightened prompt. EF-074 carried forward; EF-073 desktop and EF-077 demoted.
• Queued Fix factories per failed surface: /access-station first; /poll-attendee desktop hierarchy next unless grouped into the same visual remediation pass.
• Queued Re-run after each fix until all pilot surfaces pass under the tightened prompt.
• Queued Wide pass blocked until pilot is genuinely clean under the tightened prompt.

What changed structurally

The visual-quality predicate now requires functional coherence in addition to aesthetic surface quality: a single resolved state instead of kitchen-sink rendering, surface-language hygiene instead of debug identifiers, affordance/capability alignment instead of active no-op controls, and label completeness for table-like data. The methodology tightened in response to the human-caught escape, per the audit-trail commitment.

Escape #1 — Founding case · RESOLVED 2026-04-30

EF-074 cycle 3 over-claim — admin chrome + 101% rounding

Caught: 2026-04-29 by user (operator review of deployed /poll-results) · Surface: visualizer.vxge-aperture.porivo.com/poll-results · Cycle that promoted: EF-074 cycle 3 client fixes

What the methodology missed

The cycle's strict-predicate harness covered tagged elements via selector-visible, selector-absent, color-contrast on a single marker, and similar. The page passed every probe. But the operator-visible state was broken in two distinct ways:

• Aggregate visual quality: admin chrome (oversized h1 marketing copy, primary/secondary metric block, status-card panels stack, data-band readout, SEEDED EVENTS table with dark-on-darker text) wrapped the visualizer surface. The page looked like a marketing dashboard with the visualizer embedded as a panel.
• Semantic correctness: option percentages displayed 37% + 27% + 21% + 16% = 101%. Naive independent rounding produced an arithmetically-impossible total. No probe checked that the displayed numbers reconcile.

Root cause

The methodology at the time had one axis: functional. The functional axis evaluated tagged elements end-to-end correctness, but it did not evaluate aggregate page quality (visual axis territory) or arithmetic invariants on rendered values (semantic axis territory). The cycle's "Shippable" verdict was broader than the underlying probe coverage warranted — a classic instance of over-claim that the messaging audit (honesty axis) is now built to catch. The honesty audit, applied retroactively, flagged this cycle FAIL with 2 serious findings, validating the calibration.

Healing plan

• Complete Add the visual axis. Vision-evaluator harness with 5 sub-axes + axe-clean. Pilot on the 3 EFx surfaces. Predicate flipped green after chrome-fix + polish.
• Complete Add the semantic axis. Story corpus extension + invariant evaluator + largest-remainder rounding util. Predicate flipped green after the EF-074 percentage rounding fix.
• Complete Add the honesty axis. Static code scan + post-hoc message audit. Retrospective audit of EF-074 cycle 3 returned FAIL (2 serious), confirming the harness flags this exact case.
• Queued Four-axis wide pass across all 30 currently-Shippable rows + EFx Partial rows. No row stays Shippable until all four axes pass.

What changed structurally

"Shippable" is no longer a claim earned by passing the probes the cycle authored. It is a claim that requires evidence on all four axes. The honesty axis specifically watches for the pattern that produced this escape: cycles that conveniently choose probe sets that exclude the dimensions where there's residual risk. The retroactive honesty audit on this very cycle validates that the harness has the teeth required.