Detect a missing CLI before worker startup, keep Live usable through the foreground poller, and surface actionable status in Live and Live Lab.\n\nAI-assisted implementation.
Three new rules and three widenings, all from confirmed eval-corpus
escapes found by eye:
script-error (quality, error severity, URL engine): pageerror listener
attached before goto catches uncaught exceptions AND parse errors (a
syntax error fires during the initial parse, long before load). Deduped
by message, capped at 3. A JS typo was silently deleting whole pages.
content-hidden-at-rest (quality, error, URL engine): after the main
at-rest scan, an instant-scroll reveal sweep (bypasses scroll-behavior:
smooth, which silently defeated the first sweep design) gives every
IntersectionObserver reveal its chance to fire, returns to top, then
measures the share of text characters still at opacity 0 / visibility
hidden. display:none / [hidden] / aria-hidden subtrees stay out of the
denominator. Fires above 30% with a 200/150-char floor. Calibration on
30 corpus samples: broken repro holds 83% after the sweep, all clean
samples (including 0.75-0.93 at-rest reveal pages) drop to <= 7%.
edge-flush-cards (quality, warning, browser): cards with their own
opaque background or 2+ borders inside a horizontal scroller, flush
against one edge of the clip box at rest (< 8px, > -24px so deliberate
mid-card peeks stay exempt) while keeping a gutter on the other side.
Grouped per scroller. Repro: transit-mobile pager whose first snap
panel is 407px wide inside a 390px clip. New --viewport WxH CLI flag
makes mobile-width URL scans reachable (--viewport 390x844).
Chip/badge contrast widening: the SAFE_TAGS styled-button exception in
checkColors now covers any text-bearing element painting its own opaque
background at >= 9px font, not just a/button. The shipped miss: a span
SEV-2 chip whose white text lost a specificity fight and rendered
muted-on-red at 1.2:1. Static adapter also resolves var() own-bg via
the custom-property map so the gate engages on token backgrounds.
background:none cascade fix: the background shorthand now resets
background-color/-image when it names neither (and no var()). Exposed
by the chip widening: pre code { background: none } left an earlier
surface color standing and manufactured 1.1:1 phantom findings.
text-overflow inline-owner widening: inline elements have no client
geometry (clientWidth 0) so the scrollWidth path never saw them, and
their block parent owns no direct text. New branch measures the inline
rect against the nearest block container's padding box (16px floor,
transform-path exempt). Repro: nowrap span.v spilling 45px past its
grid cell.
The round-3 nav-CTA contrast escape (val-a22-opus obs 003 header CTA)
was verified already covered at HEAD by the earlier parseAnyColor
oklch fallback; both engines fire 3.6:1 on the repro, no change needed.
FP sweep across 36 val-a21/a22/a23 samples: new rules fire only on
their repros (script-error also catches a second genuinely broken
sample); static-engine delta is limited to the chip repro plus two
borderline-but-real chip findings on one sample.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Regular-use guard ahead of the realistic-lane validation: forced
creativity must never supersede user input or product context.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Opus probe (r10-opus-wireframe): control articulates loose skeletons,
wireframe arm names them (dubbing script sheet, timecode gutter spine)
and diffs against the standard stack every time. Targets Paul's
layout-diversity question: concept-atom commitment with template
skeletons underneath.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Paul: 'way better to have the first iteration land fully committed to
the concept, because that's the genuinely hard part. the next pass can
make sure it is clear and effective.' The check selected against the
original lektor site itself, the campaign's 10/10 reference.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
heading-rhythm (quality): a heading binds to the content it introduces,
so its rendered space above must exceed its space below. Browser-only:
measures real getBoundingClientRect gaps (margin collapsing, flex rows,
and section padding make authored margins untrustworthy), merges eyebrow
labels into the heading cluster, requires same-column edges, and exempts
first-in-container headings, bounded bands, and small cards. Fires only
when 2+ headings on a page invert the rhythm.
blinking-cursor (slop, advisory): a decorative blinking caret (solid
block, underscore bar, or block glyph) bound to an infinite blink
animation in the landing region of a page. Real editable surfaces
(contenteditable, role=textbox, inputs) are exempt; round pulsing dots
stay with the pulsing-dot rule.
Verified against eval corpus repros: heading-rhythm fires on the
val-a18 observability sample Paul flagged (6 headings, 0px above vs
40px below) and blinking-cursor on the val-a19 hero terminal cursor;
10 other samples across both runs stay clean.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The inverse-probe extracted this from Paul's NewRelic review and it won
in the batch4 forward test; it was never ported. Craft diagnosis: pages
lose on rhythm monotony (one treatment uniformly applied), not defects.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The skill's decide-then-build step opens the built HTML artifact with a
DIRECTION CONTRACT comment (UNIQUE / NOT-TEMPLATE / OWN-WORLD / STORY /
FIRST VIEWPORT / FORM). Until now nothing ever judged the finished build
against that contract; the eval harness proved sample contracts promised
radical compositions while the build shipped the standard template anyway.
The Stop deep pass now extracts the leading contract comment from each
session-touched HTML file (marker match in the first 200 chars, body
capped at 1800 chars) and appends a contract-audit section after the
detector findings: audit the render promise by promise, naming the two
observed failure shapes (a promise not in the pixels; a contract whose
own plan is the standard template wearing the concept's nouns). Zero
extra API calls; the audit rides the existing single Stop emission and
fires at most once per file per session via a contractAudited flag on
the same session cache entry the finding dedupe uses.
Ported from the eval harness reference implementation
(extractDirectionContract / composeContractAuditMessage in
impeccable-evals runner/workers/anthropic-native.ts). No hooks.json
changes needed: Claude Code and Codex both already dispatch Stop to
hook.mjs.
Tests: 163 -> 179 in tests/hook.test.mjs (extraction unit coverage plus
Stop-pass integration: present/absent/once-per-session/non-HTML/
malformed/oversized). hook-build 18/18, build:skills prose gate clean.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Paul: no stochastic challenger-assignment mode (unreproducible bad draws
= undebuggable bug reports); keep the weigh-off. Every roll now prints
its key so any field report can be replayed with --from.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Contract-probe campaign findings (evals repo, notes/fable-oneshot-craft-plan.md):
a single model's resonance ranking is deterministic (30/35 identical
concepts across 16 framings); dice must come from the script, mirroring
the palette-seed result. Derived candidates stay grounded in the
audience's world + subject's cultural home; challengers win only on
identification x clarity; incumbent-with-deliberate-idea overrides the
roll. Validated at contract level on 01-observability + r10-lektor
(teletext ranks #3 for lektor; assigned index 3 produced it).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Capture below-fold selected targets and each progressive variant reliably for portable evidence bundles. AI-assisted implementation under maintainer direction.
Four gaps found shipping in Opus 4.8 eval samples during human review:
1. low-contrast (extended): the browser adapters parsed text/own-bg
colors with parseRgb only, so Chrome's oklch()-serialized computed
colors silently skipped every contrast check — a flat dark-on-dark
nav CTA (broader nav selector beating the button class) shipped at
1.5:1 undetected. checkElementColorsDOM and readOwnBackgroundColor
now fall back to parseAnyColor. Near-threshold ratios print two
decimals so a 4.497 finding no longer reads "4.5 needs 4.5".
2. NEW numbered-section-labels (slop, advisory): tiny (<=13px) styled
numeric index labels riding beside section headings, repeated across
2+ sections with distinct indices. Sibling of repeated-section-kickers
(which deliberately excludes bare numeric labels); handles both the
direct prev-sibling shape and label-before-heading-wrapper shape.
List/nav/table/card-item numbering is exempt.
3. side-tab (extended): the vertical pseudo-element stripe scan required
the stripe to touch both corners (top/bottom 0 or height 100%), so a
left accent bar inset a few px from each end evaded it; small end
insets (<=20px each) now count. Added a browser-side pseudo-element
check (getComputedStyle(el, '::before'/'::after')) since runtime-
assigned custom-property colors are invisible to the text scanner.
Selection-state exemptions stay as narrowed: only aria-selected=true /
aria-current / active-class markers exempt, plus button/link
affordances on the horizontal variant.
4. NEW repeated-container-text (quality): the same literal string (>=4
chars, contains letters) rendered 3+ times at 3+ structurally distinct
positions inside one bordered/elevated container. Parallel/templated
repetition (table cells, calendar grids, nav lists, identical sibling
rows) never counts — structural signatures, not word lists.
Verified: each rule fires on its repro sample via the file:// browser
scan; clean eval samples add no new findings (the new low-contrast hits
on other samples are genuine sub-AA oklch button pairs). Full test
suite green; browser bundle regenerated; README/homepage rule counts
bumped 49 -> 51 (docs-integrity test enforces them).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Four lines from the r10 dual consultation (codex gpt-5.6-sol + gemini
3.5-pro on the actual HTMLs) and the hero-probe micro-eval: the probe
isolated a first-viewport monoculture (same split template in every
sample, control and skill alike) and showed these lines break it while
codex's raw 15-liner alone does not. The incumbent sentence swap fixes
the r10 root cause both consultants independently identified.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Add repeated production annotation timing, real app-server cache telemetry, the context-delta decision, and correct --judge=false handling.\n\nAI-assisted: OpenAI Codex.
Keep short crash-recovery leases without allowing a healthy worker to queue its own generation twice, and surface non-monotonic benchmark journals as errors.\n\nAI-assisted: OpenAI Codex.
Replaces the a14/a15 attempts (both deleted). Diagnosis: incentive
stacking; the placeholder-completion MUST plus the image tool turned
'bolder' into full-bleed photo insertion. Scope preservation is the
missing rule, not imagery policy.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Return after a durable starting record, overlap app-server initialization with page startup, dynamically reclaim generation after worker failure, and cap hard-crash leases at 15 seconds.\n\nAI-assisted: OpenAI Codex.
x02 a14 rerun: 3/3 samples still imported photos — the unscoped MUST in
the Persuade mode block overrode the existing-worlds principle. Scoping
keeps the greenfield ablation win, frees iteration asks.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Remove stale and contradictory experiments, order the surviving decisions by impact, and replace synthetic claims with current production evidence.\n\nAI-assisted: OpenAI Codex.
Replace stale startup and synthetic claims with production browser timing, matched architecture comparisons, Accept latency, and honest run counts.\n\nAI-assisted: OpenAI Codex.
x02-tidewater-bolder eval: 3/3 skill-on samples imported photography into
a photo-free seed system (0% arena vs competitor, which amplified the
seed's own vocabulary instead). One sentence, shape-level, no examples.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Count variants only inside the active generation wrapper so deferred carbonize markers cannot create false fast-path results.\n\nAI-assisted: OpenAI Codex.
Paul BakausandClaude Fable 5