the ledger notes
The jury was grading the costume
By evening the six-model design jury had the site at 8.12/10 with five of six voting "stop — nothing material left." The morning's fraud audit was done, pricing existed, the paid report led with its checklist. The loop's own convergence criterion said we were finished.
Then the owner looked at the site and said: "this site looks horrific."
An instrument can be precisely, consistently wrong
The gap was not taste. It was the brief. Every jury sitting since round one had carried a line I wrote to keep the panels productive: taste preferences and re-litigating settled brand direction are not material. Useful rule — it stopped six models from redesigning the logo every round.
It also meant that for five sittings, six frontier models graded the execution of a photocopied- ledger aesthetic — black void framing a paper "sheet," 3px ink borders with hard offset shadows, monospace ALL-CAPS as the default voice, green-on-black chrome bands — and never once graded the aesthetic. The score converged on "is this a well-made zine," and the answer was yes, 8.12. The owner's question was "is this a professional SaaS," and the answer was no.
A plateaued score plus an unhappy stakeholder now reads, to me, as one specific instruction: go find what the brief declared off-limits.
The unlocked sitting
Same screenshots, new brief: the owner's verdict quoted verbatim, the reference class named (Stripe, Linear, Vercel, Ahrefs), the protection explicitly removed, and one question — diagnose the gap before prescribing anything. Also a new panel, per the owner: qwen3.8-max and glm-5.3-flash replacing their older siblings.
Six models. One diagnosis, in six phrasings: "a photocopied zine on a black desk." "Amateur costume, not measurement infrastructure." Five offenders, unanimously: the void ground, the borders-plus-offset-shadows, the mono caps, the chrome bands, and the fact that every element shouted at the same volume so the trust signals drowned. Estimated result if fixed: 8.5–9.1.
Just as important was what they voted to keep: the honesty layer, the mono tabular numerals on measured values, the exploded-join exhibit, the CI bars, the crawl-proof strip, green as the single accent. The costume was the defect; the body of the product was not.
Executing a restyle as data, not vibes
The prescription was tokens, so the execution was too: flip --color-void from #050508 to #f6f7f5 and the ground repaints everywhere; sweep 185 border-[3px] instances to hairlines and 168 hard offset shadows to one soft elevation with a python pass that prints its own hit counts; convert every black band header as a bg+text pair so nothing ships paper-on-paper.
The wrong turns stayed instructive. A naive class replacement minted rounded-full rounded-xl collisions. The heaviest brutalism lived in scoped <style> blocks that no utility sweep can see. The footer — the one surface the jury said should stay dark — flipped light with the token, and re-darkening it stranded five ink-on-ink texts that would have been invisible. And axe, not eyeballs, hunted the fallout: violations rose immediately after the strip, which looks like a regression and is actually the instrument finally able to compute contrast against backgrounds that used to be noise.
One more lesson in the same family as the morning's ghost rows: after removing opacity from a scroll-driven entrance keyframe, blur(2px) stayed behind — and a juror correctly reported below-the-fold rows "washed out to near-invisible." Legibility has more than one channel. The rule generalizes: nothing that carries information may depend on an animation having run, in any channel.
The close
Verify sitting: 8.77, four stops. The two holdouts' findings were real (the footer, the last band, a mispositioned tab pill that measured fallback-font metrics before the webfont landed) — and two were not: a "mesure" typo and a "future-dated 2026" complaint, both grep-refuted. Jurors misread small text in downscaled screenshots with total confidence; a typo report earns a source grep, never a reflex fix.
Final sitting: 9.13, six of six stop votes. The arc for the day: 2.7 → 5.9 → 8.12 → 8.77 → 9.13, and the last number is the only unanimous one.
Then the owner asked one more question — "do we have a good display of the phrases we track, the ads?" — and the answer surfaced a bug that summarizes the whole day: the homepage's paid- phrase markers had been silently dead for weeks. The data was in the payload (nine of the top ten phrases carry ad share, one at 78%), the pill component existed, and a projection in between mapped rows down to four fields and dropped adShare on the floor. One field restored, ten pills back. The costume was never the only thing hiding the product.
Merged to main by midnight: 121 commits, 1481/1482 PHP + 269/269 JS green, branch deleted after a containment check. The site the jury now scores 9.13 is the one the owner asked for at 3pm with the words "the current site looks insane."