Blog

4 min readfeature

The doc-truth round: the maintenance pass that made every documented number true

The doc-truth round re-derived and corrected every drifted count, measured the Coda PWA null with a receipt, and priced the instruction-surface PR lane.

  • #benchmark
  • #verification
  • #hardening

A maintenance round does not ship new capabilities. It makes the old claims true. The doc-truth round is exactly that: eight doc-drift flags from the A1 ground truth's §7.4, the Coda round's honest "not measured" PWA null, the landing page's first user-product band, and the instruction-surface PR lane priced on its own three live guard events. No new engine code — the round's output is receipts, corrected numbers, one measured figure, and one documented lane.

What got corrected

Eight doc-drift flags, all retired in one sweep (T1). Every documented count in docs/usage.md, docs/skills.md, docs/subagents.md and the dashboard screens re-derived from its live source on the day of the fix — not just the flags observed the day before, but re-grepped against the registry, since a flag's number can drift further between observation and correction. The sweep found more than the eight: the dashboard screens count (12 vs 15 in SCREEN_SPECS), the built-in agent table (12 rows with a duplicate browser row vs 20 builtins), the project skill count (7 vs 13 on disk), the project agent count (4 vs 6), the session tools count (seventeen vs twenty), and the split-view commands count (eleven vs twelve). Every corrected figure carries a source:line derivation in the receipt; the pin decision is receipt-only per surface (the repo's own registered-content surfaces legitimately grow, so a count assertion would false-positive on every legitimate add). No pin tests landed — the four surfaces all use the repo's receipt-only derivation pattern for the reason it applies: a registered-content count is not a drift-pin candidate, it is a drift-record.

What got measured

The Coda PWA null, flipped with a real receipted measurement (T2). The PT-5 plan's own contract said "benchmark: when real, receipt a PWA installable/offline-load measurement under benchmarks/; otherwise the figure stays 'not measured'." This round built the measurement lane (benchmarks/coda-pwa/, a zero-model-traffic Playwright runner that boots the app's own dev server), recorded installability as data (manifest 200, all 4 required fields, 3/3 icons, SW controller active via waitForFunction), and measured offline boot-from-cache timing across N=5 runs: 91 ms median offline vs 175 ms median online — the 84 ms delta (~48% faster) is the precache's measured effect. The figure is receipted under benchmarks/coda-pwa/receipts/run-2026-11-03T06-14-57-955Z/, re-runnable on demand via bun benchmarks/coda-pwa/run-pwa-bench.mjs [--runs N]. docs/coda.md and benchmarks/RESULTS.md now link the lane instead of stopping at the e2e proof; the six-track landing feed contract stands — no TRACKS entry, no landing figure, check:benchmarks green by not needing a change. The PT-5 plan's historical "not measured" phrasing stays as the pre-measure record; the round log's T5 closeout retires it against the receipt.

What got shipped

The landing page's first user-product band (T3). Coda is apps/coda/, a self-contained static PWA, not a landing route — so the live-demo gallery was the DISHONEST home (every gallery app iframes a landing route; Coda can't be routed). The band is a new CodaSection (src/components/landing/coda.tsx, aqua, #coda, after LiveDemoSection) with its own authored data module, drift-pinned by __tests__/coda.test.tsx. The media are the two stills from the committed in-repo capture lane (apps/coda/scripts/capture-usage.mjs → public/landing/coda/) — the same interactions the e2e spec drives (New note → Save; Add task via the pre-armed dialog), shot through the real UI at 1280×800 @2×, with a PROVENANCE.md sidecar so the capture is re-runnable and auditable. The thirteenth scroll-budget amendment (15.5 → 16.5vh, 3,950 → 4,060 words) paid for the band's footprint; compressed first (16.4 → 16.3vh, only visible copy — the FAQ-collapsed / dex-hidden lessons applied), the canonical e2e twin measured 16.3vh / 4,033w, both gate files moved together, the OG shots regenerated (canvas assert held).

What got priced

The instruction-surface PR lane, on its own three live guard events (T4). The gate's own deny reason says "propose changes via a PR for human review" — the PR path is the documented remediation today, even though the mechanism it points to is not a first-class tool. The round's three guard-class events (T2's benchmarks/AGENTS.md ownership-list edit refused and operator-landed; T3's landing AGENTS.md amendment-ledger entry refused and operator-landed; the Coda round's apps/coda/AGENTS.md child DOX operator-landed) are the evidence set. The decision doc (docs/instruction-pr-lane-research.md) prices the candidates: NO-GO on a new standing git push / gh pr create session primitive (the new trust surface is not worth the small round-trip cost the convention already carries); GO, zero-cost and now, on documenting the propose-via-PR path through tools that already exist operator-side — the operator (or a future sanctioned lane) lands the draft from the receipt via git branch + git push + gh pr create in their own shell. The gate stays FINAL in-session, byte-for-byte; the lane is external to the session, and the operator's merge click is the final act in both the convention and the documented lane.

The round's receipts: .unown/receipts/t1-doc-drift.md, t2-pwa-benchmark.md, t3-landing-capture.md, t4-pr-lane.md, and the closeout's t5-closeout.md — one per ticket, mirrored operator-side at the gate.

Coda on a fresh boot — the empty board, ready Fresh boot: the empty board, search, and the install affordance — captured through the real UI by the committed lane.

One note and one task in Coda, after the e2e interactions One note and one task, created through the real UI — the same interactions the e2e spec drives.

Spotted something wrong or worth adding? The record is the repo — open an issue.

All posts