Owns: root *.md, AGENTS.md, agents/ Must not touch: computedriven.com/ (lane E), RRABBIT/ (lane A), studbook/ (lane B). Status: ✅ READY NOW. Unblocked, and it clears debt that is actively misleading other sessions.
On 2026-08-15 CLAUDE.md — the file injected into every session in this tree — was found still teaching the shared-Supabase data layer, with the full migration table and runnable supabase start commands. It had been abandoned sixteen days earlier. Every session in that window began from a false premise.
That is the failure this lane exists to prevent. A ruling that is not swept into the documents is a ruling the next session does not have.
REVISION_REGISTER.md — the ranked list, with what is fixed and what is not.
STACK_COMPLETION.md §0/§6 — [TRAVIS], not yoursThe ratings take Fly deployment breadth as an input, and the wind-down ruling changed it. They are stale, not wrong. The document already reserves the blended-vs-per-layer question for Travis. Flag it; do not recompute it.
AGENTS.md is stale and nobody has audited itFound in passing, never swept: it describes WebHost.Systems as "Convex + Vite/React/Clerk", while CLAUDE.md says Supabase backend + Vite/React/Supabase Auth. One of them is wrong and possibly both. It also predates every 2026-08-15 ruling. 95 lines — read all of it against the tree.
docs/spec/ folders were never swept~14 of them. WebHost.Systems/docs/spec/30_DATA_MODEL_SUPABASE.md is Supabase by name, and its v1.0 spec is otherwise the most complete hub design in the portfolio. Others certainly assume Supabase or a Fly deployment.
Do not mass-edit. Some of these are dated design records where a supersession banner is the honest fix; others are live guidance that should be corrected. Tell the two apart, and say which you did.
ampersand-supabase/ is marked archivedIt is archived by decision. Nothing has checked that the directory says so to someone who opens it.
Never checked. Curl the served pages rather than grepping the tree — the local-grep error mode is recorded and has produced a wrong claim here before.
CONFIDENTIALITY.md §1 lists seven adversaries (T1–T7) and none are declared in scope. Every mechanism choice in class B is downstream of that answer. This is a [TRAVIS] ruling, but the document that makes it easy to make — the concrete question, the consequence of each answer — is yours to write.
Retract in place, by name. Never silently edit a claim away. This tree's retractions are its
most valuable records — the ledger that retracted two of its own findings, the plan that retracted its own pricing twice.
Do not rewrite dated analyses. A gap review records what was seen at the time; rewriting it
fabricates a review nobody performed. STACK_ARCHITECTURE_GAP_REVIEW.md, ECOSYSTEM.md and STACK_PLANNING.md got supersession banners for exactly this reason. Say which half survives.
Never hand-type a derived number. The law count drifted twice — 116 in four places in
STACK_SITEMAP.md and two more in STACK_COMPLETION.md, where the suites derive 118. The suites count themselves: node test/laws.mjs && node test/compose-laws.mjs from AmpersandBoxDesign/box-and-box/. Quote the output.
A boundary in a document is a decision, not a measurement. Plan §8.3's out-of-scope list was
inherited as fact and produced a conclusion off by four orders of magnitude. If a scope looks load-bearing, ask rather than assume.
Date every claim, and give it the rung it has earned: spec → in_tree → live_local →
live_deployed → external. When in doubt, downgrade.
REVISION_REGISTER.md has no open row that is not explicitly [TRAVIS], and each closed row says what was done and how it was checked. Keep the register — it is a standing document, not a one-off. The failure it catches is silent and cumulative.