What slipped through — a documentation sweep after the 2026-08-30 wave

Roughly forty-five lanes landed on 2026-08-30 and the .md corpus grew faster than it was reconciled. This is a read-mostly sweep of that corpus for three things: claims today's landings falsified, commitments nobody executed, and decision records that do not carry the rulings made on them.

It is grouped by disposition, not by topic: what was amended in place, what is open and needs an owner, and what is stale and was deliberately left alone. Every amendment cites its superseding evidence. Every size is a number somebody measured or estimated, attributed to whoever said it — never a feeling. Where an item could not be settled from the tree, it says so rather than guessing.

Two commits carry the amendments: a4ceb704 (decision records) and 79522879 (falsified claims).

---

1. Amended in place — 18 edits across 17 files

1.1 The decision records now carry their own rulings

Ember ruled all seven packet questions on 2026-08-30. Every ruling lived only in WAVE.md; not one had been carried into the record it rules on.

Three consequences worth reading off that table:

• D2 is ruled and enforced nowhere in the protocol. The only live check on MAX_FEE_BPS = 500 is a shell guard at tools/release/stage-devnet-sponsored-market-open.sh:84-85. A founding that does not go through that one script is unbounded. §6 of 0014 priced the real fix at one lane, protocol-tier: one const, one refusal discriminant, the Lean bound and its two native_decide boundary theorems, a census regeneration.
• E3 is the one packet question still genuinely open — seal-rent beneficiary, leaning collector-keeps, final call deferred. Recorded on OMISSION_INDEX P-006, which is where a CloseSeal implementer would look.
• E5 was accepted conditionally, and the condition is a charter requirement: the debtor must always be able to settle unilaterally, including when the fee recipient's token account has vanished.

1.2 Claims today's landings falsified

---

2. Open — needs an owner

Cross-checked against docs/ledger/SESSION_STATE_2026-08-31.md's lane queue and today's commits. Sizes are the source document's own.

2.1 Owed by a ruling that has now been made

2.2 Named today, carried nowhere

Each of these was named by a lane in its yield and appears in no queue outside docs/ledger/SESSION_STATE_2026-08-31.md, which is rewritten at each compact.

2.3 Structural: the reference index does not know today happened

docs/reference/decisions.md is generated and stops at decision 0013. All four of today's decision records — the fee rate, the dead markets, checked release identity, cache-read authentication — are absent from the index that ships to the public site. docs/reference/README.md's totals (13 programs, 141 routes, 212 refusal codes) are likewise pre-BASIS-ENUM.

Not regenerated here: whole-tree generators are barred at this lane count (tools/genref/generate.sh swept eighteen lanes' refusal codes into one reference once already), and the one authorized quiet-tree regeneration is TRADE-2's, at the cut. This should be on the cut checklist alongside the refusals regeneration, not discovered afterwards.

2.4 The gate list nothing points at

docs/ledger/VALIDATION_BACKLOG_2026-08-31.md is 453 lines of release prerequisites — nine numbered convergence gates, a seven-row deployment disposition, twenty named seeds — and WAVE.md, GOAL.md, README.md, AGENTS.md and docs/INTENT.md reference it zero times. ASPIRATION_ARCHAEOLOGY flagged this on 2026-08-30 ("someone should merge or cross-link it") and nothing did. It also still points the release CU gate at the M-61 sweep — that is, at the tier §1.2 just demoted.

---

3. Stale, listed and deliberately not fixed

These need more than a sentence, or belong to an owner who must make a call.

---

4. ~/dev/dragons-clutch — host-repo drift (listed only, not edited)

The successor subtree was cut to bb6d4edb today. The wrapper's own prose was not, and it describes a project three days behind.

None of the above was edited: the host repository is out of this lane's scope by charter, and its README and SECURITY are launch-surface copy.

---

5. What this sweep did NOT cover

Stated so the next reader knows the boundary rather than inferring coverage from silence.

• ~/dev/dragons-clutch/archive/ — gen-1, already swept by the Aspiration Ledger. Not re-dug.
• docs/evidence/ASPIRATION_LEDGER_2026_08_27.md and docs/ledger/board-archive-2026-08-27.md — excluded as archives. The ledger is itself a sweep artifact and the board is explicitly not authority. Their M-numbers were used as citations, not audited.
• docs/compost/ and COMPOST.md — the compost rules are their own process and were not evaluated.
• GOAL.md — a running done-log with stacked "old thrust" sections. It is historical by construction, so staleness there is not a defect. Not swept.
• node_modules — 2,000+ vendored READMEs, excluded.
• The 205-row raw commitment inventory. A full sweep of every forward promise in docs/decisions, docs/design, docs/evidence, OMISSION_INDEX and VALIDATION_BACKLOG produced roughly 205 rows across 63 files. §2 carries the ones whose truth changed today or that nothing outside a pre-compact file carries. The remainder — every U- row in the omission index, the whole of VALIDATION_BACKLOG, LIVENESS_CENSUS's Q1–Q10 costed queue, ORPHAN_DESIGNS_TRIAGE's thirteen sized rows, ASPIRATION_ARCHAEOLOGY's three tiers — is genuinely open, already written down in a document that owns it, and was not re-litigated here. Re-verifying 205 rows against the tree is a lane of its own; this sweep did not do it and does not claim to have.
• Whether the code matches the docs. This is a documentation sweep. Where it says a commitment is unbuilt, that is a grep for the named symbol (for example MAX_FEE_BPS, which exists in no .rs, .lean or .ts in the tree), not a behavioural test.
• Three cross-file ordering constraints surfaced by the commitment sweep and left un-audited, because each is a design decision rather than a doc defect: Dealer entry must not be built before the exit-affordability invariant (DEALER_EXIT_AFFORDABILITY_V1.md:275, LIVENESS_CENSUS:401,425); the Q1 ruling must precede the refcount which must precede CloseActivation (LIVENESS_CENSUS:408, RELEASE_LINEAGE_MIGRATION_V1.md:1346-1353); and Reaffirm lands before or with lineage commit 8 (RELEASE_LINEAGE_MIGRATION_V1.md:1167-1172). They are recorded here so a lane taking any one of them finds the other two.

---

6. The three that would cost the most if they stayed lost

1. Every "Hot CU" number in the tree belonged to a route that was demoted this morning. tools/gauntlet/hot-cu drives the Registry Hot continuation, which runs a constant +35,127 CU above the production top-level route, and its README, the gauntlet README, TIERS.md and VALIDATION_BACKLOG's release gate all presented its output as the Hot route's compute. CONTINUATION_ROUTE_FIX_OR_RETIRE sized the repair at half a lane-hour and named no owner, which is exactly how a correction goes missing.

1. An evidence document still carried the plan ember reversed. TRADE_DIRECT_ACTIVATION_WALL_2026_08_29.md described selecting maker keys for their CU draw and publishing them as "selected for CU" — rigging the demo and labelling the rig. The reversal was recorded in docs/ledger/SESSION_STATE_2026-08-31.md, which is rewritten every compact; the plan was recorded in an evidence file, which is not. The durable copy said the wrong thing.

1. A ruled fee band that nothing enforces. MAX_FEE_BPS = 500 is in four documents and zero programs. Its only live check is a bash comparison in a single staging script — a second author for a protocol bound, in a tree whose whole defect taxonomy today was two authors for one fact.