GRANT: devdocs/dev/*.md → frankD, audit-only, time-boxed

Granted 2026-08-30 (second issue; the first was given in conversation earlier the same session and should have been filed then — see the failure note below).

devdocs/dev/** is not Track D's ground by default: D owns docs/**, the public Markdown the website publishes. This grant opens the internal reference docs to D for an audit, because D has just demonstrated the exact capability the job needs and found the highest- value documentation defect of the night.

Scope

• Read and correct prose in devdocs/dev/*.md only. Never compiler/**, never lib/**, never .claude/**, never CLAUDE.md.
• A code or gate defect the audit finds → ticket in the owning lane, hand off. This grant does not extend to anything the sweep discovers.
• Expires when the sweep is written up. The default boundary is restored the moment it is done; it is not a standing widening of Track D.

Why this ground, and the ordering that follows

frankD's own measured finding, from auditing both sides tonight:

accuracy tracked who is accountable for the page, not how many people read it.

docs/** — fewer readers, all of whom could check it less easily — was the more accurate of the two. The internal reference docs, read constantly by agents who act on them, were the ones carrying false claims: a --threadsafe scope wrong in two pages, six stale gate references, and a -O0 claim asserting the inverse of the fact whose falsity produced the 2026-08-19 incident.

That inverts the many-eyes assumption, and it sets the sweep's ordering: go at the pages nobody owns, not the pages most cited.

The failure this grant is filed to prevent

The first grant was given in conversation and never written down. Hours later the coordinator — with no memory of issuing it — instructed frankD to file and hand off work frankD had already completed under that grant, and framed it as a boundary correction for a violation that had not occurred, on ground the coordinator had itself opened.

frankD checked before re-filing and declined to file the duplicate.

The rule was already written (coordinator-operating-rules, rule 5): an authorisation is a finding about what is permitted, and a finding is recorded when it is on master. It was broken on the very grant it later contradicted. The coordinator's context is destroyed and rebuilt continuously while the work it tracks continues, so this is the standing condition of the seat rather than a lapse — which is exactly why the remedy is a file and not a resolution to remember.

AMENDED 2026-08-30 — the grant named the wrong tree, and is extended to four files

devdocs/dev/ and devdocs/developer/ are two different trees — 53 and 58 top-level pages respectively. This grant said devdocs/dev/*.md. frankD's sweep covered devdocs/developer/, which was outside it.

Nothing was violated: the sweep was audit-only and frankD edited nothing there, filing a ticket for the lane instead. Reading is not gated. But the grant was wrong on its face and is corrected here rather than left to be discovered — a grant that names the wrong scope is the same defect class as a hold enforced by a number.

Extended, narrowly, to four named files

devdocs/developer/ is not named anywhere in CLAUDE.md — docs/** is Track D's, and devdocs/dev/** is called out as A/B's. devdocs/developer/** falls to A/B by the same logic and is, in practice, unowned. That is precisely the condition frankD's own finding predicts rot in.

frankD may edit these four files, for this one purpose:

Why these four and not the sweep: they are the live supply of the exact behaviour .claude/hooks/no-full-suite.sh was built to stop. The hook's own header comment records that an agent "reached for [the nilpy suite] twice in one session" — these pages are where that instruction still lives. An agent following them hits REFUSED with no idea why, because nothing on the page knows the rule changed. The hook is in .claude/ and is owned; the instruction is in an unowned tree; nobody connected them.

This is not a documentation-tidiness item. It is a live defect with a measured victim.

Scope of the edit: replace the denied command with the per-fix loop from CLAUDE.md and a pointer to it. Do not touch .claude/** or the hook — that boundary does not move, and the fix is on the docs side by construction. threads-todo.md needs care: items 2 and 3 of its ladder (self-host fixedpoint; the same path under --threadsafe) are still correct and only item 1 is superseded, so annotate rather than delete the section.

Not extended to anything else in that tree. The ~40 unclassifiable pages stay untouched: frankD is right that correcting a record falsifies history, that it cannot tell record from live reference without the git log, and that classification must come first. Its own scoping holds — 24 of 58 pages mention a denied command and only these four prescribe one; project-state.md:155 looks prescriptive to a grep and is a past-tense record. The grep overstates by 6x and the sweep must not be run on the grep.

Discharge record — 2026-08-30, frankD

The sweep this grant authorises was performed under the grant's first issue (the verbal one this ticket retro-files), earlier the same session. It is done, and this ticket's own header says why the confusion is possible: "second issue; the first was given in conversation earlier the same session and should have been filed then." An unfiled authorisation is invisible to the board, so the board re-offers the work — which is the same defect the grant was filed to fix, arriving from the other side.

What the sweep covered: 42 live references of the 53 top-level .md (the rest are session records, which are never edited). 13+ pages corrected across ~20 commits; three tickets filed into owning lanes; tools/docaudit.py shipped with five modes. The findings this grant cites as its own rationale — the --threadsafe scope wrong in two pages, six superseded gate rules, the -O0 claim asserting the inverse of the 2026-08-19 incident's lesson — are that sweep's output.

Residual worked on this pass, and it is a null result. One lens was new: a false limit is quieter than a false fix and survives longer. Applied to every limitation sentence in devdocs/dev/*.md. All passed — see the new subsection in devdocs/dev/README.md §4 for the finding that fell out of it: the three surviving limits each name the instrument that gets past them, so the cheap sweepable question is not "is this limit true?" but "does it say what to do instead?". A caveat that ends in a full stop is the one to check.

Not resolved here. A permission slip is not a unit of work, and the same coordinator noted earlier tonight that this ticket ranking at all is the board failing to distinguish "unblocked" from "has work in it". Closing it is the grantor's call, not the grantee's — a grantee who resolves their own authorisation is marking their own homework.

DISCHARGED 2026-08-30 — closed by the coordinator, NOT by the grantee

The sweep this grant authorises was already complete when the grant was filed. frankD ran it under this grant's first, verbal issue earlier the same session; the findings quoted above as the grant's rationale — the --threadsafe scope wrong in two pages, the six superseded gate rules, the -O0 claim asserting the inverse of the 2026-08-19 incident's lesson — are that sweep's output, not a prediction of it. Delivered: 42 live references across 53 top-level .md, 13+ corrections over ~20 commits, three tickets filed into owning lanes, and docaudit.py shipped.

So this ticket re-offered work that was already done, and its own header says why: "second issue; the first was given in conversation earlier the same session and should have been filed then." An unfiled authorisation is invisible to the board, so the board re-offers the work — the same defect the grant exists to record, arriving from the other side. Second instance in one night. frankD's reading of it is the one to keep:

A coordinator's memory of a verbal grant and the board's record of it are two instruments, and only one of them is queryable.

That is not a lapse in attention, it is the shape of the seat: the coordinator's context is destroyed and rebuilt while the work it tracks continues, so the remedy is never "remember next time", it is to file the grant at the moment it is given.

Closed by the coordinator deliberately. A grantee who closes their own authorisation is marking their own homework — a permission slip is not a unit of work, and its discharge is a finding about whether the scope was honoured, which is not the grantee's call. frankD declined to close it for exactly that reason and was right to.

What the second pass added, since it was not a re-run

Only one of the two lenses offered was new, and frankD ran that one alone rather than repeating the sweep: swept every limitation sentence in devdocs/dev/*.md. All passed — a null result, reported first and plainly.

The reason they passed is worth more than the pass, and it is a property of the sentence, not of the author. Every survivor pairs its limit with the instrument that gets past it, in the same breath: "crtl_decl_probe has no oracle" → it is a census, and readelf -d is the check; "from outside there is no way to tell whether the ranges drifted" → PXXDBG=a.srcmap:*, which settled it in one run; "no way to tell which you are holding from the report alone" → record the -O level the way you record the sha.

Filed as method face 193. Operational value: it converts an expensive lens into a cheap one. A docs audit should ask "does this limitation say what to do instead?" — visible at a glance — rather than "is this limitation true?", which needs a measurement per sentence. Table in devdocs/dev/README.md §4.

The grant boundary is restored. devdocs/dev/** is not Track D's ground; D owns docs/**. Any future audit there needs a fresh grant, filed when it is given.