the adversarial-review floor assumes below-the-line artefacts are covered by their gates, and today they were not

the adversarial-review floor assumes below-the-line artefacts are covered by their gates, and today they were not

adversarial-review-scales-with-blast-radius draws a floor, and gives a reason for it:

Artefacts below that line — ledger captures, comments, routine pull requests — are never blocked on adversarial review: they are covered by the gates they already have, and capture in particular must stay frictionless.

Two clauses, and they are doing different work. "Capture must stay frictionless" is a cost argument and it is sound. "They are covered by the gates they already have" is an empirical claim, and it is false.

The counter-evidence

Four errors on 2026-08-22/23, every one in a below-the-line artefact, every one passing record-lint, docs-lint and lint-reviews:

• A capture filed as a standalone process observation when it refines iss-213, which is what gives it its weight. Routing, not form.
• A proposed new principle that one-canonical-primitive forbids, because enforcement-claims-are-facts already owned the ground.
• Two ADR links in an issue body resolving to nothing, because links_resolve does not run over .abcd/work/ (iss-2608230752354927).
• Record repairs and a capture written into a shared checkout another session was about to branch from (iss-2608230847432285).

The first two are the interesting pair. Both are errors of routing and relation — which record this belongs to, which existing record already owns it — and no deterministic gate can see them, because both artefacts were perfectly well-formed. A linter checks whether a record is shaped correctly. It cannot check whether it is the right record.

What follows, and what does not

The floor should stay where it is. Its own argument for the placement is correct and this changes none of it: mandatory review on captures would suppress the ledger, recording a finding must cost less than fixing it, and reviewers trained on mandatory verdicts over trivia skim the one that matters.

What should change is the stated reason. Below-the-line artefacts are unblocked because the friction would cost more than the errors do — not because their gates cover them. Keeping the false clause is itself an instance of the class recorded in iss-2608230847432286: a reader who believes the gates cover these artefacts stops compensating, which is the precise harm that record describes.

The bound, offered weakly

A peer session already present and already communicating is a reviewer at near-zero friction, so the cost argument the floor rests on does not apply to it. That suggests a standing convention: announce what you are working on to concurrent peers and engage with what they announce.

This is offered as a proposal, not a finding, and the evidence for it is thin. iss-2608230847432286 records why: four concurrent sessions of one model in one repository is not the independence evaluator-outside-the-loop asks for, a shared blind spot would have been invisible to all four, two of the catches were a peer naming a record the author had not read (recall, not review), and at least one was luck. Peers also produced errors — a record-merge proposal that did not survive scrutiny, a threshold that collapsed on a wider sweep, a misidentified record owner.

A second reading is available and the record should not suppress it. Today's coordination ran entirely on a convention with no mechanism behind it: four sessions wrote in one repository, nobody lost work, and what held was "a diff you did not make is a peer's work" — a sentence in AGENTS.md that nothing enforces. That supports either conclusion. The convention is sufficient and needs no machinery, or the convention got lucky four times and the sample is too small to tell which. The isolation property the same document claims did fail on the day (iss-2608230847432285), and the convention is what absorbed it, so at minimum the two are not independently reliable.

Routing, per itd-84: the capability half is already owned by itd-33 (coordination contract) and iss-2608220750029993 (presence detection), so a second intent beside itd-33 would be the third copy. The stance half is a principle bound plus a line in AGENTS.md § Concurrent sessions, whose five existing concurrent-session records are all hazard-framed and none of which treats a peer as a resource. Verdict SPLIT, HOLD on the intent.

Promotion trigger, per recurrence-is-signal: if this recurs among sessions that are not coincidentally co-present, the convention has earned more than a caveat.

Grounds

• pursued: a principle whose stated reason matches its evidence keeps readers compensating for the gates' routing blind spot; shown wrong if a later reading of the principle again infers that record-lint or docs-lint vouch for a capture's routing