A claim about how the code behaves is executable, or it is not made

The rule

A comment, doc or record that states how some part of the codebase behaves is backed by something that runs, or it is not written.

The forbidden shape is a list: "these are the readers and this is what each does", "these are the places this value is copied", "every surface handles this the same way". Such a list is words. Nothing exercises it. When the code moves, the list stays and becomes false, and it is believed because it is specific.

Where the fact is worth stating, state it as a test that establishes it by running the thing. Where it is not worth a test, do not state it: say what the code in hand does and stay silent about the rest.

Why

Measured, not reasoned. On 2026-08-30 two unrelated claim families on two unrelated branches each took FOUR rounds to correct, and each correction introduced a fresh error:

• an enumeration of what nine record stores' readers do with a duplicated key;
• an enumeration of where a three-token vocabulary is copied.

The error rate did not fall with effort, which is the signature of a format problem rather than carelessness. A correction to an unverifiable list is itself unverifiable, so each round produced a new claim nothing could check.

Making the first one executable settled it in one round AND explained the four failures. The written version had one row per store; several stores have several readers, and those readers DISAGREE — one refuses a malformed record, another quietly keeps the first value. A list shaped one-row-per-store cannot hold that, so every hand-written version had to drop a row. Four rounds of checking whether rows were right never asked whether the shape could carry the truth. The executable version has fourteen rows for nine stores.

It then caught two false statements in the orchestrator's own briefing, which four review passes reading prose had not.

The gate

• Given a comment, doc or record stating how a component behaves, when it is reviewed, then it either cites something that runs and would fail if the statement stopped holding, or the statement is removed rather than corrected.
• Given such a claim found to be wrong, when it is fixed a second time, then the fix is to make it executable or delete it, never to reword it again.
• Given an enumeration made executable, when a new member appears, then a gate fails rather than the list ageing.

Fit

This is the general rule; two adopted principles are its special cases, and neither covers it. enforcement-claims-are-facts forbids describing a gate that does not run — a claim about EXISTENCE, judged when written. This rule is about a claim STAYING true: a comment can be exactly right the day it is written and false six months later, which that principle does not reach because it was a fact at the time. guards-prove-themselves requires a refusal path to ship with a test that watches it refuse — the same instinct, scoped to guards. Filing the general form beside two adopted specialisations is structure rather than a third copy, and if it is adopted the two should say they are instances of it.

Staging

Adopted 2026-08-31 as a STUB. The rule stands and the two adjacent principles now point at it as instances. Nothing else is worked through: the sweep is unstarted, the gate is prose rather than a detector, and the relationship between this record and its two special cases is asserted rather than reconciled. Deliberate — the cold-reading experiment has the floor, and this waits for it.

The rule binds new work immediately. The existing corpus does not conform and is not assumed to: a sweep of the whole codebase for prose that states how other components behave is the work this intent carries, and it has not been done. Two sites are already converted (the record-store reader table) or removed (the vocabulary copies); everything else is unsurveyed.

The sweep is deliberately not automated first. What counts as "a claim about behaviour" versus ordinary description is a judgement the corpus has to teach, which is script-first-mvp: survey by hand, grade what is found, and let the detector follow the corpus rather than lead it.