A fact a source can settle is read, not recalled

The rule

Before stating what a record, a rule, a spec, a command or a file says, open it and read it. Recall is for locating the source. It is never for quoting one.

The bound matters, because "verify everything" is paralysis rather than rigour. The rule binds an assertion that is about to be written into a record, a commit message or a brief, or used as the grounds for a decision. Conversation may be provisional, and the honest form there is to say which it is: "the intent says X" and "I recall the intent saying X" are different sentences and only one of them survives being wrong.

Why

Measured, in one session, four times. Each was a document the author had read carefully hours earlier, and each was stated with confidence.

• The grounds put to the maintainer for degrading a gate said the design licensed it "when a signature proved noisy in practice". The design's actual condition is degradation only on observed noise, and the evidence in hand was a constructed corpus over an instrument that has never run. The maintainer caught it by asking whether the claim was really true.
• A captured issue asserted that no non-test file in a package referenced the agent definitions. One already declared the directory constant and read it. The builder that the capture briefed caught it.
• A residue paragraph claimed an enumeration "cannot fall behind the surface it guards". Two written lists survived inside it, and the guard's own header said so in capitals.
• A synthetic battery's result was reported as "measured: they do not", of an instrument no reading had ever been dispatched to.

The pattern is the finding. Confidence tracked familiarity, not accuracy. Every error was about material read recently and attentively; recency produced confidence, and confidence displaced the check. The documents nobody had looked at that day produced no false claims at all, because nobody felt able to assert anything about them.

The cost is asymmetric to the point of being decisive. Opening the file costs seconds. A recalled claim enters a ledger record, briefs an agent, or becomes the premise of a maintainer's decision, and is then acted on by people who cannot tell recall from reading because the sentence looks identical either way.

Why this is not a memory item

It already was one, and that is the evidence. The session that produced the four errors above carried an agent-memory note from a previous session saying, in substance, to check a principle's letter before citing it. It was in context from the first turn. It did not bind.

A rule enforced only by an agent remembering to apply it fails precisely when the agent is confident, which is exactly when it is needed. That is the promotion signal memory-graduates-to-record describes, in its stronger form: not merely a lesson recalled twice, but one recalled and then not followed, which shows the home rather than the lesson to be at fault.

The gate

• Given a statement about what a record, rule, spec, command or file says, when it is written into a record, a commit message or a brief, or used as grounds for a decision, then the source was opened for it, or the statement is not made.
• Given a fact stated from recall in conversation, when it is stated, then it is marked as recalled rather than checked, so a reader can tell the difference before relying on it.
• Given such a claim found to be wrong, when it is corrected, then the correction says whether the original was read or recalled, because a corpus of those is what would justify automating any part of this.

Fit

itd-195 is the complement and neither contains the other. That rule governs a claim about how code BEHAVES, and its remedy is a test. This one governs a claim about what a source SAYS, and its remedy is to read it. A quotation cannot be made executable and a behaviour cannot be settled by quotation, so the two are disjoint halves of one instinct: do not write what you have not checked.

Staging

Adopted 2026-08-31 at the documented-protocol rung. No detector, and it is not obvious one is possible: distinguishing a recalled assertion from a read one is not visible in the text, which is the whole difficulty. What could be mechanised later is narrower, a check that a record citing another record by id quotes text that is actually in it.

The existing corpus does not conform and is not assumed to. The four errors above were corrected; nothing else has been surveyed.