[finding] The teaching corpus has no lexical anti-drift pin, so a filter operator's prose can contradict its driver forever — the reachable half of "no gate can catch it"

[os-project-manager]

Filed while repairing #13539 (PRs #13577 and #13581). ⛔ Nothing was edited for this; it is a finding for triage to grade.

#13539 states that no gate could have caught the $exists mis-teaching, "because the teaching corpus does not execute". That was tested rather than assumed while repairing it, and it splits in two:

True as stated. The one gate that reads the offending block — os:check, via packages/spec/scripts/check-skill-examples.ts — type-checks the TypeScript inside a fenced example against the live spec. The false line in content/docs/protocol/objectql/query-syntax.mdx was a // comment inside an os:check block: the gate ran on that exact block, reported it green, and was structurally blind to the sentence that was wrong. A type checker cannot express "has a value" versus "the key is present"; both readings type-check identically. The same holds for the markdown table rows, which are not code at all.

False as a limit. The recurring failure is narrower than "prose disagrees with behaviour", and the narrow version is mechanical: a lexical pin over the teaching corpus, forbidding the co-occurrence of a filter operator with phrasing that names a semantic the platform does not implement — $exists beside "field exists", "existence check", "key presence", or "MongoDB $exists".

The precedent is already in the tree and needs no invention: scripts/check-role-word.mjs is a shrink-only word ratchet over exactly the right roots (content/docs and skills, with generated references/ excluded because the spec source is the fix site there), with a baseline, a --self-test, and a maintainer-only widening path. What it does not have is a second word: WORD is a single module-level regex, so a second term is a change to its shape rather than a row in a table.

Why this is worth a card rather than a note

The corpus is where AI authors read the contract. A false line there is copied into customer projects and cannot be discovered by running anything. The two skills/** sites in #13539 were false for a day and shipped to customers; the class survives for months elsewhere precisely because the only instrument is someone reading it.

What a taker should weigh, honestly

• Cost is a new validation surface. A new gate needs registration, a self-test, a workflow entry and a baseline. That is why it was ⛔ not smuggled into $exists docs and two PUBLISHED skills still teach key-presence — false since the has-value alignment landed, and no gate can catch it because the teaching corpus does not execute #13539's wording repair, which is bounded to files that were false.
• A word ratchet cannot judge sense. It catches spellings, so it is only as good as the term list, and each term costs a baseline entry for the legitimate uses (has(record.foo) in objectstack-formula is real key existence and must stay green).
• The alternative worth pricing against it is generalising the term list rather than hard-coding $exists: the same shape would have caught the retired $regex spellings and the visibleWhen claims that docs(ui,protocol,lint): a form SECTION visibleWhen binds current_user — retire the claim from the three prose sites #13074 left unswept #13532 swept by hand.

Refs

#13539 (the repair, and the claim this tests) · PR #13529 (the alignment that made the lines false) · #13531 (the same operator's stale cells inside packages/spec, already dispatched) · scripts/check-role-word.mjs (the precedent) · packages/spec/scripts/check-skill-examples.ts (the gate that ran green over the false line).