Every threshold in the instruction is decidable, after two of them were not

Run: gate/decidability.py → gate/decidability.json Date: 2026-09-20, extended 2026-09-21. INSTRUCTION.md v1.0.0 → v1.2.0.

Question or issue resolved

INSTRUCTION.md was written to remove discretion from the Nooscope method by reducing every judgment to a signalling question with a fixed threshold. Writing that document does not establish that it worked. A threshold that states a rule but still needs the rater to decide what the rule means has removed nothing, and it is worse than an openly acknowledged judgment call, because it reads as settled and so never gets argued about.

This audit asks whether each of the 35 questions is decidable as written — before two raters are put in front of it, and at a cost of an afternoon rather than a cycle.

The findings

Seven of 35 questions failed the first pass. Two were real defects; five were a defect in the audit.

The two real ones were in the document as published:

• 7.3 required the caveat field to name what a critic should attack first, and judged a list that did not "has not been written in good faith." Good faith is an open predicate wearing the clothes of a rule. Two raters can disagree about it indefinitely while each believes they followed the instruction. Replaced with a closed list: the caveat must name a specific widener, a specific parameter departure, a specific rejected channel or trial class, or a specific source limitation. Naming none of those is No.
• Domain 6, the type and tier decision tree, stated its thresholds as consequences — "No → Type C or E" — rather than as answer rules, which is the form the other seven domains use. The intent was clear and the wording was not, and a rater who reads seven domains in one grammar and the eighth in another is being invited to improvise. Restated: each question now names the record field whose presence decides it.

The other five were the audit under-recognizing closed forms. Its detector accepted = Yes and missed two constructions that are just as closed: a decision arrow, and a flat bolded No. It reported 3.3, 6.1–6.4 and 8.4 as defects when 3.3's threshold is the single most closed sentence in the document. The audit needed auditing, which is the ordinary condition of a first-run instrument and the reason the fix below exists.

After both fixes: 35 of 35 decidable, 0 undecidable, 0 requiring data the record does not carry.

A clean sweep is also what a broken detector returns, so the detector is now tested before it is trusted. Four control questions with known answers run first — an open predicate that must be caught, a bare appeal to the assessor that must be caught, a numeric threshold that must pass, a closed enumeration that must pass. If any control comes out wrong the audit refuses to run rather than reporting a number. Widening the detector to recognize arrows and bolded answers was a correction; the open-predicate list it screens against was not touched, so a question containing one is still a defect however it is phrased.

One reclassification is recorded rather than buried: question 7.3's required input moved from human_judgment to reading_record. That changed because the question changed — a threshold naming four specific record contents is readable from the record, and one saying "in good faith" is not — and not to improve the count.

A second defect class, found by the register rather than by the audit

The audit passed INSTRUCTION.md 3.1 — the geometry assignment — on both runs. The conformance report then showed two readings with no geometry at all, and the reason was not a missing record. METHOD.md §3.3 named four geometries "or another named arrangement", and that escape clause sat in the middle of a decision the instruction describes as mechanical. A rater meeting an unfamiliar array could name it whatever seemed reasonable, and two raters could name the same array differently while each followed the instruction exactly.

The audit could not see it because it was testing the wrong half. A threshold can be perfectly evaluable and still undecidable if the set it maps onto is open. Evaluability is a property of the sentence; completeness is a property of the codomain, and v1.0.0 of this audit tested only the first.

Three changes followed, and the middle one is the substantive one:

• The audit now screens for escape phrases in enumerations — or another, or other, such as, etc, including but not limited to — as a defect class of its own, with two added controls: an enumeration carrying an escape clause must be caught, and a closed presence test must still pass. Six detector controls now run before any result.
• METHOD.md §3.3's geometry set is closed as of 0.3.7. An arrangement outside it blocks the reading and is added by amendment. That is slower by design; the alternative is a vocabulary that grows silently and cannot be grouped or compared afterwards.
• probe_linear was added for penetrating silicon probes with contacts along a shank — the arrangement the register's two mouse readings were actually taken on, and which the old set had no room for. It is kept distinct from mea (planar, slice and culture work) and seeg_depth (penetrating at clinical macro scale), because those are three different sampling geometries and PCIst depends on which was used.

INSTRUCTION.md is now v1.2.0 with 36 questions, all decidable. The new question 3.5 makes the closure operational: a census returning an unmapped channel type is not issuable.

The conclusion

The instruction is executable. That is a weaker claim than it sounds and is worth stating precisely: every threshold can now be evaluated mechanically from data the register already carries, which means a disagreement between two raters can be traced to a specific question rather than dissolving into temperament. It does not mean two raters will agree. Decidability is necessary for reliability and nowhere near sufficient, and the only evidence that will settle reliability is two independent executions, which METHOD.md §9.3.1 now provides a disclosed ladder for.

Two things follow.

Three defects, found three different ways, and only one of them by the audit. The open predicate and the inconsistent grammar came from the text audit; the open enumeration came from the live register contradicting a document that had just passed. An instrument that checks a standard will only ever find the defects it was built to look for, which is an argument for running the conformance report against real records as well as auditing the prose — they fail differently.

The defect rate in a first draft of this kind of document is about six percent, and both defects were in the two domains where the underlying judgment is least numeric — free-text caveats, and a decision tree. That is where to look first in any future amendment, and it is an argument for running this audit on every version rather than once.

This audit belongs in CI. It costs under a second, it has controls, and it turns a class of defect that would otherwise surface as unexplained rater disagreement months later into a failure at the moment the text is edited. It is now wired in as a test.