core/docs/specs/postures.md

Posture Statements

Authority: R-1, R-5, R-6, R-11 (docs/assessment/50-rulings.md), ruled 2026-07-28. What this file is for: four decisions that cost nothing to make and something real to leave silent. Each names a criterion — the evidence that would change the posture — because an unstated boundary reads as an unexamined one, and a future reader cannot tell a deliberate limit from an accident.

None of these flips a flag. Three record where CORE deliberately stops; the fourth records a proposal withdrawn because measurement dissolved its premise.

---

P-1 · Efferent action is deferred, with an entry criterion (R-1, ruled A · G-12)

CORE's telos ends at articulate / learn / replay. AGI-grade generality ordinarily implies acting on the world, and no system-level statement existed either way.

Posture: deferred, in scope eventually, and nothing moves until the criterion is met.

Entry criterion. No efferent surface beyond typed, trace-folded tool operators (ADR-0018) until (i) the chooser (G-4) exists and is governed, and (ii) an efferent falsification bench exists at the standard docs/specs/ sets for M2.

Both halves are load-bearing. Without (i) CORE would act with no principled account of what it should do next — the Candidate Register's open seat. Without (ii) it would act with no way to be shown wrong, which is the one thing this architecture never permits of a serving capability.

Why deferral rather than exclusion: declaring efferent action out of telos is coherent and cheap, but it forecloses a scope decision on a system whose chooser has not been designed. Deferral costs nothing exclusion saves and preserves the option. ADR-0211's prohibition on motor/efferent units in the falsification bench v1 is narrower than this and remains in force — it is a bench-level rule, not the system-level statement recorded here.

---

P-2 · Identity enforcement stays scoring-only until a discrimination bar is cleared (R-5, ruled A · G-11)

identity_wave_gate is off; identity_action_surface is off and documented "NOT authorized for live activation." That is a deliberate, well-reasoned posture. What was missing was the criterion, and a posture with no criterion decays into an unexamined default.

Posture: scoring without blocking, until live refusal is authorized by measured discrimination.

Authorization bar. Live identity refusal requires a discrimination report on a named, held-out corpus of benign and adversarial traffic, showing a separation between the two on the certified metric, at a floor named before the run. γ_id is the only certified metric today; identity_action_surface's thresholds are uncertified placeholders and ADR-0246 §6.3 shows it refuses benign and adversarial traffic alike on the declared placeholder frame.

Why the bar is stated in advance: the pre-registration discipline that made ADR-0252 §5's NO-GO full credit applies here for the same reason. A refusal gate authorized on a floor chosen after seeing the numbers is not evidence, and identity refusal is a surface where being wrong is expensive in both directions — a false refusal is a denial of service to a legitimate user, a false accept is the attack succeeding.

Scoring-without-blocking stays honest only while the path to blocking exists. This is that path.

---

P-3 · Non-text ingest is deferred, with the falsification bench as its standard (R-6, ruled A · G-17)

59 sensorium modules exist and reach no serving path; projection heads do not exist. The position paper is honest about this. It was neither built nor deliberately postponed — which reads as drift.

Posture: explicitly deferred. Not a gap; a decision.

Entry criterion. A non-text modality enters serving when it meets the same falsification standard as text: a named held-out corpus, a falsifiable predicate set with holds/bites pairs, and wrong=0-or-refuse on the serving path. No modality is admitted on the strength of the substrate being able to represent it.

Why this criterion and not a roadmap: the sensorium's existence is not evidence that it works, and the arc that produced this file spent its length distinguishing built from proven. Naming the bench as the standard makes non-text ingest earn its way in on exactly the terms text did, and makes the 59 modules a capability awaiting evidence rather than an unexplained absence.

---

P-4 · The interim fabrication gate is WITHDRAWN — measurement dissolved its premise (R-11, ruled B → second ruling)

R-11 ruled B: measure first, then re-ask. The instrumentation ran on 2026-07-28. It did not return a rate; it returned a reason the question could not be asked in that form, which is the more valuable outcome and exactly what an instrumentation step is for.

The proposal. While the G-2 fabrication fixes are held for their ADR, serve an interim posture: refuse to hold a reading outside the verified 19-construction inventory.

What the measurement found.

distinct serving-path inputs measured	11,199 (deduction + curriculum practice corpora)
clauses, at the reader's own sentence split	23,562
clauses outside the verified inventory	0

That zero is not a clean bill of health, and reporting it as one would be the error this file exists to avoid. It is an artifact with a cause:

One of the 19 verified constructions is atom_fact, whose template is {p} — a single slot with no literal anchor. It matches any string.

So "outside the verified inventory" is not a well-defined syntactic property. Nothing is outside it. The proposed gate would refuse nothing while appearing to be a safety mechanism — a mechanism whose failure state is indistinguishable from its success state, which is this repository's dominant defect class.

This reframes G-2, and the reframing is the finding. every dog is a mammal → member(every_dog, mammal) is not the reader admitting an out-of-inventory construction. The surface is admissible; the reader assigns it the wrong relation. Likewise asserted(furthermore). The defect is in the mapping, not in the admissibility set — and no gate over the admissibility set can catch it.

Second ruling (DELEGATED, 2026-07-28): option A is WITHDRAWN as unimplementable as specified — not deferred. A deferred option is one that could be built later; this one cannot be built at all against the inventory as it stands, and recording it as "deferred" would leave a future reader to rediscover why. Option C (no interim gate; the fixes stay held for the fabrication ADR) is the operative posture.

What the fabrication ADR inherits, stated so it is not lost. It must define the admissibility boundary itself, because the verified inventory does not supply one. Specifically: either atom_fact is narrowed so that admissibility is decidable, or the guarantee is relocated from admissibility to mapping correctness — a check that the relation assigned to an admissible surface is the one that surface licenses. The second is the shape the evidence points at.

Method note, recorded because the number nearly shipped wrong. The first two passes of this measurement were both invalid — one matched templates against whole multi-sentence inputs (reporting 100% out-of-inventory, when every clause was in-inventory), the other misread the sentence-splitter's tuple and measured "." 23,562 times. Both produced confident, plausible, wrong numbers. The catch-all was found only by a non-vacuity check — asserting the matcher could still say no to "most birds can fly" — and it could not. A measurement that cannot fail is not a measurement.