Fourth episode: the first in-pass adequacy assessment

Pass 20260819T105132Z-2a6408 (codex, independent of both authoring sessions) on kb/notes/instantiation-alone-cannot-model-agent-learning-across-sessions.md; revision committed as 25873411. This is the run that met the repair-dispositions proposal's adoption criterion and supplied the decisive evidence for ADR 066's no-new-machinery design. The three calibration readouts are recorded in the README's completed thread; this file consolidates what the episode established beyond them.

Acceptance, not rejection, as the outcome

Disposition keep, no GLOBAL defeat, all six premises HOLDS at closing. The pass applied the two-stage structure unprompted — "pricing signatures route the immutable-class model to adequacy assessment; they do not prove adequacy" — and decided adequacy on the entered declared-use commitment (fence integrity, not exception frequency). The open prevalence dimension did not force a defeat: the criterion returned a calibrated outcome on a partially open record, the exact question the operator scheduled this run to answer. The falsifier stays live in the report's Open items: a corpus showing ordinary unmarked class change in the compared systems would dissolve the fence.

Comprehension under attack

Closing accessibility proposed glossing the domain-pricing criterion as "common and cheap enough"; the pass refused the gloss as a misstatement — "domain-attested marking routes; declared use and consequence bounds decide." An external attacker read the criterion note correctly and defended its content against a simplification that would have reintroduced the defeated pricing-sufficiency reading.

Advisory/authoritative separation held

The pass classed the workshop record as advisory context only — not warrant — and still independently found and repaired the scope defect flagged earlier in this thread: proposal-selection stated as necessary for all cross-session learning. The repair matches what the routed-attention note recommended (general requirement restated as an evidence-responsive retained update; proposal-selection narrowed to the readable-artifact subtype; governance — permissions, review, versioning — separated from the loop operating within it), arrived at from the note and reports alone. Convergence without warrant-borrowing is evidence the finding was in the material, not in this workshop's framing.

Residuals (recorded, unactioned)

Two friction warrants — why governance devices are necessary rather than merely possible, and how recurrence is distinguished from one-off need before promotion; six connect candidates deferred to editorial selection; the criterion-note gloss accessibility item. These belong to the instantiation note's future maintenance, not to this workshop.