Genre-drift cohort: the hypothesis is refuted, and a better claim replaces it
Scored 2026-08-21 from the 23 run readouts in adr-066-test-runs.md. The question, as the workshop README posed it: do keep-reframes systematically move claims from causal/ontological to diagnostic/procedural framing? When it was written, two or three reframes had run and all of them drifted.
The tally
Of 23 runs, 16 changed a title. Six were plain keeps at the existing path (runs 4, 10, 12, 16, 19, 23), and run 11 was a merge whose source was deleted rather than renamed. No run was excluded for missing titles, though three old titles survive only as truncated filename slugs (runs 7, 15, 22); their genre is unambiguous from the surviving text.
Coding each title move as DRIFT (toward diagnostic/procedural), NO DRIFT (genre unchanged), or COUNTER (toward causal/ontological):
- DRIFT: 5 — runs 1, 2, 3, 9, 18
- NO DRIFT: 5 — runs 6, 7, 14, 15, 17
- COUNTER: 3 — runs 8, 13, 20
- Lateral, into CONDITIONAL, which the three-label scheme does not cover: 3 — runs 5, 21, 22. Forced into the scheme they read COUNTER, giving 5/5/6.
Drift is a minority under either coding: 5 of 16, against 11 that hold genre or move the other way.
Why this refutes the hypothesis rather than leaving it open
"Systematically" is the load-bearing word, and three clean reversals defeat it. Run 20 turned an ontological identity ("storing LLM outputs is constraining") into a scoped causal claim; run 8 turned a causal claim ("can turn tools into hidden schedulers") into an ontological one ("remains scheduling behind a tool interface"); run 13 turned a normative imperative ("reinforce facts at point of use") into a refuter-bearing predictive comparison. A machinery-wide attractor toward diagnostic/procedural framing would not produce those, nor the five genre-preserving repairs alongside them.
Two things make the drift column weaker still.
The coding is sensitive at the margin. Runs 9 and 18 land on "does not certify" and "do not establish" — negative universal epistemic claims. Read as denials inside the causal/ontological family rather than as diagnostics, DRIFT falls to 3 of 16, and all three (runs 1, 2, 3) sit in wave 1, the exact window the hypothesis was generalized from.
The original count had no denominator. The README recorded "two keep-reframes", run 2's readout called itself "third instance for the cohort thread", and third-episode-criterion-note-reframed.md independently calls itself "third case" for the same thread — the pre-wave count was both uncontrolled and internally inconsistent, because only drift-positive cases were being counted. Runs 5 through 23 supply the missing denominator and it is large.
What survives is narrower: all three uncontested drift cases are notes about KB and authoring method, where "analysis should separate" is a natural landing. The runtime, storage, and prompt notes did not drift. A weak domain-specific tendency is compatible with this corpus; a machinery-wide one is not.
The claim that fits the data
Reframes systematically trade universality for conditions and negations. Genre movement is a byproduct of where the surviving warrant sits, not an attractor. Fifteen of the sixteen title changes narrow scope, add a condition, or weaken strength while keeping the claim's kind:
- Conditions hoisted into the title: run 5 ("when required evidence fits"), run 21 ("when they suppress fallback search"), run 22 ("until write invariants stabilize"), run 3 (scoped to the consumption path).
- Actor or quantity narrowed: run 8 (cross-task transition policy only), run 14 (harmful retained inflow, not gross throughput), run 17 (model-resolved indirection only), run 7 (distinct rather than independent).
- Strength weakened to a limiting or negative form: runs 1, 6, 9, 15, 18, 20.
Run 13 is the single exception, and it moves to a stronger, more testable claim by adding a comparator and a refuter.
Six of the sixteen landings are negative or limiting universals. That is the cohort's largest single cluster, and it is exactly the shape a counterexample-driven repair produces: a defeated universal survives as a claim about what its evidence does not license.
The series verdict already names the mechanism, arrived at independently — "large-radius foundational notes state universals whose defeats are counterexamples, and counterexamples route scope and category repairs." instance-shaped defeats cannot license a mode conversion, so they route scope and category repair; genre follows the warrant. Runs 19 through 23 produced no prevalence and no priced-exception shapes at all, which is why those runs cluster so tightly on conditionalization.
What this does to the episode record's story
episode-record.md proposed the causal story behind the original hypothesis: "question-framings survive better than causal framings because a diagnostic question cannot be false", and "the installation's goals made diagnosis the attractor". The cohort falsifies the second half — diagnosis is not the attractor. The first half survives in weakened form, with a different subject: what survives review is the negative or conditional framing rather than the question framing, and it survives for a related but distinct reason. It states less, so there is less for a counterexample to reach.
That weakened claim is not free of the original worry. A repair that buys survival by stating less is the failure mode narrowing bought to survive review is paid for in content already names, and the refuter and citer guards exist to price it. The cohort says the guards are being applied to the right thing — condition-and-negation trades, not genre — and that no run in the series bought survival with vacuity, since run 3 was blocked from re-hedging and run 13 moved upward.
Status
Closed. The question is answered: refuted as posed, replaced by the conditionalization claim above, which is itself a restatement of the series verdict rather than a new finding needing its own evidence. No promotion is proposed — the transferable content is already carried by the narrowing note and the series verdict.