# MOTOR-01 RESULT — the motor claim FAILS as registered; what survived is better (2026-08-27)

Run 2: qwen2.5:7b, three TRUE floor states (n40/n50/n70, concurrent arm-A
variance 0.00000/0.00023/0.00000 at 21 reps — deeply collapsed, eligible by
screen). All six arms, 21 reps, permutation tests, manip checks all logged.

## The pre-registered verdict

- **P1 (frame rotation re-opens spread, C1 > A AND C1 > D2, in >=2 of 3
  states): FAILS — 1 of 3.** n50 passed both (C1 var 0.00524 vs A 0.00023,
  p=0.0001; vs D2 0.00152, p=0.0007). n40 failed the load-bearing contrast:
  the irrelevant-puzzle null re-opened spread MORE than the rotation (D2 var
  0.00848 vs C1 0.00630, p=0.556). n70 missed both (p=0.116/0.088), and the
  text null D1 out-varied C1 there (0.01206 vs 0.01102).
- **P4 (nulls stay quiet): FALLS.** At deep floors, both nulls re-opened
  spread and moved means. Per the prereg's own words: the effect is context
  perturbation, not frame rotation, and the motor claim dies. **The guild
  case gains nothing from this experiment, and we say so** (also the
  prereg's own words).
- C2 > C1 at all three levels (demand account favoured); C2 cells at n50/n70
  flagged [CELL SUSPECT] by the mechanical check and treated accordingly.
- n50 stands as a fragile single-state anomaly, recorded, not argued from.

## What the data actually found (the bones-not-the-goal pattern again)

1. **Perturbation-escape generalizes.** A state collapsed across 21
  independent attempts is re-opened by ANY exchange — an irrelevant puzzle,
  a weather summary — not only by task-relevant frames. Run 5's single-model
  boundary observation, now at 21 reps across three levels. Collapse is
  shallow: the basin holds against repetition but not against almost any
  disturbance.
2. **Content has a distinct signature that perturbation lacks.** Arm B
  (true hint) at n50 and n70: mean up +0.10, variance 0.00000. Content does
  not scatter the state — it RE-COLLAPSES it into a better basin, uniformly,
  21/21. Perturbation decoheres; information re-coheres higher. In the
  electrical language: perturbation is heat, content is directed current.
  This is the cleanest new observation of the day and was not a
  pre-registered target.
3. **The instrument held.** Eligibility screen voided run 1's false
  positive; suspect-cell flags fired live on two C2 cells; the aggregation
  rule prevented a 1-of-3 anomaly from becoming a headline.

## The Faraday answer, revised by data

The walk question was: what reverses saturation into signal? The tested
answer (frame rotation) failed. The measured answer: **almost anything
disturbs a collapsed state back into spread (cheap, undirected, heat), but
only new information moves it to a better basin (directed, and it arrives
with LESS spread, not more).** The motor, if it exists, is not rotation of
the question; on this evidence it is the injection of content — and that is
the generator run forward, not backward. The reciprocity may be false at
this layer, which is itself a finding about the analogy's limits.

## Bounds

One model family for floors (qwen); toy domain; rotation wording three
phrasings per arm; C2 suspect cells; single day. Raw: motor-01-results.json,
logs, PREREG + Amendment 1 all committed.

---
## LOGGED CORRECTION 1 — 2026-08-27, post-hostile-interpretation-review
The pre-registered verdict (P1 fails 1/3, P4 falls, motor dies) SURVIVED
review; all arithmetic verified exact. The three salvage findings did not all
survive:

1. **Finding 1 downgraded to PROBE-DEPENDENT escape.** "ANY exchange" is
   falsified by this run's own data: two probe phrasings (C1-p0 and D2-p0 at
   n40) held the exact floor 7/7 while similar-length probes escaped. Also a
   token-only null (prepended text, no model turn, length-logged) was never
   run, and the permutation tests pooled three prompt strata against one
   (p-values overstate; within-phrasing variance still clears A's floor, so
   the direction stands). Restated: most but not all exchanges re-open a
   collapsed state; escape varies by probe; mechanism unresolved.
2. **Finding 2 WITHDRAWN as stated.** The "content re-coheres" signature has
   an unexcluded trivial mechanism: the hint ENUMERATES items (probe_picks
   logged), so 21/21 identical answers may be copy-through, not basin
   migration. B ran one phrasing against C/D's three (unfair variance
   contrast), and the n40 counterexample (B var 0.0054 ~ C1's) was omitted
   from the claim text — that omission was my drafting error. Downgraded to:
   an unexplained uniform cell at 2 of 3 levels. Decisive follow-up designed:
   B' arms with non-enumerating hints ("the optimal value is V") and a FALSE
   enumerating hint — if false enumeration also yields variance 0 at the
   false value, pinning wins and the finding dies.
3. **Finding 3 split.** The eligibility screen and aggregation rule genuinely
   held (run-1 false positive killed; n50 contained). The manipulation-check
   VOIDING did not: C2 phrasing-1 failed ok:false at all three levels and was
   voided three different ways (0, 3-of-7, 0 reps). Instrument defect,
   confessed here: the voiding unit (rep vs phrasing vs cell) was never
   specified and the harness does not enforce ok:false => void. C2 numbers
   ("C2>C1 at all levels") are computed over partially-void cells and are
   hereby marked unreliable. Fix required before any rerun.
4. n50 re-read as a testable MODERATOR, not an anomaly: it was the only
   state with nonzero concurrent-A variance — "the motor only turns where a
   crack already exists." Falsifiable with a graded-floor ladder, same rig.
5. Reciprocity sentence rescoped: "false for this model family at this
   layer." One family, one domain, one day.

Follow-ups queued (not run): token-only null; B' non-enumerating + false
hints; stratified permutation as primary statistic; graded-floor ladder;
one non-qwen floor family. The harness ok:false=>void assertion precedes all.