# PREREG — MOTOR-01: can a frame rotation reverse saturation into signal?
*Registered 2026-08-27, before implementation and before any run. Amendments
after this point are logged, never silently edited.*

## The claim under test (the Tesla/Faraday reciprocity, operationalized)

The generator (measured, sealed 2026-08-24): spread in unaided attempts
predicts responsiveness to injected content. The motor (this experiment):
**a content-free change of frame re-opens spread in a saturated state.**
Induction analogy taken literally: a steady field (pushing the same question
harder) induces nothing; a changing field (rotating the question) does.

## Design

Rig: the sealed saturation rig (seeded knapsack generator, DP true optimum,
no judge model, 5 reps per cell). Subjects and states reuse the sealed
ladder so prior baselines are the controls.

Per saturated state, four arms, all followed by a RE-ASK of the ORIGINAL
question whose answers are what gets scored and variance-measured:

- **A. Baseline**: re-ask only (replication of sealed baseline variance).
- **B. Content** (known result, run as anchor): true hint, then re-ask.
- **C. FRAME ROTATION** (the motor): a probe exchange that changes the frame
  without carrying content — e.g. "solve the inverse problem: the WORST valid
  selection," or "state two maximally different near-optimal packings," or a
  domain re-skin (ship's hold, ration kit) with identical numbers. The probe
  answer is discarded; then re-ask the original. Rotations carry NO values,
  weights, or optimum information beyond what the original prompt carries.
- **D. Null-rotation control**: an interlude of matched length and zero
  relevance (weather paragraph), then re-ask. Distinguishes frame-change from
  mere context perturbation / context-length effects.

States: floor-saturated (qwen2.5:7b mid30 — collapsed on a wrong basin;
gemma3:4b any level — decoder-collapse boundary case, run labelled as such),
and solved-saturated (glm-5.2 n20/n40, haiku-4.5 two levels). Spread
control: qwen n16 (rotation at a spread state should not obviously help and
may hurt — recorded either way).

## Pre-registered predictions

- **P1 (the motor)**: at floor-saturated states, arm C variance > arm A
  variance (spread returns), and arm C mean >= arm A mean (the motion is not
  purely destructive). Operational threshold: variance rises above the 0.0005
  band used for spectrum labels in the sealed runs.
- **P2 (the asymmetry that makes it falsifiable)**: at solved-saturated
  states, arm C does NOT re-open spread (variance stays < 0.0005, mean stays
  within 0.01 of baseline). The motor moves collapsed states, not solved
  ones. If rotation degrades solved states, that is a FAILURE of the motor
  frame and gets reported as such.
- **P3 (content-freedom)**: rotations verified by construction to carry no
  task content; any C effect with D null is frame, not leakage and not mere
  perturbation.
- **P4 (the null that must hold)**: arm D moves neither variance nor mean
  beyond baseline noise at any state. If D re-opens spread, the effect is
  context perturbation, not frame rotation, and the motor claim dies.

## Manipulation check (mandatory per the 2026-08-23 invariant lesson)

Before any arm comparison is reported: verify the probe actually changed
behavior in the probe turn itself (the rotated answer differs structurally
from the original-task answer). An unapplied manipulation voids the cell.

## Guild stake, stated up front

If P1+P2 hold, "send the converged model where it diverges" stops being a
metaphor: guild rotation becomes a measured operation on saturated states,
with the variance band as its trigger and its receipt. If P4 falls, the
guild case gains nothing from this experiment and we say so.

## Honest bounds, declared now

Toy domain; 5 reps; one rotation family per arm (rotation wording is a free
parameter and a known demand-characteristics surface — the wringer decides
whether the C prompts pass); floor states available in only one-two model
families; gemma cells are boundary-labelled and never counted toward P1.

---
## LOGGED AMENDMENT 1 — 2026-08-27, pre-run, per hostile wringer review (verdict: RUN WITH FIXES; full review in session record)

1. **Rep protocol (was FATAL, unspecified):** every rep is an independent fresh
   session containing exactly [probe exchange] + [original question]. No prior
   attempt in context, no sequential re-asks, sampler settings identical to the
   sealed baseline runs.
2. **Probe answers never enter re-ask context (was FATAL, leakage):** re-ask
   sessions contain the rotation PROMPT plus a fixed neutral stand-in
   "(answered)" — never the probe's actual answer. Any probe that elicits an
   item set over the instance values is BANNED. The manipulation check runs in
   separate throwaway sessions, decoupled from treatment.
3. **Arm C split (was FATAL, demand characteristics):** C1 = frame-only
   rotation, one question one answer, no variety demand (domain re-skin;
   approach-only probes) — the ONLY motor-evidence arm. C2 = variety-demand
   rotations (inverse, two-packings), run as the demand-characteristics
   CONTROL; C2 > C1 variance is evidence FOR the demand account. Three
   phrasings per rotation type.
4. **Null upgraded:** D1 = irrelevant text WITH a model turn (summarize it).
   D2 = matched-effort irrelevant task (unrelated puzzle, model answers).
   Motor evidence requires C1 > D2, not C > weather.
5. **Concurrent eligibility screen:** a cell enters P1 only if its arm A
   variance THIS RUN < 0.0005; otherwise VOID, recorded, never relabeled. All
   comparisons vs concurrent A.
6. **Power:** 20 reps per cell for variance claims; permutation test (pooled
   C1+A rep scores, arm labels permuted, one-sided alpha 0.05); P1 requires
   the effect in >=2 of the eligible floor states; single-rep-driven
   confirmations reported as fragile.
7. **P2 demoted** to a boundary condition with zero evidential weight (ceiling
   arithmetic, per runs 1-2); the guild stake rests on P1 + P4 + the P2
   degradation clause. A sub-ceiling solved state added if one exists on the
   ladder.
8. **Manipulation check made mechanical:** probe item-set != modal baseline
   item-set (or predeclared string check for selection-free probes), applied
   before re-ask results are unblinded.