# 002 qa-head-dose — RESULT (harvested 2026-08-01, tier T3-exploratory)

**Verdict: NEGATIVE across the full dose range — a supervised role head reaches 100%
held-out accuracy without inducing ANY transferable role code.**

## Numbers (per dose; 1.7e8 tok each, budget-matched to each other)
| λ_qa | aux head val acc (final) | battery primary linear | primary mlp | val_f1 |
|------|--------------------------|------------------------|-------------|--------|
| 0.01 | 0.5283 (failed/collapsed) | 0.492 [0.483,0.499] | 0.489 | 0.5754 |
| 0.1  | 0.9978 | 0.489 [0.473,0.506] | 0.477 | 0.5754 |
| 1.0  | **1.0000** | 0.495 [0.483,0.507] | 0.484 | 0.5732 |
(λ=0 point: baseline A_D_s0 primary 0.501 from 001's run. Battery mechanical verdict
INSTRUMENT_FAILURE per the organism-genitive gate, as pre-flagged; primary AUC read per prereg.)

## Interpretation (scoped)
- **Manipulation check passes at λ≥0.1**: the linear aux head on z decodes agent-vs-patient
  at ~1.0 on held-out propositions (same vocab/families). So z DOES carry linearly-readable
  role information for the training distribution — when something asks for it.
- **Zero transfer**: battery primary (novel vocab + novel constructions) flat at chance at
  every dose; no dose-response; no reconstruction cost.
- Combined with 001 (contrastive variant of the same phenomenon): **two independent
  objective types are satisfied by non-generalizing, lexically-anchored role codes.**
  Training pressure produced role *decodability*, not role *abstraction*. The pressure
  that matters is evidently cross-lexical/cross-construction invariance, which neither
  objective supplies.
- λ=0.01 aux-head collapse (0.99 mid-training → 0.53 final) noted as-is: low-dose head
  learning is unstable; not interpreted further.

## Brier (prereg-lite)
(a) monotone dose-response (0.40) → F · (b) λ=1 primary CI-lo>0.6 (0.40) → F ·
(c) aux acc>0.9 at λ=1 (0.65) → T · (d) val_f1 within 0.03 across doses (0.60) → T.
Mean Brier **0.166**.

## Caveats / follow-ups
- Aux held-out split is proposition-level, same vocab/families — it certifies decodability,
  not abstraction (that contrast is the finding). Single seed per dose; shares 001's
  mix-in design.
- Feeds directly: 003 (structured-decoder = generation-side pressure with word-level
  slots), 006 (phase transition — should sweep CROSS-LEXICAL pressure, updated design
  note), 096 (toy theory now has two shortcut-satisfaction instances to formalize).