# 013 head-knockout — RESULT (harvested tick 13, 2026-08-01, tier T3-exploratory)

**Verdict: the 42 role-consistent heads are CAUSALLY INERT for reconstruction — role-
structured attention (012) is pattern without function at the decoder's readout. Bonus
first-measurement: SONAR's role-preservation floor on canonical transitives = 0.981.**

## Numbers (2000 items, parser gate 1.000 on gold inputs; prop-cluster bootstrap CIs)
- Baseline: role preserved **0.981** (swap 0.0045, drop 0.0145; chrF-vs-input 94.6;
  per-family 0.970-0.988 incl. passive 0.983).
- Ablations (delta preserved vs baseline): L11h5 alone **+0.000** [-0.002,+0.002];
  top-10 role **-0.001** [-0.004,+0.003]; **all-42 role +0.001** [-0.006,+0.007] (fair
  control random-42: -0.001/+0.003). Every role condition ≈ every control.
- Only CI excluding zero: **surface-top-10 ablation, -0.0055** [-0.010,-0.001]
  degradation (and random10_s3 -0.004) — small but sign-consistent with roles riding
  the reproduced SURFACE form, not a role channel.
- cos(z_ablated, z_clean) ladder confirms ablations bite: 0.9998 (1 head) → 0.987 (42).

## Interpretation (scoped)
Combined 012+013: SONAR's encoder computes role-structured attention, but that structure
is not written into anything the decoder's role fidelity depends on — ablate it all and
reconstruction roles are untouched. Meanwhile the decoder preserves roles at 98% on
canonical transitives WITHOUT any state-level role code (011): it reproduces the surface
form nearly verbatim (chrF 94.6), and role rides on the reproduced word order/construction.
The 0.981 floor also usefully bounds the fabrication concern for short in-distribution
sentences (C4's ~20-30% figure is free-form text; here swap errors are 0.45%).
Epiphenomenal-attention is the parsimonious reading; alternatives (redundancy across >42
heads; mean-ablation too gentle) noted as unexcluded.

## Brier: (a) 0.50 → T (0.25); (b) 0.35 → F (0.1225); (c) 0.12 → F (0.0144);
(d) 0.45 → T (0.3025). Mean **0.172**.

## Follow-ups: 014 (swap-patching) is now the sharper causal tool — patch attention/states
between swap-mates and see WHERE the decode flips; redundancy check (ablate 42 + top
surface simultaneously) is a cheap add-on if the thread is promoted.