# RESULT — 012 head-role-circuits

*Ran 2026-08-01 (tmux c100_012, box GPU2/CVD=1 capture + CPU analyze; whole chain ~15 s).
Tier: **T3-exploratory**. **Promotion candidate: YES** — first attention-level positive in
the program: SONAR attention IS role-structured even though the linear state/pooled code is
not (binding_death v4 / 011). Frozen gates & predictions in PREREG_LITE.md. Raw:
`out/results.json`, `out/perm_control.json`, `out/attn_meta_agent_patient.json`; per-head
span-stats npz retained on box (~1.5 MB).*

## Headline — POSITIVE (surprising)
Of the 384 SONAR text-encoder self-attention heads, **42 are role-head candidates**: their
agent→predicate vs patient→predicate attention has the SAME sign in ≥4/5 constructions,
Bonferroni-significant (z* = 3.83, 384-way) in ≥4 of them with that sign, AND role-consistency
> surface-consistency. The top head **L11h5** attends agent→predicate far more than
patient→predicate in **all 5** families — active +0.325, passive +0.277, cleft +0.079,
objrel +0.427, nominal +0.161 (agent→pred ≈ 0.11–0.44 vs patient→pred ≈ 0.01–0.06), per-family
z = 29/27/20/32/21. Both **agent-preferring** (L11h5, L10h12, +) and **patient-preferring**
(L7h15, L11h6, −) role heads exist. Held-out split: the top head's family-sign pattern agrees
**5/5** on the held-out half of propositions (survives).

Crucially the pattern is NOT grammatical-subject tracking: in the passive the subject IS the
patient, yet these heads still route the AGENT (in the oblique by-phrase) to the predicate.

## Gates
- **G-align: PASS** — span alignment **1.000** overall and in every family (active/passive/
  cleft/objrel/nominal all 1.000). Predicate spans taken from the generator's exact VERBS table
  (past form; nominalization head for the nominal family), so no fuzzy matching.
- **G-power (positive control): PASS, strongly** — surface-consistency finds **216** Bonferroni-
  significant + surface-consistent heads (304/384 pooled-surface-significant). Position/order
  attention is ubiquitous, as expected; the instrument is not dead.
- **Label-permutation null (decisive control, added post-hoc):** randomly swapping the
  agent/patient label per item → **0 role-candidates across 5 seeds** (vs 42 with true labels).
  The 42 are genuinely tied to thematic role labels, not a metric/bootstrap artifact. (Sign-only
  role-consistency still gives 14–31 heads at 5/5 under permutation — chance/surface sign
  agreement — but the significance × role>surface conjunction collapses to 0, which is exactly
  the discriminating power the candidate definition was designed for.)

## Discriminant that makes this real
The modal role-consistency across 384 heads is **3/5** (214 heads) — precisely the value a pure
surface/position head must take (agent is surface-first in active/cleft/nominal, surface-second
in the parity-flipped passive/objrel). 129 heads sit at 4/5 and **41 at 5/5**. A surface head
cannot exceed 3/5 role-consistency by construction, so the 4–5/5 heads carry genuine thematic
structure beyond surface order. All 42 candidates have surf_cons = 3/5 (surface-inconsistent).

## Interpretation (and what this does NOT mean)
- **Complementary to binding_death / 011, not contradictory.** Those certified there is no
  *linear, construction-invariant role code* in the token states or at the pooling input. 012
  is about the **attention computation**, a different channel. Role-consistent routing in
  attention is fully compatible with the pooled z having no linearly-decodable role axis — the
  routing can be washed out / non-linearly encoded by mean-pooling, or used for something other
  than writing a role variable into the residual stream.
- **Mechanism is likely cue-exploiting, not amodal.** A single head realizes agent→pred across
  constructions by using each construction's overt role-marking morphosyntax (subject position,
  the by-phrase, "whom", the possessive 's). That is legitimately *role-consistent attention*
  as prereg'd, but it is not evidence of an abstract, cue-free role variable.
- **Pattern, not causal usage.** This is an attention-structure claim. Whether these heads
  causally determine decoded role fidelity is untested here (→ 013 head-knockout, 014 swap-patching).

## Brier (honest ledger) — mean **0.3798**
| # | prediction | p | outcome | sq |
|---|---|---|---|---|
| a | ≥1 Bonferroni role-consistent head, role>surface | 0.12 | **TRUE (42)** | 0.7744 |
| b | ≥20 surface-consistent heads | 0.93 | TRUE (216) | 0.0049 |
| c | top role head survives held-out split | 0.40 | TRUE (5/5 agree) | 0.3600 |

The (a) miss is the substantive calibration update: I carried the program's strong "no role
code" prior into the attention channel and set p=0.12; the attention channel is in fact richly
role-structured. Lesson: a certified null on the *linear state code* does not license a low
prior on *attention structure* — they are different measurement surfaces.

## What did NOT run / limitations
- n=1 SONAR checkpoint, English, agent_patient only. Forward (noun→pred) direction is primary;
  reverse (pred→noun) stats captured and in `results.json` but not the headline.
- "Role" here is agent-vs-patient thematic contrast operationalized as span→span mean attention;
  it does not separate a unified amodal agent-detector from a bag of construction-specific cue
  heads (see mechanism caveat). Attention structure ≠ causal contribution to the pooled z.
- Bootstrap SE via 500 cluster resamples of 200 propositions; Bonferroni (max-statistic) over
  384 heads; conjunction across families is an intersection test (no extra correction).

## Verdict & follow-up
**POSITIVE — role-consistent attention heads exist in SONAR (42/384, certified vs a strong
surface positive control and a label-permutation null).** Fund follow-up? **Y**:
1. **013 head-knockout** and **014 swap-patching** are now sharply motivated — ablate/patch the
   L11h5-type heads and read decoder role fidelity: do these heads *cause* role reconstruction,
   or is their routing inert w.r.t. the pooled z (which would reconcile with the linear null)?
2. Characterize the mechanism: is one head an amodal agent-detector, or does each lean on a
   single cue (by-phrase, subject-adjacency, 's)? Cue-lesioned stimuli would separate these.
3. This is the H-POOL story's missing half: role lives in the *computation/attention*, not in
   the *linear geometry* of the states.