# 014 swap-patching — RESULT (harvested tick 14, 2026-08-02, tier T3-exploratory)

**Verdict: role identity is causally LOCALIZED at filler token positions through L22,
then falls off a cliff — filler-span patches flip the decoded roles at ~0.95 from L0 to
L22 and at 0.00 at L24, identically in actives and passives. The commitment to the
pooled summary happens in the last two layers, not gradually.**

## Numbers (43 kept pairs/family after piece-match exclusions; 95% CIs ~[0.88, 1.0])
- **Filler-span patch flip-rate**: 0.95 at L0-L22 (active AND passive; dip 0.88 at L22
  passive), **0.00 at L24** (decodes revert to the ORIGINAL sentence's roles, garble
  only 0.05-0.07 — the patch is simply ignored).
- **Half-swap control (one filler only): 100% garbled at L0-L22** (parse fails/entity
  dropped — you cannot change one entity independently; causal echo of C2
  entity-entanglement), collapsing to 0.47 garble at L24.
- Random-span control: flip 0.00 everywhere, garble ≤0.05. Parser gates pass; baseline
  decode sanity 0.95-0.98.

## Interpretation (scoped)
Through 22 of 24 layers, WHO-did-what in the decode is carried by the filler tokens'
own states — swap them and the reconstruction swaps; swap one and the sentence breaks
as a unit. By L24 individual filler states no longer matter: the sentence-level
commitment (what pooling will read) has already been broadcast/settled, so a late local
patch can't reach the decode. This completes the mechanistic arc 011→014: role is
(i) absent from the state code as an abstract variable (011), (ii) present in attention
patterns that are causally inert (012+013), (iii) carried lexically-and-locally per
token until a late, abrupt distribution step (014). The "binding" SONAR exhibits at
decode time is the surface form riding on locally-stored lexical identities — never an
abstract role variable at any stage. (Caveats: n=43/family after strict piece-matching
— exclusions 130/200 logged; L24-vs-L24n patch surface distinction; greedy decode only.)

## Brier: (a) 0.92 → T (0.0064); (b) 0.45 → F (0.2025 — profile is flat-then-cliff, not
declining); (c) 0.40 → T (0.36); (d) 0.55 → F (0.2025 — profiles near-identical).
Mean **0.193**.

## Follow-ups: fine-grained L22-L24 sweep + patch-at-L24n(post-LN) to pin the commitment
step; larger n via relaxed piece-matching with alignment interpolation. Strong candidate
for the mechanistic arc writeup (011-014 as one figure).