# RESULT — 011 v4-ln-delta (binding-death v4, LN-consistent contextual delta at pooling input)

*Ran 2026-08-01, CPU only (tmux c100_011, box, 339s wall). Tier: T3-exploratory.
**Promotion candidate: YES** — this is the first properly-certified instrument to answer the
OPEN H-POOL question (v2/v3 were both INSTRUMENT_FAILURE). See PREREG_LITE.md (+ its pre-run
amendment) for the frozen gates/predictions. Raw numbers in `out/results_v4.json`,
`out/v4_norm_profile.json`.*

## Headline
At the LN-consistent pooling input (post-LN L24n), with the static lexical component removed
at the correct scale, **there is NO linearly-decodable, construction-invariant role code** —
the probe reads *surface order* at ceiling and *anti-transfers* on role, the exact signature
the flipped-parity guard exists to catch. This is the cleanest evidence to date that C1's
pooled-z null is **stack-deep, not a pooling / LayerNorm / scale artifact** (the v3 failure mode).

## Two constructions, two outcomes
### delta-LN = L24n − LN(L0)  →  **CLEAN_NULL** (certified)
- **All instrument gates pass**: G-lint (β-free term ratio **1.05x**; see norm note below),
  G-repro (**0.5084** [0.489,0.527], target 0.5084 ✓), G-power (certified — power curve reaches
  AUC>0.9 by **10%**-norm; 5% point 0.82), G-order (**1.0** — order readable at ceiling, so the
  construction is NOT over-aggressive).
- **Role primary dif**: AUC **0.4597** [0.453,0.466]; same-parity **0.988**, flipped-parity
  **0.108**, median flipped-pair **0.013**; permutation p=**1.0**. Broadcast controls all ≈0.5
  (eos 0.508, lang 0.506, rand 0.485, pooled 0.495) — no global/broadcast role code either.
- **Genitive positive control** (surface-reader signature): within-family dif **1.0**,
  cross-family **0.62**, cross-family-flipped **0.50** — probe generalizes within a family
  (memorizes order) but not across parity ⇒ it is an order reader, not a role reader.
- **Reading**: same-parity 0.99 + flipped 0.11 + perm p=1 ⇒ the certified probe reads WHICH
  FILLER IS SURFACE-FIRST perfectly and has its sign flipped when parity flips. Role is not
  linearly present at the pooling input, in a scale-consistent construction that provably
  reproduces C1 and can detect a planted 10%-norm signal.

### projection = h − (h·ê)ê, ê=unit(LN(L0))  →  **INSTRUMENT_FAILURE** (low power)
- **G-power FAILS**: the probe needs a **40%**-norm planted signal to reach AUC 0.9 (5% → 0.46,
  20% → 0.54). Cause: ê=unit(LN(L0)) removes only ~**6%** of the L24n norm (L24n is nearly
  orthogonal to the static direction), so proj ≈ raw L24n and inherits raw-state low power.
- Its dif (0.44, same-parity 1.0, flipped 0.067) shows the same order-reader shape, but the null
  is **not trusted** — the cell cannot distinguish absence-of-role from insufficient-power.
- **Lesson**: projecting the static direction out of L24n does not sharpen the cell; only the
  β-free normalized delta (delta-LN) reduces lexical variance enough to become a usable probe.

## Norm-lint note (documented, not hidden)
delta-LN raw subtracted terms are **0.47** (L24n) vs **3.60** (LN(L0)) = 7.7x, which would trip
a naive >5x lint. The entire mismatch is the shared additive LN bias β (‖β‖=2.55) which **cancels
exactly** in the subtraction: delta-LN = γ⊙(normalize(L24) − normalize(L0)); the governing lint on
the β-free terms is **2.48 vs 2.57 = 1.05x → PASS**. Contrast v3's 800x, *non-cancelling* mismatch
that injected −(32×embedding+sinusoid) into the delta. LN validated: LN(L24_raw) == stored L24n to
max rel err **2.9e-6**.

## Brier (honest ledger) — mean **0.267**
| # | prediction | p | outcome | note |
|---|---|---|---|---|
| P1 | probe-power certifies for BOTH variants | 0.92 | **F** | over-predicted; proj failed certification (needs 40% norm). Registered 5%-bar also F for both (0.82 / 0.46). Main miss. |
| P2 | role signal found at L24n (delta-LN) | 0.15 | F (correct null) | prior held: no role code |
| P3 | order still readable from delta-LN (≥0.9) | 0.60 | **T** | under-confident; order = 1.0 |
| P4 | projection variant null (dif CI-hi<0.65) | 0.80 | T | proj dif 0.44 (but null uncertified) |

The P1 miss is the substantive calibration update: I assumed a scale-consistent construction would
be a powerful probe; the *projection* construction is not (it barely alters L24n). Only delta-LN,
which subtracts the normalized static, becomes a certified instrument.

## What did NOT run / scope
- Single SONAR checkpoint, English, 5 agent_patient families + genitive, frozen lexical split.
  Single surface (L24n) by design — no layer sweep, no Bonferroni (stated up front).
- Linear probes only; oracle filler-position readouts measure information *presence*, not
  decoder-usable info. Nonlinear probes at L24n are the natural next step (manifest 016).
- delta-LN is one LN-consistent construction; γ⊙(n24−n0) is a specific static-removal, not the
  unique one.

## Verdict & follow-up
**delta-LN: CLEAN_NULL — a certified negative for linear role binding at the pooling input.**
Combined with the mid-stack negative (BD_RESULTS §1c) this closes the "pooling/LN destroyed a
role code that was there" story for the linear regime: the code is not linearly present at the
pooling input either, in a construction that reproduces C1 and passes a planted-signal power test.
**Fund follow-up? Y** — (1) promote this delta-LN cell to the H-POOL record (it is the gate v3
could never pass); (2) nonlinear/MLP probe at L24n (016) on the delta-LN readout with the same
power certification; (3) the more valuable direction remains objective-side (001/002): whether any
training objective can *induce* a construction-invariant role code, since there may be nothing to
localize.