# 016 nonlinear-probes-l24n — RESULT (harvested tick 16, 2026-08-02, tier T3-exploratory)

**Verdict: NEGATIVE across all three remaining probe families (kernel-RFF,
focal-conditioned MLP, bilinear-relational) — no transferable role structure at any
certified surface. The nonlinear families instead recover the SURFACE code at ceiling
(same-parity 1.000, flipped 0.06-0.14), the sharpest surface-reader signature yet.**

## Numbers (primary = cross-construction + lexical holdout, 20 pairs, prop-cluster CIs)
| cell | cert. sensitivity (plant amp) | primary | flipped | same-parity |
|---|---|---|---|---|
| kernel L8/L16/L22 | 8.0 / 8.0 / 8.0 | 0.452 / 0.437 / 0.471 | 0.06-0.12 | **0.999-1.000** |
| mlp L16/L22 | 4.0 / 8.0 | 0.451 / 0.484 | 0.09 / 0.14 | 1.000 |
| mlp L24n | UNPOWERED | (0.415 — not interpreted) | — | — |
| bilinear L16/L22 | 2.0 / 2.0 | 0.472 / 0.410 | 0.32-0.42 | 0.55 |
| kernel/bilinear L24n | UNPOWERED | not interpreted | — | — |

- **L24n never certified in any family** (even the LN-consistent rotation plant) —
  reported UNPOWERED, not null; the post-LN surface remains probe-hostile (a finding
  about LN geometry in itself, consistent with 011/015's marginal certs).
- Mid-stack nulls are licensed only at the documented sensitivity (2-8× plant amplitude
  — i.e., these families would miss role codes smaller than the certified plant; stated
  per prereg amendment).

## Interpretation (scoped)
Five probe families (linear, trained attention-pooler, kernel, focal-MLP, bilinear) now
agree: SONAR token states contain no cross-construction-transferable role code down to
the certified sensitivities, while every sufficiently powerful family finds a PERFECTLY
transferring surface-order code (same-parity 1.0 / flipped ~0.1 = the pure anti-transfer
signature). Readout exhaustion is about as complete as probing can make it; the 011-016
arc's conclusion stands on token-state content, not readout choice.

## Brier: (a) 0.32 → F (0.1024); (b) 0.75 → F under the honest strict reading (L24n
never certified; pre-flagged leniency not used) (0.5625); (c) 0.68 → T (within 1.0)
(0.1024). Mean **0.256**.

## Incidents: initial chain aborted on an over-strict all-families gate (prereg said
report-unpowered-and-continue); salvaged same-day via bounded power tuning (kernel/MLP
certified at higher amplitudes) + bilinear full run. Manager repeatedly idle-waited on
watchers against instruction (~150k tokens) and was stopped; tick harvested. PROTOCOL
lesson already logged at t14 now enforced: no watcher-managers.

## Follow-ups: none needed for the arc; the LN-geometry probe-hostility of L24n could
seed a small methods note (feeds 092 norm-linter and 097 theory).