# n214 — Dressed-speed gate, leg 1: executed, INCONCLUSIVE by its own clause — with a robust adverse trend and an implementation discrepancy to resolve

**Date:** 2026-07-11 · **Gate:** `gates/dressed-speed.json`, sha256 `4a44340f…d8c65`, frozen before execution (motivation: the guard term is Newton's own back-action — nothing introduced for the speed; PASS band [0.98, 1.02]; FAIL wording frozen; INCONCLUSIVE = non-convergence class).

## What ran

The declared computation: two-phonon shear sums in the dressed vacuum, dressing ω²(k) → ω²(k) + s_g·w_A(k) (the monitor shifts what it watches; the blind channel gets nothing), light cone = the staggered sector's dressed speed, s_g swept 0 → 16 within the stability domain.

## The two findings

**1. The trend is robust and adverse.** Grid-checked at 24³/36³/48³ (0.3% at the last doubling for s_g = 0; 4.2% for s_g = 2, inside the 5% clause): **the guard-type dressing monotonically LOWERS the shear speed** — 1.437 → 0.928 → 0.284 raw across s_g = 0 → 2 → 8 at 48³ — moving *away* from the cone at every step. The curve never approaches the PASS band from below with this dressing class. If this survives reconciliation, leg 1 fails and the frozen wording fires.

**2. But the gate cannot adjudicate tonight: the implementations disagree.** At the shared undressed point (s_g = 0, same complex, same weights, same formulas line-by-line), the Python execution gives c_g² = 1.44 (grid-stable to 0.3%) where the n202 Rust gave 3.66 (also grid-stable). Same declared computation, **factor ≈ 2.5 apart — one of them has a bug**, and until it is found, *both* n202's speed map and tonight's dressing curve carry an unresolved-implementation flag. This is the audit's §5 lesson operating as designed: the second backend caught what one alone could not. Per the frozen INCONCLUSIVE clause (numerical non-convergence, declared), **leg 1 returns INCONCLUSIVE — no verdict, no fired condition, no rescue.**

## The immediate next step (before anything else in this sector)

Reconcile: a single minimal test point (one complex, one q, one μ, both backends instrumented term-by-term — |M|² sums, R₀, I₀ separately) to localize the divergence. Whichever implementation is wrong gets the defect recorded and repaired fail-closed; then the leg re-runs. The trend's direction (dressing ↓) must be re-confirmed in the corrected implementation before it means anything.

**Class: [A-model execution, honestly stalled]; the adverse trend [flagged, pending reconciliation]; the gate remains frozen and open. Nothing widened; nothing claimed.**