ARM CC (code-channel probe) — frozen results. 2026-08-18.
Registration: arm_cc_preregistration.txt, commit 285deca (pushed to public
remote before first invocation). Collection: 45/45 launched invocations
returned, no runtime deaths (arm_cc_collect.jsonl, verbatim). Scoring:
arm_cc_analysis.py (committed with registration; smoke test 4/4 bins).

REGISTERED VERDICT: P-CC1 HOLDS (3 of 3 cells) — the modal valid executed
output remains the T(k*,N) anchor bucket at every cell.
Margin disclosure (registered convention): N=21 mode is heavy (9/13 valid,
margin 8 — nine exact 2.1000000 emissions). N=13 and N=31 modes are the
anchor value at ONE-SAMPLE margins (2/12 and 3/12) over dispersed sub-anchor
values; the modal-identity claim is thin at those two cells.

Cells (valid / sampled, anchor-rate, argmax-rate, above-rival):
  N=13: 12/15, anchor 2/12, argmax 1/12, above-rival 0/12
        sums max = 1.7761424 (the family argmax V(3,4), EXACT)
  N=21: 13/15, anchor 9/13, argmax 0/13, above-rival 0/13, max = 2.1
  N=31: 12/15, anchor 3/12, argmax 0/12, above-rival 0/12, max = 2.5833333

MANDATORY DECOMPOSITION ROW (registered): no valid program output exceeds
the family argmax anywhere (0/37 above rival). Exactly one output exceeds
the anchor at all: a single N=13 program computing the family argmax
construction exactly. Sharpens arm CH: chosen argmax cannot be built by
direct emission (30/31 invalid, S3.5); delegated to an executed program it
is reached once in 37 and never passed.

Failure taxonomy: valid 37, geom_invalid 7, blocked_import 1 (random,
against the stated math-only contract), timeout 0, other bins 0.

Structural k = k* in 18/37 valid outputs (3/12, 10/13, 5/12). Versus the
code-free bare arm (anchor concentration 56-86% by cell): pooled
code-channel anchor-rate 14/37; the lost mass sits entirely BELOW the
anchor (greedy radius-growth heuristics) — dispersion, not escape.

Registered framing branch (P-CC1): anchoring extends into the math-only
code channel at these cells; S2.1's "a code channel routes around the
anchor" is weakened from an assumption to a measured negative for this
proposer. Scope: stdlib-math programs only; the library-assisted channel
(numpy/scipy) is registered as a further arm, not measured here.

Corpus: 468 + 45 = 513 invocations.
