Found 2026-09-09 while running the rung-163 bench rows through the policy path on rift_O4d 56bb68c (#301 + #305). Parked deliberately behind the multipeak ladder result (RO, 2026-09-09): this is a gate defect in the policy path only and does not touch the multipeak controller.
Symptom. With the PolicyConfig defaults #301 introduced, the LOCAL branch accepts bench row 1 of /scratch/richard.oshaughnessy/policy_batch_bench/likedata_snr160.pkl (seed 7 recipe) at lnL 12351.971, while the whole-window exact reserve at refine 32 (19617 nodes, check at 16 agreeing to 6e-10) and the peak-local reserve both give 12335.9. Accepted value +16 nat high. Row 2: +1.9 nat. Rows 0 and 3 accept at identical values under both capacity settings. Under the pre-#301 defaults (64 / 32 / 1e-3) rows 1 and 2 declined to the reserve, so the raised capacities let the acceptance gate pass values it should refuse.
Reproduction, complete: /scratch/richard.oshaughnessy/policy_batch_bench/DEFECT_rung160_local_accept_20260909.md (CIT only): tree, pickle, row recipe, config, script /scratch/richard.oshaughnessy/jaxbl_diag/measure_plr.py, ledger subsets for both configurations, and like._batched_ledger(ra, dec, incl) for the full per-row ledger.
To do, when picked up: which of the four acceptance terms (all_axis_peaklocal.py ~862) passed on rows 1 and 2; the two-tier quadrature delta; candidate count vs max_starts and time-node count vs max_time_nodes; a single-variable test that flips the outcome; then a recommendation on whether the #301 defaults need reverting or the gate needs a term (RO decides). Add a regression test on the +16 case. Until then, no policy-path rung-160 acceptance is quotable.
Found 2026-09-09 while running the rung-163 bench rows through the policy path on rift_O4d 56bb68c (#301 + #305). Parked deliberately behind the multipeak ladder result (RO, 2026-09-09): this is a gate defect in the policy path only and does not touch the multipeak controller.
Symptom. With the PolicyConfig defaults #301 introduced, the LOCAL branch accepts bench row 1 of
/scratch/richard.oshaughnessy/policy_batch_bench/likedata_snr160.pkl(seed 7 recipe) at lnL 12351.971, while the whole-window exact reserve at refine 32 (19617 nodes, check at 16 agreeing to 6e-10) and the peak-local reserve both give 12335.9. Accepted value +16 nat high. Row 2: +1.9 nat. Rows 0 and 3 accept at identical values under both capacity settings. Under the pre-#301 defaults (64 / 32 / 1e-3) rows 1 and 2 declined to the reserve, so the raised capacities let the acceptance gate pass values it should refuse.Reproduction, complete:
/scratch/richard.oshaughnessy/policy_batch_bench/DEFECT_rung160_local_accept_20260909.md(CIT only): tree, pickle, row recipe, config, script/scratch/richard.oshaughnessy/jaxbl_diag/measure_plr.py, ledger subsets for both configurations, andlike._batched_ledger(ra, dec, incl)for the full per-row ledger.To do, when picked up: which of the four acceptance terms (all_axis_peaklocal.py ~862) passed on rows 1 and 2; the two-tier quadrature delta; candidate count vs max_starts and time-node count vs max_time_nodes; a single-variable test that flips the outcome; then a recommendation on whether the #301 defaults need reverting or the gate needs a term (RO decides). Add a regression test on the +16 case. Until then, no policy-path rung-160 acceptance is quotable.