菜单
Run and interpret a bidding A/B experiment in pons — harness choice, seed hygiene, polite scheduling, dual scoring, and the ship/reject decision. Use whenever the user asks to measure, A/B, benchmark, re-measure, or validate a bidding change, or to interpret A/B results.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
| name | measure-ab |
| description | Run and interpret a bidding A/B experiment in pons — harness choice, seed hygiene, polite scheduling, dual scoring, and the ship/reject decision. Use whenever the user asks to measure, A/B, benchmark, re-measure, or validate a bidding change, or to interpret A/B results. |
docs/measurement.md is the source of truth — read it now if this session hasn't. This skill is the run procedure.
examples/ab-* (self-play
seat-swap; check --help — most questions are a flag away). Use
examples/bba-gen via scripts/bba-gen-parallel.sh + ab-dump-diff when
you need the reference opponent — mandatory for tuning our own conventions'
strength/ranges (self-play can't punish them; pass the matching
--advertise-* so BBA reads the bids).export SEED_BASE=$(date +%s) once; all arms of this experiment
share it (paired diffs need identical deals). Next experiment → new base.
Record the base + git SHA in the results note.scripts/idle-run.sh (shared box), arms
strictly sequential; one run already saturates every core. tmux or
setsid nohup for long runs. Typical scale: 200k filtered boards/cell
contested, ~205k/arm (6400 × nproc) for bba-gen; 40k filtered ≈ 1 min/cell.single_dummy_leads prices the blind-lead seam).Advance or resume the "AI instinct bidder" effort in pons — replacing the deterministic bidding floor with a learned model (a hand→call policy net and a description language model). Use when the user asks to work on the neural/AI bidder, the policy net, distillation, the constrained sampler, the description compiler/corpus, or references docs/ai-bidder/. Loads the design, respects the invariants, and teaches the ML grounded in math/bridge/low-level.
End-to-end checklist for adding or changing a bidding convention/treatment in pons — authoring the rules, alerts, inference reading, knob, tests, and the A/B measurement that decides whether it ships. Use whenever the user asks to add, author, implement, tune, or fix a convention, treatment, response structure, defense, or any behavior of the bidding system.