Stacked on top of #3 (v0.1 fetch). Merge #2 → #3 → this in order.

Summary

• First standings file lands: leaderboard/next-event/synthetic-toy.json, with the Markov reference baseline as the inaugural entry (top-1 0.9756, top-3 1.0, n=41).
• pm-bench leaderboard <task> <dataset> [--verify] pretty-prints the table and, with --verify, re-runs scoring on the checked-in predictions to catch drift.
• Reference predictions are in-repo (leaderboard/predictions/next-event/synthetic-toy/markov-ref.csv.gz) so the loop is reproducible offline.

What's new

• pm_bench/leaderboard.py — load_board, rescore, verify, standings. Pure CPython, reads gzipped or plain CSV. Truth dispatch is keyed on dataset name; today only synthetic-toy is wired (the dispatch grows a branch per pinned dataset).
• CLI: pm-bench leaderboard <task> <dataset> prints standings; --verify fails non-zero if recorded scores don't match a fresh rescore.
• leaderboard/README.md — submission convention; how to verify locally.
• tests/test_leaderboard.py — 8 tests, including a drift canary that tampers with top1 in a tmp copy of the JSON and asserts verify flags it.

Why this matters

• Locks the leaderboard JSON schema before any external submission lands.
• Makes the Markov number the explicit floor on the leaderboard, not just a number in a README.
• Sets up v0.4's CI workflow as a one-step follow-up: it just runs pm-bench leaderboard --verify on the changed files.

Smoke

$ pm-bench leaderboard next-event synthetic-toy --verify
verified 1 entr(ies) — no drift
next-event · synthetic-toy · top1 / top3 accuracy
----------------------------------------
markov-ref  top1=0.9756  top3=1.0000  n=41

Test plan

• pytest -q — 45 passed (was 37 on PR v0.1: dataset fetch + sha256 verification + cache resolution #3)
• ruff check pm_bench tests — clean
• Drift canary: tampering with the recorded top1 produces a top1 drift message via verify()

Roadmap impact

• README v0.4 milestone marked 🟡 (scaffold + verify shipped; CI workflow on PRs is the remaining piece).