meta-watch

name	meta-watch
description	Weekly PTCG ladder meta check — analyze our latest submission's replays for opponent-archetype W-L and meta share, flag rotations vs the last snapshot, show the LB top. Use to check the meta, our submission's record, or whether the meta rotated. Triggers — "メタ確認", "メタ回転した?", "今の戦績", "weekly meta check", "what are we facing".

Weekly meta-watch

One command for the operational loop: check the meta → if it rotated, counter (/build-submit); else hold.

Steps

1. Run the watcher (auto-detects our latest scored submission; downloads its ladder replays — cached under references/raw/replays/, gitignored):

   cd workspace/exp011_meta_watch
   uv run python meta_watch.py            # or: meta_watch.py <submission_id>
   

   It prints: opponent-archetype W-L + meta share, a rotation diff vs the previous snapshot (flags any archetype whose share moved ≥15%), and the LB top 8.

2. Mind the sample size. If the latest submission has < ~15 games it says so — rotation flags are unreliable on a fresh (μ600, few-game) submission. Re-run later, or pass a more-converged submission_id (see kaggle competitions submissions pokemon-tcg-ai-battle). Our eligible pair = latest 2 submissions by time.

3. Interpret & decide:

   • No rotation (< 15% share change): hold / monitor. Note our W-L per archetype (are our eligible agents still favored vs the field?).
   • Rotation detected — first ask: temporal rotation, or ALTITUDE? (learned 07-14) The opponent mix changes with OUR rating: a submission that climbs 700→940 sees a different band (e.g. crustle 20%→0%, archaludon 9%→45%) with zero temporal rotation. If our rating moved a lot between snapshots, read the "rotation" as band change and compare against the matching band pool (exp053/054), not the previous snapshot.
   • Real rotation: identify the rising archetype and whether our agents beat it. The meta is a rock-paper-scissors (ex-beatdown → anti-ex wall → non-ex attackers →…); see [[meta-and-leaderboard]].

3b. Contingency triggers (for the LO line, v021+; set 07-14). Check these shares at OUR altitude every run; below threshold = hold:

• non-ex aggro (e.g. Team Rocket's Spidops) > 20-25% (raised 07-19 by exp062; was 10-15%, and 11% was REACHED on 07-19 with koff 0/6 live) — LO's structural weakness class: non-ex attackers bypass BOTH Crustle Safeguard and Neutralization Zone (each blocks only ex attacks) and win the prize race. Below the raised threshold the v020 swap is EV-NEGATIVE: v020 fixes the TR lane (floor 0.955) but bleeds the 53% mainstream (archaludon mirror 0.500 / alakazam_dun 0.610, silver weighted 0.634 vs koff 0.786; each +10% Spidops share closes ~0.09 of that gap).
• crustle/LO bucket > 20% (was ~7%) — mirror saturation / anti-LO rising.
• Starmie-type any clear rise (was ~0%) — flagged wall/LO killer (pilkwang). Ready fallback if triggered: v020 Archaludon — REBUILD with the updated cg first (disc727094 engine fix, 07-17; old build bundles the pre-fix engine). Counter-deck pre-building was measured and declined (dragapult 0.567 / wall 0.438 silver-band as solo candidates vs v023's 0.792).

3c. Reroll check (best-of-2 ops; set 07-14, CONSTRAINT found 07-15). LB = max of the 2 eligible submissions' INDEPENDENT ratings (complementarity is impossible), and byte-identical agents settle up to ~400 pts apart (disc712621; our own live demo: identical koff builds at 927 vs 709). Keep TWO copies of the strongest build eligible; when a slot settled clearly below the other (≥100 pts, ≥~30-50 games), reroll it by resubmitting the strongest build. HARD CONSTRAINT: a new submission always evicts the OLDER eligible slot (latest-2-by-time), so you can only ever reroll the older one. If the older slot currently holds the HIGH rating, submitting ANYTHING sacrifices it — do not submit; wait until the newer slot overtakes the older. Repeat until the final-submission target (8/2). Also: the silver cut DRIFTS over time (930.2→919.4 over 07-15→07-17; direction varies with team influx) — aim for margin above the cut, not touching it. Matchmaker mechanics are now OFFICIAL (staff, disc726690 — full notes in references/knowledge/matchmaker_mechanics_0717.md): scheduling priority = sigma + staleness, SCALED by rating (up to ~8x vs mu600) and new submissions heavily prioritized; opponents sampled from a Gaussian window around own rating (band meta confirmed). Consequence: a ticket's draw quality LOCKS IN early (path-dependent loop: early wins at high sigma → 8x more games → fast sigma shrink at the high rating). JUDGE REROLLS AT DAY 1-2 (<~800 = lagging, redraw when the eviction constraint allows) instead of waiting 2-3 days. Also: the FINAL evaluation phase keeps generating episodes (sigma-reduction goal) — lucky-high slots may partially wash out, so for the FINAL pair prefer true build strength over a lucky draw. Medal-cut recheck when top compresses: kaggle competitions leaderboard pokemon-tcg-ai-battle --download → silver = score at rank ceil(0.05·N), bronze at ceil(0.10·N).

4. (Optional) scout the top to see what new top decks look like:

   uv run python top_meta.py <top_player_submission_id> <tag>
   

   Find a top player's submission_id by traversing episodes (their submission_id appears as an opponent in our / another player's episodes; cross-ref names with kaggle competitions leaderboard pokemon-tcg-ai-battle --show). To copy a top deck, use the /extract-deck skill; to build+submit a counter, use /build-submit.

4b. (Optional) scout a top PILOT's decisions (piloting is the #1 lever; same deck, +200 LB). After caching a pilot's replays (step 4 / top_meta.py), find where their play differs from ours:

cd workspace/exp022_megastarmie
uv run python pilot_gap_scan.py <replay_dir> <name_substring> [extra_card_ids...]
# e.g. pilot_gap_scan.py top_mogja_j_0624 Mogja      (our non-ex deck pilot)
#      pilot_gap_scan.py 0625_54022035 Morita        (our own agent, for comparison)

It prints per-decision (length-normalized) action rates split W/L, plus a take-when-legal table. Interpretation rule (learned exp022): compare the top pilot to our own scan.

• Big exposure gap + matching take-when-legal rate = a draw/throughput/game-length gap, NOT a patchable decision — single-card rules won't help (don't build one).
• Normal exposure but we take it far less when legal = a real gated decision leak — the gust template (Boss's Orders, gated by plan.target>=1). Worth a single fix.
• The single-card-leak class is currently exhausted for our non-ex deck (take-when- legal matches Mogja); see [[meta-and-leaderboard]]. Re-scan only after the deck/meta shifts.

5. Report the meta table, rotation verdict, and a recommendation (hold vs build a counter). Don't submit anything here — that's /build-submit, after user approval.

Notes

• Needs Kaggle API auth + the cabt engine (via uv run).
• Snapshots accumulate in workspace/exp011_meta_watch/results/meta_*.json; the diff uses the most recent prior one.