[ejsmith]

Summary

Sucker is luck-heavy, so the Elo board mostly ranked hot streaks and activity. A newcomer who went 5–0 on coin flips could land above a player with a real 60% edge. Leaders could also stop playing and keep their spot forever. This switches ratings to Glicko-1 and ranks players by a cautious score that reflects how sure we are of their rating.

• Shown rating = rating − 2 × deviation. Players see one number everywhere: the leaderboard, the profile, their world rank, and each game's rating_before/rating_after. New players start at 900 and climb as games build confidence.
• Starting deviation 150, not the textbook 350. Dice luck keeps skill gaps small. A 350 start overreacts to early luck: replaying production history with it, 10 of 332 losses raised the loser's shown rating, by up to 68 points. With 150, none did.
• Idle decay. Deviation grows while a player sits out (variance +10² per day, capped at 150). A settled player loses about 30–45 shown points after a month idle and 75–105 after three. Their first games back count for more, so they climb back quickly.
• Lopsided-pair guard (new). Once either player has won 19 of a pair's last 20 games, games between them stop counting until that record drops below 19. The most lopsided real stretch so far is 17 of 20.
• Kept: 7-day repeat-opponent damping (now a ¼ game weight), forfeits as results, win/loss/tie only, and the 5-game minimum for the rating board.

Changes

• supabase/migrations/20261028120000_switch_ratings_to_glicko.sql
  • Adds player_standings.rating_deviation and rated_at. rating now holds the Glicko estimate. peak_rating tracks the highest shown rating and defaults to 900.
  • Moves the rating math into rate_completed_game. apply_completed_game_standings calls it, then updates streaks and milestone feats exactly as before.
  • Adds the rating_deviation_at and conservative_rating helpers. ranked_leaderboard('rating') and get_player_card compute the shown rating as of now(). The card takes its rank straight from the board, so they can't disagree.
  • Resets every rating and replays all completed games in completion order. Streaks and feats are not touched.
• supabase/schema.sql, shared/database.types.ts, and docs/community-leaderboards.md are updated to match.
• Integration and e2e expectations are updated for the new numbers.

The RPC shapes don't change, so the app and the game-action Edge Function need no updates. Rollout: apply the migration.

Known limitation

No rating formula can tell a thrown game from a real one. The guard stops a long farm against a second account. But with a 150 start, the 20 thrown games the guard still allows would put a fresh account's farmer at a shown ~1227, just above today's #1. Someone who throws only some games can also climb more slowly. A wider start (350) blunts this, at the cost of the loss-raises-rating problem above. Hiding the account from boards remains the backstop.

Testing

• npm run typecheck, npm run typecheck:edge, npm run lint -- --max-warnings=0, npm run test (217 passed), npm run test:edge (44 passed), and deno check on the integration test.
• Applied the migration to a local database inside a rolled-back transaction:
  • The replay of 90 local games matched an independent Python implementation on every rating, deviation, peak, and before/after value (0 mismatches).
  • Streaks and feats were unchanged.
  • The board, card rank, and rankings agreed.
• Ran synthetic games through the completion trigger:
  • The first 20 straight wins between a pair counted and games 21–22 did not.
  • A first-game tie shows 900 → 920.48, a win 900 → 971.24, and a loss 900 → 869.72.
  • Re-applying a rated game is a no-op, and idle decay lowers the shown rating.
• Replaying production history offline: the top two are unchanged and Update build tooling versions #3–Speed up CI caching #7 reshuffle. Changing the decay rate or the starting deviation only swaps neighbors.
• Not run locally: npm run test:integration:supabase, because it resets the shared local Supabase database. CI runs it.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-28T06:22:02.771934Z	44610bd	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.