ranking-elo-update

Goal:

• Apply deterministic Elo updates for one completed tournament batch, persist the touched hypothesis artifacts, and write the ranking update receipt.

Inputs:

• completed tournament matches
• canonical matchups
• strategy
• run_dir
• optional k_factor
• optional top_k_limit

Outputs:

• updated in-memory hypothesis ratings and match references
• updated hypotheses/<id>/HYPOTHESIS.json artifacts for all touched hypotheses
• state/ranking_update_receipts/<receipt_id>.json
• RankingUpdateReceiptContract

Context Loading:

• Open skills/shared-references/schema-index.md.
• Read packages/agent_contracts/ranking.py and confirm the exact TournamentMatchContract plus HypothesisMatchupContract shapes.
• Read packages/agent_contracts/hypothesis.py and confirm the exact HypothesisContract shape before writing any touched HYPOTHESIS.json artifact.
• Use only tournament artifacts whose winners have already been decided and serialized.

Execution Contract:

• This skill is deterministic and must not call an LLM.
• Use from tools import apply_and_persist_elo_updates as the stable invocation surface.
• The exported closeout helper is implemented in packages/run_artifacts/ranking_writeback.py.
• The helper signature is apply_and_persist_elo_updates(run_dir, matches, matchups, strategy, k_factor=..., top_k_limit=...) -> RankingUpdateReceiptContract.
• The closeout helper internally calls the canonical Elo helper, persists touched hypotheses, and writes the ranking update receipt.
• The closeout helper is idempotent for the same stable receipt ID: if the receipt already exists, it returns the persisted receipt instead of reapplying Elo deltas or appending duplicate match refs.
• Accepted strategy values are placement_tournament and ranked_tournament only.
• The closeout helper updates canonical hypothesis bundle fields for touched hypotheses, but it must not rewrite standalone REVIEW/*.json stage artifacts.
• Match refs are not a ranking update receipt. Completion requires the persisted RankingUpdateReceiptContract.

Execution Steps:

1. Open skills/shared-references/schema-index.md, then read packages/agent_contracts/ranking.py and packages/agent_contracts/hypothesis.py before writing any touched HYPOTHESIS.json.
2. Load the completed tournament matches as TournamentMatchContract payloads and the paired matchup payloads as HypothesisMatchupContract.
3. Confirm every match already has a decided winner_id and that the requested strategy is valid.
4. Call tools.apply_and_persist_elo_updates(run_dir, matches, matchups, strategy, k_factor=..., top_k_limit=...).
5. Confirm the returned RankingUpdateReceiptContract names every completed match ID and touched hypothesis ID.
6. Validate the touched hypothesis artifacts and state/ranking_update_receipts/<receipt_id>.json before declaring completion.

Artifact Rules:

• This skill updates ratings only after the tournament winner artifacts are already finalized.
• Placement rounds update placement_match_ids for the challenger hypothesis and ranked_match_ids for the defender.
• Ranked rounds update ranked_match_ids for both participating hypotheses.
• Match ref arrays must remain duplicate-free. If a completed match ID is already recorded by both touched hypotheses for the same strategy, do not apply the Elo delta again.
• Do not invent extra ranking heuristics outside the helper; the helper is the canonical Elo update rule.
• Do not replace canonical ranking closeout with ad hoc file edits; use tools.apply_and_persist_elo_updates(...) after every Elo batch.
• Review-stage synchronization is a separate concern owned by tools.sync_hypothesis_review(...); Elo writeback must not clobber those standalone review artifacts.
• Do not hand-write ranking update receipts. The closeout helper owns receipt creation.

Completion Rule:

• This skill is complete only when the closeout helper has been applied once to the finished batch, every touched hypotheses/<id>/HYPOTHESIS.json artifact has been rewritten in canonical contract form, and the corresponding ranking update receipt exists.