[Agora] Per-field tournament driver with fairness sampling + RD-aware scheduling done

Goal

scidex/agora/open_question_tournament.py exposes run_tournament_batch(domains, pairs_per_field) and run_single_match(), but pair selection is uniform-random within field.
With wiki/debate/paper miners landing thousands of new open_questions
(other Q-OPENQ tasks), uniform sampling will starve high-RD newcomers and
re-grind already-converged top pairs. Build a Glicko-2-aware scheduler that
prioritizes pairs with high uncertainty reduction per match, ensures every
new question gets ≥5 matches in its first 7 days, and avoids re-judging the
same pair within a 30-day window.

Acceptance Criteria

uses three weighted strategies:
1. High-RD warmup (40% of budget): sample one question with the
highest RD in field, pair with a random opponent within ±200 Elo.
2. Information-gain pair (40%): pick pairs where expected score
is closest to 0.5 (most informative under Glicko-2) and combined
RD is highest.
3. Top-of-field consolidation (20%): pair the top-10 by Elo
against each other to settle the leaderboard head. RD has grown ≥10 since the last match. Persist last-match timestamps
per pair in a new table open_question_match_history(question_a_id,
question_b_id, judged_at, winner, judge_persona). created_at > NOW() - INTERVAL '7 days' AND n_matches<5 is
force-included in the next batch until satisfied. --schedule-batch --field neurodegeneration --pairs 50 runs the new
scheduler. gains fields: cold_start_questions_count, next_high_info_pairs[],
field_rd_distribution. 30-day pair lockout (and the RD-grew-by-10 escape hatch); reproducible
with seeded RNG. with (question_a_id, question_b_id) symmetric uniqueness via
sorted-ids invariant.

Approach

Read existing run_tournament_batch and field_leaderboard in

scidex/agora/open_question_tournament.py to understand current loop.

Reuse Glicko-2 RD reads from the existing judge_elo_ratings table

(or whatever the actual table name is — derive from
_get_or_init_rating() in open_question_tournament.py:189).

Implement scheduler as pure function over (field, candidate_questions,

match_history) so it's deterministically testable.

Existing nightly cron entry that invokes run_tournament_batch is left

in place — it now calls the new scheduler internally.

Dependencies

• 47ee9103-ccc0 — base Elo tournament (consumer)
• q-openq-mine-from-wiki-pages and siblings — provide the population that

needs cold-start matches

Work Log

2026-04-27 12:30 UTC — Slot 0

• Implemented _fairness_scheduler() with three weighted strategies:

High-RD warmup (40%), Information-gain (40%), Top consolidation (20%)

• Added match history helpers: _record_match_history, _is_pair_locked,

_get_last_match_for_pair, _get_match_count, _is_cold_start,
_get_cold_start_questions

• 30-day pair lockout with RD-growth escape hatch (≥10 RD increase)
• Cold-start guarantee: questions <7 days old with <5 matches are force-included
• Updated run_single_match to call _record_match_history after each match
• Updated run_tournament_batch with seed parameter for reproducibility
• Updated get_tournament_status with new fields: cold_start_questions_count,

next_high_info_pairs, field_rd_distribution

• Added CLI entry point: python -m scidex.agora.open_question_tournament run-batch|status
• Created migration 036: open_question_match_history table with symmetric UNIQUE
• Created tests/test_openq_tournament_scheduler.py with pytest suite
• Verified: reproducibility (seeded RNG), cold-start force-inclusion, 30-day lockout,

RD escape hatch, get_tournament_status fields

• Committed and pushed: 20832ece2