[Senate] Persona ladders - round-robin Elo tournament across personas done

Effort: thorough

Goal

scidex/exchange/elo_ratings.py and scidex/senate/judge_elo.py already
maintain Elo for hypotheses and judges; nothing maintains an Elo *for the
personas themselves*. We have ~50 personas under personas/ (9 founding
plus dozens of scientists + philosophers); some are clearly more
informative debaters than others, and we cannot tell which without a
ranking signal. Build a persona ladder: a continuously-updated Elo
rating per persona derived from their pairwise debate-round wins (using
the weighted verdict from q-debate-evidence-weighted-vote), with a
round-robin scheduler that ensures every persona meets every other
persona over a sliding 90-day window. Top of the ladder gets first-pick
into newly-spawned debates; bottom gets retired-pending-review.

Acceptance Criteria

judge-meta arena in judge_elo.py:44) — implemented directly in
persona_ladder.py using PERSONA_ARENA = "personae". - record_pair_match(session_id, persona_a, persona_b, winner,
weight_multiplier=1.0) — uses the weighted verdict to score.
- leaderboard(window_days: int = 90, limit: int = 100) -> list[dict]
- coverage(window_days) -> dict (returns
{matched_pairs: n, possible_pairs: m, coverage_ratio: …}).
- pick_next_pair(window_days) -> tuple[str, str] — returns the
pair with the fewest recent matches plus the highest combined
Elo (informativeness ⨯ data-need). every debate_sessions with weighted_verdict_json populated and
records all C(n, 2) pairwise matches per session. POST /api/agora/persona_ladder/schedule_match accepts a topic
and creates a 2-persona debate (using
scidex/agora/scidex_orchestrator.py:run_debate with
persona_ids=[a, b]) for the next pair. from pick_next_pair and schedules debates directly via the
orchestrator. Elo trend sparkline per persona, current coverage % bar, and
"Next scheduled pair" widget. matches gets a row written to senate_alerts proposing review;
no auto-action. leaderboard, coverage, pick_next_pair, retirement alerts, and full
workflow. Elo distribution across personas.

Approach

Reuse elo_ratings.update_match_result — only difference is the new

arena name.

pick_next_pair is argmin(recent_match_count) + tiebreak on Elo;

keep simple.

Scheduling matches as Orchestra tasks (rather than firing them

directly) lets the existing fleet pick them up with the appropriate
model effort.

HTML reuses templates/agora/base.html.

Dependencies

• q-debate-evidence-weighted-vote — supplies the per-round weighted

winner.

• scidex/exchange/elo_ratings.py — Elo machinery.
• scidex.agents.select — persona pool source.

Dependents

• q-persona-disagreement-scoreboard — high-Elo ↔ high-Elo pairs are

the most informative disagreements; ladder data feeds the scoreboard's
prior.

• q-debate-replay-cross-topic — ladder Elo helps the replay engine

pick personas for new topics.

Work Log