[Atlas] Spawn open questions from falsified hypotheses + market-resolved misses done

Goal

When a hypothesis prediction is falsified or a prediction market resolves
against the consensus, the most valuable downstream artifact is the new
question that the falsification opens. ("If APOE4 → tau spread is wrong, what
DOES drive tau spread in the absence of APOE4?") Today, falsifications close
the hypothesis loop without spawning any new search frontier. This task wires
a falsification → open_question generator so every refuted prediction becomes
a ranked question for the field.

Acceptance Criteria

- hypothesis_predictions.outcome_status flipping to
falsified / unsupported (poll table; no event_bus dependency).
- markets.resolution_status='resolved' with the resolution disagreeing
with the pre-resolution consensus probability by ≥0.4. domain-expert persona (selected from DOMAIN_JUDGES in
scidex/agora/open_question_tournament.py) to generate 1-3 successor
questions that the falsification specifically motivates. Each must
include a verbatim "what we now know" + "what we still don't know"
contrast. - metadata.source_kind='falsified_prediction' or 'market_resolution'
- metadata.parent_hypothesis_id set
- metadata.field_tag inherited from the parent hypothesis
- metadata.importance_elo seeded at the parent hypothesis's prior
Elo + 50 (falsifications point at the most-bet-on questions, so they
deserve a head start in the per-field tournament). hypothesis → new question), and the new question gets a
link_type='succeeds' edge back. (SELECT id FROM hypotheses WHERE status='falsified') and emit ≥30
new open_question artifacts; assert in test. questions (dedup via question_hash). and dedup behavior.

Approach

Inspect hypothesis_predictions schema in scidex/core/database.py and

markets table in exchange.py for resolution columns.

Reuse persona selection logic from

scidex/agora/open_question_tournament.py DOMAIN_JUDGES.

Run as a daily systemd timer (scidex-openq-falsified.timer); single-shot

for backfill.

Write resulting question count + sample to

data/scidex-artifacts/reports/openq_falsified_<utc>.json.

Dependencies

• q-openq-mine-from-wiki-pages — shared dedup util
• b2d85e76-51f3 — open_question schema
• 47ee9103-ccc0 — Elo tournament reads seeded importance_elo

Work Log

2026-04-27 — Implementation (task:bbe35802)

Schema findings:

• hypothesis_predictions.status (not outcome_status) holds 'falsified'/'unsupported'. 5 falsified rows exist (test data).
• hypotheses.status='falsified' — 0 rows currently; backfill handles both paths.
• markets uses current_price + resolution_price; consensus fallback via metadata['consensus_probability'].

Deliverables:

• scidex/agora/open_question_from_falsified.py (378 LoC):

- process_falsified_prediction / process_market_resolution per-source processors
- run_poll(since_hours) / run_backfill(limit) batch runners
- Idempotency: source_id + source_kind guard prevents re-processing
- Dedup: SimHash question_hash from open_question_miner_wiki
- Elo seeding: parent_hypothesis_elo + 50; answered_by + succeeds links
- DOMAIN_JUDGES persona selection from open_question_tournament

• tests/agora/test_open_question_from_falsified.py — 19 tests all pass:

field-tag inference, heuristic stub, dedup, link-graph shape, idempotency,
market divergence gate, backfill≥30 artifacts, Elo boost assertion

• deploy/scidex-openq-falsified.{service,timer} — daily at 03:00 UTC