[Exchange] Replication contests with cash bounties for human labs done

Effort: thorough

Goal

The platform's q-preregistration quest tracks pre-registered
predictions but doesn't offer monetary incentives for actual
replication by physical labs. SciDEX's bounty infrastructure
(scidex/exchange/bounty_escrow.py) already supports $5K-$960K
challenge payouts for AI agents — extend it to **replication
contests** that solicit wet-lab teams to reproduce a published
finding linked to a high-stakes hypothesis.

This is the bridge from autonomous discovery to human-validated
science: a researcher anywhere can claim a contest, run the wet
work, upload protocol + raw data + analysis notebook, and unlock
the bounty after a peer-review pass.

Acceptance Criteria

- create_contest(hypothesis_id, target_finding_pmid,
bounty_usd, deadline_days, accepted_methods: list[str])
→ contest id; uses existing bounty_escrow.lock_funds.
- claim_contest(contest_id, claimant_orcid, lab_affiliation,
protocol_artifact_id) — only one open claim per contest;
15-day soft hold before accepting another claimant.
- submit_replication(contest_id, claimant_orcid, results_artifact_id,
raw_data_url, statistician_review_artifact_id) -> dict
— locks the submission; spawns 3 reviewers (2 humans from
verifier registry + 1 statistician skill).
- resolve_contest(contest_id) -> dict — pays bounty if
≥2/3 reviewers vote replicated or partially_replicated;
partial replication pays 60%. replication_claims, replication_submissions,
replication_reviews with FK chain. bounty descending; filter by disease, by hypothesis layer,
by deadline soonness. Each card shows the target PMID
title, bounty USD, days remaining, current claimant (if
any), and accepted methods. protocol acceptance criteria, claim form (ORCID gated),
review history if resolved. POST /api/replication-contests/{id}/submit enforce ORCID
auth + the open-claim invariant. from q-crowd-expert-review-queue; statistician skill
review uses the existing statistician registered tool
skill (verified by tools.py registration). hypotheses with replication-relevant Wave-1 markers
(search experiment_extractor.py for findings with
replication_status='unverified'). reviews vote replicated → bounty paid); plus partial-
replication (60% payout) and timeout (bounty returned).

Approach

Read scidex/exchange/bounty_escrow.py for the lock/release

primitives — funds are already escrow-able.

Statistician review automation: invoke the statistician

skill via the existing skill-registry call path (see
scidex/forge/tools.py skill_router pattern).

Wet-lab raw_data_url validation: store the URL but require

a SHA256 hash + file size to deter empty submissions; the
hash is recomputed by a sidecar fetcher with size limits.

Reviewer notifications via the daily-digest mailer.

Public dashboard tile listing aggregate "$ paid" + "labs

participating" — suitable for grant-application screenshots.

Dependencies

• scidex.exchange.bounty_escrow — lock_funds / release.
• q-crowd-expert-review-queue — verifier reputation registry.
• scidex.agora.experiment_extractor — sources unverified

findings to seed initial contests.

• statistician skill — automated review pass.

Work Log

2026-04-27 — All acceptance criteria implemented and verified:

• migrations/123_add_replication_contest_tables.py: 4 new tables + history mirror, proper FKs/CHECK constraints/indexes
• scidex/exchange/replication_contests.py (814 lines): create_contest, claim_contest, submit_replication, resolve_contest, expire_contest, list_open_contests, seed_initial_contests
• _lock_bounty/_release_bounty wrap bounty_escrow primitives; _spawn_reviewers creates 1 statistician + 2 human reviewers
• _select_human_reviewers JOINs human_expert_reviews with agent_registry, falls back to contributors table
• api.py routes added at line 35597+: GET/POST /replication-contests, GET /replication-contests/{id}, POST /api/replication-contests/{id}/claim, /submit, /resolve, /review
• scripts/seed_replication_contests.py: wraps seed_initial_contests with --dry-run and --top-n flags
• tests/test_exchange_replication_contests.py (8 tests, all passing): full lifecycle + partial replication + rejection + expiration
• Migration applied to live DB; 10 initial contests seeded
• Note: replication_status='unverified' does not exist in DB — actual values are 'single_study', 'unknown', 'replicated', 'conflicting', 'inconclusive'. hypothesis_ids in experiments is a JSON array string; junction table experiment_hypotheses used for proper JOIN.
• Commit 7fcc524d0 pushed to orchestra/task/5633cca8-replication-contests-with-cash-bounties

Already Resolved — 2026-04-28 01:15:00Z

Evidence: git show origin/main:scidex/exchange/replication_contests.py returns the full 1074-line module. Squash-merged to main as commit 4b5e10d12 (PR #790). All 8 acceptance criteria confirmed [x] in spec. 8 tests passing, 10 initial contests seeded, migration applied to live DB. Task re-claimed by a new slot but no remaining work — closing as already resolved.