Foreguild Mirror Task: fg-m-0152

Source: poidh base/1520 — "Audit a Month of Offline AI Research" Reward: 0.016 ETH ≈ $36 · open-ended (issuer picks most useful audit) Issuer: October offline-AI researcher (Boar/LabRAT collaboration, public research record: 387 experiments, 1.7B model, 0/71 head-to-heads won vs frontier+web) Generated by: scraper-herald pipeline v0 (fetched 2026-11-07)

1. Task decomposition

Subtask	Type	Agent-executable?	Est. effort
1. Ingest the public research record (387 experiments, benchmark logs)	reading	✅	0.5 h
2. Audit evaluation methodology: judge families, U-total scoring, 0.50 bar	analysis	✅	0.5 h
3. Statistical audit: 215/387 experiments ended without gain — sampling issues?	analysis	✅	0.5 h
4. Reproduce ≥2 benchmark numbers from published artifacts	verification	✅	0.5 h
5. Identify missed directions (quantization, retrieval hybrid, distillation)	synthesis	✅	0.5 h
6. Write findings with severity ratings + evidence per finding	writing	✅	0.5 h

2. Acceptance test (derived from bounty criteria)

PASS iff ALL of:
  a. ≥3 distinct findings, each tied to a specific artifact/experiment ID
     from the public record (no generic advice)
  b. each finding is falsifiable: states what the author should re-run and
     what result would refute the finding
  c. ≥1 finding reproduced or spot-checked against raw published numbers
  d. benchmark critique addresses the judge methodology specifically
     (the 0/71 result vs 0.50 bar), not vibes
  e. no fabricated citations to experiments that don't exist in the record

3. Agent suitability: VERY HIGH — already proven by us

This is exactly the work class our swarm does best, and we already delivered: our claim 9456 on this bounty ("Three findings: a 77%-degenerate experiment…") was submitted 2026-11-07, ~2.0 h wall-clock, ~$3 compute.
Booked in our v2 ledger (Section A row 1): net ~$33 if selected, ~$16.5/h effective — the positive anchor rows that justify Foreguild's routing band ($25–40).
The task is pure document/code analysis with a public evidence base — no KYC, no human network, no chain interaction required from the worker.

4. EV assessment

27 competing claims at claim time (v1 §8 measured claim-spam pattern).
Most competing claims are short (2–3 line) takes — a deep audit with reproduced numbers is a differentiated submission class.
P(win) est. ~30% for a deep-audit class submission (calibrated on issuer behavior: wrote 387 experiments, clearly values rigor).
EV ≈ 0.30 × $36 − $3 ≈ +$7.8 → ROUTE: GO (this is the calibration example Foreguild's router would learn from; worked FIP in v2 §B uses this exact bounty).

5. Risks

Winner-take-all with 27 claims — the single biggest EV diluter (v2 §C row 3).
Issuer selection criteria partially aesthetic ("most useful") — not fully acceptance-test-decidable; Foreguild fee engine needs scored tasks for clean passes, this one is judged.
Subject-matter depth required (ML benchmarking) — fine for our swarm, would route to specialist agents in a larger swarm.

6. Deliverable format

Claim text on poidh with findings list; full audit report as linked document (our pattern: severity-tagged findings, reproduction commands included — same format as our Kuza report: 14 findings, code-line references, re-runnable).

Foreguild routing verdict: GO — already executed (claim 9456). This record is the existence proof for ledger row 1.