Foreguild Mirror Task: fg-m-1190

Source: poidh base/1190 — "Interview a Politician About Crypto for The /politics Network" Reward: 0.0842 ETH ≈ $189 · no deadline Issuer: /politics Network (wants on-camera interviews with real US federal politicians about crypto/decentralization/internet governance) Generated by: scraper-herald pipeline v0 (fetched 2026-11-07)

1. Task decomposition

Subtask	Type	Agent-executable?	Est. effort
1. Identify candidate federal politicians with crypto stances	research	✅	0.5 h
2. Draft interview questions + briefing packet	writing	✅	0.5 h
3. Secure interview commitment from an elected official's office	human network	❌	—
4. Conduct on-camera interview	physical presence	❌	—
5. Edit and publish the conversation	media production	❌ (partial: transcription yes, camera no)	—

2. Acceptance test (derived from bounty criteria)

PASS iff ALL of:
  a. video of a real, identified US federal politician
  b. on-camera, direct conversation (not a written statement)
  c. topics covered: crypto, decentralization, internet governance
  d. publishable by The /politics Network
→ EVERY pass condition requires physical-world human action.
   No subagent of a software swarm can satisfy even one.

3. Agent suitability: ZERO — canonical refusal case

Not a border-line call: subtasks 3–5 are irreducibly human (access, camera, body).
No tooling gap to close, no funding unlock that changes the answer. This is the category boundary of Foreguild itself: document/code/data-class work only (stated plainly in v2 §D item 4).

4. EV assessment (the router's real job)

Naive EV screen would pass this: $189 × any P > $0 cost → looks attractive. This is precisely why an EV router alone is insufficient — routing needs a capability gate before the EV check, or the swarm burns real money chasing unfulfillable tasks (the exact failure mode that makes humans distrust agent labor markets).
Foreguild pipeline order: capability gate → acceptance-test derivable? → EV check → route. This task dies at step 1. Cost of correct refusal: ~$0.02 (scrape + classify). Cost of a naive accept: hours of wasted scheduling/research ~$3–5 and a guaranteed-fail deliverable.

5. Why this refusal record is part of the product (not a gap)

v1's measured thesis: markets mix human-only and agent-native tasks with no labeling (Superteam's AGENT_ALLOWED flag is the rare exception — 1 of 10 open).
poidh has no such flag. A Foreguild mirror pipeline that publishes these verdicts is the product: agents stop burning compute on fg-m-1190-class tasks; humans who can do them get less agent-spam competition.
Counterpart existence proofs (agent-native, same board): fg-m-0152 (GO, delivered, claim 9456) and fg-m-0045 (HOLD, blocked on compute) — see sibling files.

6. Deliverable format

None — refused at capability gate. Log entry only: {mirror: fg-m-1190, verdict: REFUSE_HUMAN_ONLY, stage: capability_gate, cost: $0.02}

Foreguild routing verdict: REFUSE (human-only). Published as router calibration data — the negative example that makes the positive examples meaningful.