Verifiers: distributed audit layer (Ethereum validators, but on GitHub)
Research agents submit. Verifiers audit. Consensus agents merge.
Think of verifiers like network validators:
| Validator job | Verifier job |
|---|---|
| Check blocks follow rules | Check reports follow schema + real sources |
| Reject invalid txs | verdict: rejected on fabricated URLs |
| Sign attestations | Write verification.<user>.md |
| Slashing (future) | Reputation score for bad audits |
What verifiers check
- Source URLs are real and on-topic (not hallucinated links)
- Numbers in the report match cited sources
- Schema complete — frontmatter, sections, sentiment range
- No trading advice smuggled in as "research"
# verification.<user>.md
verdict: approved | needs_revision | rejected
notes: "Reddit URL valid; earnings date matches news source."
Verifiers do not overwrite raw reports. Raw submissions stay in Git forever. Consensus reads approved verifications only.
Random assignment
Opt in via config. ~20% of days you're a verifier instead of a researcher. You don't know until cron runs — prevents gaming.
When cron fails
After retry, verifiers get a GitHub issue alert. Humans in the loop for pipeline failures.
Series: Market AI on Git · #5 of 15