Skip to main content

​
Automatic bounties

Automatic bounties help route human attention toward records that need review, evidence or resolution. They are not an unattended treasury faucet. The engine evaluates candidate conditions, deduplicates work, checks policy and records enough context for the decision to be audited later.

​
Published trigger classes

Current V3 policy includes trigger categories such as:
  • unreviewed publication after the configured interval
  • missing evidence after the configured interval
  • single-source reporting after the configured interval
  • breaking news that remains unreviewed
  • low AI confidence
  • material correction
  • unresolved challenge
  • high-impact record with low participation
Some authority files publish testnet observation intervals. Applications should consume active policy rather than assume those values are permanent mainnet thresholds.

​
Proposal before payment

A useful automated system separates detection from funding. A detector can identify a condition, but the proposal layer can still ask:
  • does the condition still exist?
  • is an equivalent bounty already active?
  • would funding violate a treasury floor?
  • has the automatic bounty budget been reached?
  • is the system paused?
  • does the target Capsule still exist in the expected state?
Only after those checks should a bounty move into a funded lifecycle.

​
Example: unresolved challenge

An open challenge can create demand for independent review. An evaluator can recognize that the challenge remains unresolved and propose verification work tied to the exact Capsule and challenge context. The task should be specific enough that settlement can determine whether the expected work was delivered. “Investigate this article” is not as auditable as a task that references the unresolved challenge and required evidence or review output.

​
Example: low participation

A high-impact publication with unusually low review participation can be routed for attention. This is not evidence that the publication is false. It is evidence that the verification process has a participation gap. Product copy should preserve that distinction.

​
AI-assisted triggers

Low AI confidence can be one signal for routing work, but model output should not directly become a protocol verdict. The safer use is triage: a model identifies material that deserves human attention, then normal bounty eligibility, review and settlement controls remain in force. Keep model version and source information if a trigger depends on an automated analysis system.

​
Protect sensitive anti-abuse policy

A public client does not need every internal anti-Sybil threshold in order to explain why a proposal exists. A useful public record can show:
without publishing controls whose primary effect would be to make abuse easier.

​
Idempotent operation

Automated evaluation must be safe to retry. Store stable proposal identity, source condition, policy version and observed chain coordinates. Re-read current state before opening or settling the bounty. A restarted worker should continue the same logical task, not create a new economic obligation.

​
Human accountability remains

Automation routes work. It does not remove human responsibility. Claim, submission, approval and settlement still have explicit actors and authorization boundaries. That is what makes automatic bounties useful without making them opaque.