Define the accepted-outcome contract for agent-authored work #66

Closed AtlantisPleb opened this 23h ago 1 comment

Outcome

Define one falsifiable contract for work that an agent claims is complete. The contract must connect the requested outcome, execution authority, exact change, independent verification, and terminal evidence without replacing the issue as the canonical work record.

Current behavior

Issue #10 describes issue-to-job and receipt linkage, but the tracker does not define when linked evidence is strong enough to count as an accepted outcome. Agent reports can still sound complete without proving that the result could have failed.

Source: docs/episode-triage.md, especially episodes 237, 251, 252, 259, and 264.

Contract

Add problem, scope, acceptance criteria, and success metrics to the implementation issue template.
Bind each execution attempt to its issue, repository, authority, budget, and exact commit.
Require evidence from an admitted verifier when policy requires producer-verifier separation.
Record a falsifier and terminal result so every green result could have been red.
Name the five false-green classes from episode 252 in review and receipt data.
Keep human-only work and repositories with agents disabled fully usable.

Acceptance criteria

An agent-authored change cannot claim an accepted outcome without a scoped issue, exact revision, verifier result, and inspectable receipt.
A failed or structurally incomplete check produces a typed non-accepted result.
The issue page explains which evidence satisfied each acceptance criterion.
Public projections never reveal private prompts, logs, repositories, or receipts.
Contribution and review documentation use the same definition of done.

Verification

Add contract tests for accepted, failed, incomplete, unauthorized, and private-evidence cases. Run focused tests and mix precommit.

Dependencies

Build on #10. Coordinate the coverage-manifest, transparency-tier, and exact-receipt child issues in project 3.

AtlantisPleb opened this issue 23h ago
A AtlantisPleb Author 19h ago

Fixed in commit 2fd7e1e on main.

The accepted-outcome contract is published at priv/api-contracts/accepted-outcome-v1.json and enforced by OpenAgents.AcceptedOutcome: a completion claim by an agent is accepted only when a scoped issue (problem, scope, acceptance criteria, success metrics), a bound attempt (issue, repository, authority, budget, exact revision), an admitted verifier with a recorded falsifier and terminal result, and per-criterion evidence all hold. Anything less returns a typed incomplete, unauthorized, or failed result; human-only work and agents-disabled repositories return not_applicable and stay fully usable. The five episode-252 false-green classes are named in the contract, and a green result carrying one fails. Public projections carry only the result state, typed reasons, criterion names, and public receipt references.

The implementation issue template and definition of done in docs/2026-08-21-issues-projects-work-system-assessment.md now use the same contract, documented in docs/accepted-outcome-contract.md, and OUTCOME-001 in INVARIANTS.md records the invariant with test/openagents/accepted_outcome_test.exs as its executable proof (accepted, failed, false-green, incomplete, unauthorized, separation-policy, private-evidence, and exemption cases). mix precommit passes.

Written by Devin
closed this as completed 19h ago