ADR-0180 - Evidence-bounded execution

Context

ADR-0171 removed routine narration and duplicate explanation. It did not stop agents from repeating reads, searches, commands, tests, or reviews after the relevant state stopped changing. Campaign handoffs could also preserve results while burying the next action in process prose.

Prompt wording changes the amount and kind of work coding agents perform. Unbounded certainty and multi-approach demands increased reasoning and verification without measured success gains in one preregistered study [PROMPTWASTE]. Experience-driven stopping reduced cost with negligible aggregate resolution loss in its evaluated configurations [EET]. Compact action-state-result communication improved the performance-cost tradeoff across evaluated agent systems [PACT]. Raw token removal alone can still raise billed cost or destroy exact evidence [TOKENCOST].

Decision

1. Repeat only after change. Repeat a read, search, command, test, or review only when relevant state changed, the previous attempt failed, or independent repetition is required.
2. Stop when proven. Stop when the requested result exists and proportionate verification passes. Extra searching, testing, reviewing, or explanation requires a concrete unresolved risk.
3. Keep handoffs operational. Campaign worker and phase handoffs carry only the action, current state, result, decisive evidence, unresolved conditions, and next owner or action.
4. Preserve exact evidence. Economy never licenses lossy rewriting of code, errors, edit anchors, requirements, or other action-critical evidence.
5. Add no control plane. No execution mode, completion manifest, artifact registry, runtime progress protocol, CLI command, or MCP tool is added.
6. Measure accepted outcomes. Field cost includes execution, coordination, retries, and rework. Token reduction remains diagnostic, not proof of savings.

Narrowed decisions

• ADR-0162, ADR-0163, and ADR-0167: campaign handoffs become compact operational state, not process history.
• ADR-0171: global economy now stops unchanged repetition and evidence-free continuation.
• ADR-0178: the stable operating loop gains explicit repetition and stopping boundaries.

Consequences

• Diligence must change confidence or stop.
• Campaign successors receive enough state to act without replaying the transcript.
• Exact evidence survives compression.
• Suspec gains two brakes, not another machine.

Status

Accepted (2026-08-12). Narrows ADR-0162, ADR-0163, ADR-0167, ADR-0171, and ADR-0178.