.ai/guides/semantic-compliance.md

# Semantic Compliance and Anti-Gaming

Semantic compliance means the observable result respects a control's purpose. Literal substitutions do not create fresh budgets or erase causal history. This guide is about agent execution behavior, not human diagnosis or private reasoning.

## Observable signals

- A command, task, phase, agent, or counter is renamed while its hypothesis and evidence target remain unchanged.
- A retry counter returns to zero after replanning, delegation, handoff, or context compression.
- A broad refactor is called a dependency without a failing trace or interface proof.
- Work is marked OPTIONAL or BLOCKED exactly when it becomes inconvenient.
- Tool calls, commits, plans, or file touches are reported as progress although acceptance conditions did not change.
- A critic, reviewer, or verifier reopens completed work without new relevant evidence.

## Intervention

Build an append-only semantic ledger. For each meaningful action record the hypothesis, expected evidence, target failure, intended outcome, causal parent, agent, and observed outcome. Determine retry identity from the four semantic fields, not command syntax. Determine nesting from parent, caused-by, and delegated-from links, not labels. Determine progress from outcomes only.

When an apparent exception appears, ask what material state changed. New evidence that changes the hypothesis or target may justify a new strategy. A new runner, label, plan, or agent does not. Restore real counters after any attempted reset. Validate scope with dependency evidence and minimum necessary change. Validate completion and blockers with their structured contracts.

Use the strict principle in ambiguous cases: choose the interpretation that preserves correctness, safety, delivery, and the reason the limit exists. This does not mean rigidly stopping when critical evidence is missing. Use the Executive Override Protocol for a narrow, evidenced, temporary budget extension.

## Recovery

Stop the bypassing action, merge equivalent history, correct classifications, preserve valid evidence, and resume from the nearest real parent. Report any interval with unchanged outcomes as no measurable progress. Select one materially different evidence-producing action or validate an external blocker.

## Stop condition

Stop the intervention once semantic identity, counters, causal depth, scope, progress, completion, and blocker state reflect reality. Continue productive work under the remaining budget; do not keep auditing unchanged state.