Reference

Policies Reference

Replay policies, fixture provenance, side-effect classes, capture/redaction policy, validity labels, and adapter capability status.

Policies are part of the evidence. Every replay, experiment, and report should name the policies that shaped the output.

Replay Policy

{
  "model": "recorded-only",
  "tool": "pause-for-fixture",
  "drift": "continue_to_terminal",
  "maxSteps": 100,
  "temperatureOverride": 0
}

Model Policy

ValueMeaningCurrent local engine
recorded-onlyUse captured model outputs.Supported.
determinism-probeRepeat under same policy to detect variance.Schema-supported; executor-dependent.
pinned-liveRe-run with a pinned provider/model/config.Fails without explicit executor.
compatible-liveRe-run with a compatible live model.Fails without explicit executor.
provider-fixtureUse provider-supplied fixture output.Schema-supported; executor-dependent.
blockedDo not allow model execution.Useful safety state.

Tool Policy

ValueMeaningCurrent local engine
recorded-onlyOnly use recorded tool outputs.Errors on missing divergent fixtures.
pause-for-fixturePause when a divergent output is missing.Supported and recommended.
analyst-fixtureAllow analyst-supplied fixtures.Modeled through fixture provenance.
simulatedAllow simulator outputs.Modeled through fixture provenance.
live-readonlyAllow live read-only tool execution.Fails without explicit executor.
live-explicitAllow explicit live tool execution.Fails without explicit executor.
blockedBlock tool execution.Errors on missing divergent tool output.

Drift Policy

ValueUse when
stop_on_first_divergenceYou only want to isolate the earliest changed event.
continue_to_terminalDefault for RCA because downstream effects matter.
continue_until_budgetFuture-facing mode for bounded long replays.

Fixture Provenance

ValueMeaning
recordedOutput copied from captured trace evidence.
analyst_fixtureOutput supplied by a human analyst or approved source.
ai_fixtureOutput simulated by an AI model.
simulatorOutput generated by a declared simulator.
live_readonlyOutput from a live read-only tool.
live_explicitOutput from a live side-effecting tool with explicit approval.

Side-Effect Classes

ValueMeaning
noneNo runtime side effect.
readReads local or external data without mutation.
writeWrites local state or files.
external_mutationSends, deploys, deletes, purchases, or mutates an external system.
unknownNot classified. Treat conservatively.

Managed runtime helpers infer side effects from tool names as a fallback. Override this for important tools.

Capture Policy

{
  "defaultAction": "capture",
  "rules": [
    { "path": "customerEmail", "action": "mask", "reason": "PII" },
    { "path": "authorization", "action": "drop", "reason": "secret" }
  ]
}
ActionCurrent implementation
captureCapture after default pattern redaction.
dropReplace field with null.
hashReplace field with [HASHED].
maskReplace field with [REDACTED].

encrypt, artifact_only, and metadata_only are not capture policy actions in v0.3.

Default redaction masks emails, phones, API-key-like tokens, JWTs, credit-card-like numbers, and secret-looking object keys.

Validity Labels

LabelTrigger
confirmed_by_replayMetrics exist and no stronger caveat is detected.
sensitive_to_fixtureFixture-related metrics are positive.
model_nondeterministicVariance-related metrics are positive.
diverged_but_comparableAny diff has diverged_but_comparable.
non_comparableAny diff has non_comparable.
unsupportedNo evidence supports stronger labels.

Adapter Capability Status

StatusMeaning
native_fullRuntime primitives support native replacement or safe interception.
managed_replayAdapter can capture and steer, but runtime limitations remain.
observability_onlyUseful trace capture without deterministic replay.
unsupportedPublic API does not expose the primitive.

Interception modes include native_replace, proxy_replace, post_result_replace, block_only, and observe_only.

On this page