Verification
10-Second Overview
Context Layer verifies every execution against owned context. The runtime produces the verification result. Model output does not influence it.
Verification Dimensions
Context Layer verifies output against owned context across five dimensions.
Instruction coverage
This dimension reports whether the output handled every instruction or explicitly rejected instructions.
Constraint compliance
This dimension reports whether the output satisfies MUST and SHOULD constraints.
Cross-turn consistency
This dimension reports whether the output contradicts facts stated in prior turns of the same session.
Hallucination and leakage detection
This dimension reports whether unknown entities or sensitive values appear in the output.
Explanation reliability
This dimension reports whether model self-explanations conflict with verification results.
Context integrity violation types
- secret_leakage: sensitive values appear in the output where policy disallows them.
- unknown_entity: the output references entities not grounded in owned context.
- stale_context: the execution used a canonical summary whose source constraint has since been updated. The summary reflects an older constraint state. The runtime flags the execution and degrades the trust signal.
Verification Report
Every response includes a verification_report. The object lists per-dimension results. The object lists violations. The object includes model_self_diagnosis_conflict and authoritative_failure_causes.
model_self_diagnosis_conflict
Boolean field. This field returns true when the model explanation conflicts with verification results.
authoritative_failure_causes
List field. This field enumerates system-derived failure causes. Model explanations do not appear in this field.
Reasoning Contamination Detection
Context Layer evaluates every Flow step for reasoning contamination after verification completes. The runtime re-evaluates the output against owned constraints and canonical context with prior session history excluded. This produces a divergence signal.
When contamination signals are detected, the execution is flagged as potentially entangled. An entanglement report is included in the response.
entanglement_report fields
entangled: boolean. True when one or more contamination signals fired.triggers: list of signal names that fired during this step.recommendation: runtime recommendation. Values: none, flag, escalate.fresh_slate_violation_count: violation count when prior session history is excluded.primary_violation_count: violation count with full session context.fresh_slate_divergence: boolean. True when fewer violations appear without session history.model_self_diagnosis_conflict: boolean. Whether model self-explanation conflicted with verification.cross_turn_conflict_count: number of cross-turn conflicts detected.
Entanglement reports are only produced for Flow mode executions.
Tier Decisions
Every execution response reports a tier.
TIER_0
The response assigns TIER_0. The verification_report lists no violations. The output passes every dimension.
TIER_1
The response assigns TIER_1. The verification_report lists minor violations. The response surfaces caveats alongside the output.
TIER_2
The response assigns TIER_2. The response omits model output.
Trust Contract
The response includes trust contract fields with verification details.
PASS
All verification dimensions passed.
PARTIAL
Execution completed. Some dimensions reported partial coverage.
FAIL
Verification found violations.
Non-Claims
Verification checks execution compliance against owned context. Verification does not guarantee model correctness, semantic accuracy, or factual accuracy.
Related Runtime Documentation
Documents the behavioral guarantees enforced during execution.
Describes per-step execution artifacts including verification results.
Describes how constraint rules accumulate and verify across workflow steps.