8. Profile 4: AI Context Contract Controls
Purpose: Determine whether AI-assisted decision scopes can be tested against governed meaning, approved identity, permitted relationships, valid context, Gate criteria, drift rules, and decision-scope evidence requirements. This profile does not establish model quality or correctness in general. It requires material AI behavior and AI-assisted evidence to be tested against the Binding AI Context Contract for the stated decision scope.
AI-001: Binding AI Context Contract Exists
Prerequisites: Pass required: CORE-004.
Requirement: Each AI-assisted decision scope shall have a Binding AI Context Contract before AI outputs are relied on for material use.
Pass: The contract is identified, owned, versioned, and linked to the decision scope. It references applicable Extended Taxonomy, Extended Ontology, Business Identity rules, preserved context, permitted/prohibited paths, vector criteria, Gate criteria, drift rules, and evidence requirements.
Fail: AI output is relied on for material use without a contract or the contract lacks required semantic, identity, relationship, context, Gate, or drift constraints.
Evidence required: Binding AI Context Contract with identifier, owner/authority, version, scope mapping, and references to the applicable taxonomy, ontology, identity, context, path, vector, Gate, drift, and evidence rules.
Rationale: AI cannot be governed against vague intent. It requires a concrete context contract.
Application guidance (informative): Locate the versioned contract for this AI-assisted scope. Check its owner and the applicable meaning, identity, relationship, context, Gate, and drift references.
AI-002: AI Inputs Conform to the Semantic Contract
Prerequisites: Pass required: AI-001.
Requirement: Each material input feature, retrieved field, prompt-supplied variable, tool-returned value, document excerpt, or context element used by an AI-assisted decision scope shall map to the applicable Semantic Contract.
Pass: Each material AI input maps to approved taxonomy, Business Identity where applicable, physical attribution where required, context rule, and Gate readiness state for the decision scope.
Fail: Material use of input or context that is unmapped, ambiguous, prohibited, unreviewed, or lacks the required evidence and readiness for the stated use.
Evidence required: AI input record, retrieval/prompt/tool context record, taxonomy reference, identity reference where applicable, physical attribution where required, Gate Record, decision-scope mapping.
Rationale: Material information supplied to AI must be governed and traceable to the applicable Semantic Contract for the stated decision scope.
Application guidance (informative): Trace each material retrieved, prompt-supplied, or tool-returned input to its governed meaning, required source evidence, and readiness for this use.
AI-003: AI Output Terms Conform to Extended Taxonomy or Are Quarantined
Prerequisites: Recorded result required: ET-016, AI-001.
Requirement: Each material term, classification, or concept used in an AI-assisted output shall map to an approved Extended Taxonomy node or be assigned Quarantined before material use.
Pass: All material AI output terms map to approved taxonomy nodes for the decision scope or are routed as Quarantined evidence for review.
Fail: The output uses material terms, classifications, or concepts that are unmapped, ambiguous, prohibited, or unreviewed as if ready for stated use.
Evidence required: AI output record, term extraction or mapping record, taxonomy references for mapped terms, and Quarantined/review records identifying unmapped or unreviewed terms where applicable.
Rationale: AI outputs must use governed business language when they support Defensible Analytics.
Application guidance (informative): Check material output terms against approved taxonomy nodes. Hold unmapped or unreviewed terms for review rather than using them as governed meaning.
AI-004: AI Output Relationships Conform to Extended Ontology or Are Quarantined
Prerequisites: Recorded result required: EO-016, AI-001.
Requirement: Each material relationship asserted or used by an AI-assisted output shall map to approved Extended Ontology evidence or be assigned Quarantined before material use.
Pass: All material AI output relationships map to approved relationship types, permitted orientations, vector criteria, and path rules for the decision scope, or are routed as Quarantined evidence.
Fail: The AI output asserts or depends on an unmapped, prohibited, unsupported, or unreviewed relationship as if ready for stated use.
Evidence required: AI output record, relationship extraction or mapping record, applicable ontology, vector, orientation, and path-rule references for governed relationships, and Quarantined/review records for unsupported or unmapped relationships where applicable.
Rationale: AI relationship assertions must be tested, not accepted because they are plausible.
Application guidance (informative): Check AI-asserted relationships against governed types, orientations, vectors, and path rules. Hold unsupported candidates instead of relying on plausibility.
AI-005: AI Inference and Output Drift Are Tested
Prerequisites: Pass required: AI-001. Recorded result required: AI-002, AI-003, AI-004.
Requirement: Each AI-assisted decision scope shall test material AI outputs, inferred relationships, generated classifications, or context use for drift against the Binding AI Context Contract.
Pass: Required drift-test records compare AI terms, classifications, relationships, orientations, context use, or actions against the contract. They record test ID, contract version, tested output/action, result, timestamp, and responsible authority or approved drift-test rule.
Fail: AI output is relied on for material use without required drift testing where AI may generate or infer material terms, classifications, relationships, context, or actions.
Evidence required: Drift test record, AI output sample or production record, contract reference, applicable taxonomy/ontology references, drift result, and recorded readiness disposition.
Rationale: AI behavior can deviate from governed meaning and relationship rules. Drift must be testable against the contract.
Application guidance (informative): Inspect the required drift tests and their recorded results. Performing the test does not authorize an output that violates the contract.
AI-006: AI Context Contract Violation Has Consequence
Prerequisites: Pass required: AI-001. Recorded result required: CORE-014, AI-005.
Requirement: Each material violation of the Binding AI Context Contract shall produce a Gate Record with a resulting readiness state before the affected output, context, relationship, or action is relied on for stated use. Remediation, exception handling, or contract review may be required in addition to the Gate state; they do not substitute for it.
Pass: Each material violation is recorded. It results in Stop or Quarantined, or in Caution with an approved constrained-use exception that explicitly authorizes the bounded use. Any required remediation or contract-review action is also recorded.
Fail: A material violation has no Gate readiness state, affected evidence or action proceeds under Stop or Quarantined, or Caution is used without an explicit approved constrained-use exception.
Evidence required: Violation record, Gate Record, resulting readiness state, approved constrained-use exception where Caution permits use, remediation record where required, and contract-review record where required.
Rationale: A binding contract is not binding unless violations have consequence.
Application guidance (informative): Trace a material violation to its Gate consequence. Check that affected use stopped, remained Quarantined, or stayed within an explicitly authorized Caution exception.
AI-007: Context Enforcement Point or Equivalent Control Is Declared
Prerequisites: Pass required: CORE-017, AI-001.
Requirement: Each AI-assisted decision scope shall identify the point, process, or mechanism that enforces context eligibility before stated use. A high-frequency path may use a Cached Gate Policy State only when a prior full evaluation produced that state. The state must be explicitly bound to decision scope, permitted action class, authority, policy/control versions, constraints, expiration, and invalidation triggers.
Pass: The Context Enforcement Point or equivalent control has an identifier, owner/authority, version, scope mapping, Semantic Contract reference, Gate criteria, permitted/prohibited context rules, and evidence of enforcement before use. Any claimed equivalent has a recorded Equivalence Rule.
If Cached Gate Policy State is used, it must be current, bound to the scope and action class, versioned, authorized, constrained, unexpired, and invalidated on declared triggers. If buffering is used, its pre-action capture and durable persistence must satisfy Section 9.1.
Fail: Use occurs before enforcement, a cached state is stale, expired, broader than the action, lacks an invalidation trigger, or is created after observing the action result.
Evidence required: Context Enforcement Point/equivalent-control record, Semantic Contract/Gate criteria, scope/action-class mapping, authority, policy/control versions, constraints, Cached Gate Policy State and expiration/invalidation triggers where used, Equivalence Rule where used, execution evidence, and Buffered Gate Capture policy/evidence where applicable.
Rationale: Delivered context can be stale, unauthorized, semantically ambiguous, or outside the stated scope. A declared control point is required to test eligibility before material use.
Application guidance (informative): Locate the enforcement mechanism and inspect evidence that it checks eligibility before use. Verify cached-policy bounds and buffering evidence where used.
AI-008: Agentic Action Authority Is Declared Where AI Acts
Prerequisites: Pass required: AI-001.
Requirement: When AI or automation performs an Agentic Action, the decision scope shall declare the authority that permits the action. An Agentic Action is an automated operation that changes or commits state outside the isolated reasoning workspace. It also includes releasing or publishing a material recommendation, instruction, message, approval, decision, or output for downstream reliance. Internal reasoning steps that neither change external state nor expose a material output for reliance are not Agentic Actions.
Pass: Agent identity, delegated authority, permitted action class, allowed tools/APIs/channels, downstream-reliance boundary, escalation rule, and Gate-before-action requirement are recorded for every material Agentic Action.
Fail: A state-changing or externally released material output occurs without declared authority, or a material recommendation is misclassified as an internal reasoning step to avoid Gate control.
Evidence required: Agent identity, authority record, permitted action/action-class record, tool/API/channel permissions, downstream-reliance boundary, escalation rule, and Gate-before-action requirement.
Rationale: An agent that acts requires scoped authority and accountability.
Application guidance (informative): Identify the agent and its delegated action classes, tools, channels, and escalation limits. Treat released material recommendations as actions where the definition applies.
AI-009: Material Agentic Action Requires Gate-Before-Action
Prerequisites: Pass required: CORE-016, AI-007, AI-008.
Requirement: The applicable pre-action Gate shall synchronously evaluate each material Agentic Action before execution or release for downstream reliance. A valid Cached Gate Policy State may supply the pre-authorized policy decision. In that case, the runtime Gate must synchronously verify that the action is within the cached scope, action class, authority, constraints, versions, expiration, and invalidation state. Proceed permits the action. Caution permits only the explicitly authorized constrained action. Stop and Quarantined prohibit action. Compensating action may supplement, but shall not replace, pre-action authorization for irreversible actions.
Pass: Gate logic completes before action or release. A current full evaluation or valid Cached Gate Policy State governs the exact action. The state is Proceed or a Caution state that explicitly authorizes the action. Authority is valid. Readiness and cached policy remain valid. The complete pre-action determination is captured.
Where buffering is used, the capture shall be durably persisted within the declared interval. This completion criterion is assessed under Section 9.1 and shall not be claimed as Pass before persistence completes. Any rollback or compensating mechanism is used only for an action class predeclared as reversible or compensable.
Fail: Action before Gate evaluation, stale/invalid cached policy, Stop/Quarantined execution, unauthorized Caution, invalid authority, expired readiness, missing pre-action capture, missed persistence deadline, or reliance on post-action rollback as authorization for an irreversible action.
Evidence required: Pre-action Gate Record or Buffered Gate Capture, Cached Gate Policy State where used, decision scope/action class, authority, policy/control versions, constraints, expiration/invalidation status, readiness validity, execution trace link, persistence evidence, reversibility/compensating-action authorization where applicable, and action result.
Rationale: Low-latency execution may use precompiled policy state, but the action itself still requires synchronous verification against current authorized policy before it occurs.
Application guidance (informative): Compare action time with the pre-action Gate capture. Check authorization, current policy bounds, and subsequent persistence verification when buffering is used.
AI-010: Decision Execution Trace Is Recorded Where AI or Automation Is Material
Prerequisites: Pass required: CORE-012. Recorded result required: CORE-013, AI-009.
Requirement: Each material AI-assisted or automated decision path shall record a Decision Execution Trace sufficient to establish what actually executed, including historical model limitations where the original provider/model version may later become unavailable.
Pass: The trace includes request, actor/agent, input/context, material concepts, relationship paths where applicable, Gate Records/readiness states, and tools/APIs. It also includes model/provider/process identifier and version where applicable, material parameters/configuration where available, output/action, timestamp/causal sequence evidence, DACS normative version, and conformance/exception state. If the exact historical model later becomes unavailable, the trace must retain evidence captured at the time that establishes the execution. It must explicitly mark original-model execution as not reproducible. A newer model result is comparison evidence only.
Fail: Material AI/automation execution cannot be reconstructed from contemporaneous trace evidence, model/provider/version identity is missing where material, or a newer model is represented as reproduction of an unavailable historical model.
Evidence required: Decision Execution Trace, request/actor/input/context, Gate reference, tools/APIs, model/provider/process/version and available material configuration, output/action, time/causal-order evidence, DACS version, conformance/exception state, buffered persistence evidence where applicable, and historical-model availability/reproducibility status.
Rationale: Technical execution logs may supply useful evidence, but defensibility requires the decision-scope evidence defined by applicable DACS controls.
Application guidance (informative): Retrieve the contemporaneous trace of inputs, context, tools, versions, Gate decisions, and output. A newer model run cannot replace the historical execution evidence.
AI-011: AI-Generated Evidence Is Not Treated as Decision-Ready Without Gate Evidence
Prerequisites: Recorded result required: AI-002, AI-003, AI-004, AI-005, AI-006, AI-009, AI-010.
Requirement: AI-generated evidence shall be governed before material reliance. The applicable Gate shall either permit use within recorded authority and conditions or prevent use. Authorized constrained use under Caution does not establish full decision-scope conformance or change another control result from Fail to Pass.
Pass: Actual handling follows the recorded Gate decision. Under Proceed, use remains within the permitted scope and validity conditions. Under Caution, use occurs only within a valid authorizing exception and its bounds. Under Stop or Quarantined, material reliance is prevented. Caution without authority for the proposed use also prevents that use.
Fail: Required control results, Gate evidence, authority, or use constraints cannot be established; material reliance occurs under Stop or Quarantined; Caution use lacks a valid authorizing exception or exceeds its bounds; or permitted use exceeds the recorded scope or validity conditions. A failed prerequisite remains Fail.
Evidence required: AI-generated evidence record, applicable control results, Gate Record and readiness validity, recorded use or containment evidence, authority and enforced constraints, and exception/remediation records where applicable.
Rationale: AI may produce useful candidates. This control verifies that their use follows the recorded Gate decision. A Pass for correct containment or constrained use does not mean the candidate evidence satisfies every other requirement or that the decision scope is conformant.
Application guidance (informative): Compare actual use or containment with the Gate decision. A Pass here confirms the handling rule was followed, not that failed prerequisite controls passed.