AI can reason probabilistically. Assurance should be grounded in evidence.
When AI agents can move money, access customer data or change infrastructure, a persuasive answer is not enough. HAIEC applies defined assurance rules to qualified evidence about consequential action paths and their authority. The result is a scoped conclusion your team can inspect, revisit and use to decide what happens next.
Let AI help run the workflow. Do not let it approve its own consequential decisions.
Ask your AI to run HAIEC AI Assurance through a connected agentWould you trust an AI's answer to approve its own action?
This is not an argument against using AI. It is a question about who decides. A model can produce a confident explanation of why an action is fine. Confidence is not evidence. When the action is consequential, the determination deserves a method whose inputs and rules can be named, versioned and inspected.
Deterministic assurance does not remove AI from the loop. It removes the model from the seat of final judgment over its own consequential choices.
One agent action, two ways to assess it
A support agent can call a refund tool. The refund handler accepts a recipient parameter. Here is the same scenario assessed two ways.
Option A
A persuasive AI-generated answer
"The agent is allowed to issue refunds. The tool requires authentication, so this action is authorized."
- •Sounds complete. Provenance unstated.
- •"Allowed" mixes technical access, policy and delegation into one word.
- •Cannot be re-derived from named inputs; a second run may answer differently.
An LLM explanation is context, not malicious. It is simply not the determination.
Option B
HAIEC evidence-bound assessment
Input scope, claim, evidence basis and proof frontier are visible. No fabricated verdict.
Technical access does not automatically delegate a business choice.
Fixed evidence, claim and snapshot produce a repeatable scoped conclusion. A new issuance may get a new timestamp and receipt hash; the conclusion basis is what repeats.
Five practical benefits
Consistency
The same qualified evidence, claim, rule version and scope produce the same scoped conclusion. Reviewers argue about the evidence, not about which run to believe.
Reproducible explanation
A reviewer can trace a conclusion back to the exact evidence roles, versions and scope that produced it, instead of trusting an opaque narrative.
Authority checks
The evaluation separates technical access from policy authorization and delegated choice, so a powerful credential is not mistaken for an approved business decision.
Evidence and receipt integrity
Where artifacts, evidence sets and receipts are issued, their integrity is recorded so a change after issuance is detectable. Integrity is not a truth claim.
Change-aware review
Frozen evaluations keep their original basis. New code, policy or models are assessed under a new scope, so drift is compared rather than hidden.
Deterministic evaluation is not deterministic AI behavior
The method is deterministic; the system under review is not. A model can change its answer. The evaluation of the evidence about it does not, given the same qualified inputs.
- Source-visible and code-capable paths can be established early, before every connector exists.
- Policy authorization is real where qualified policy evidence supports it.
- IAM effective grant is a separate plane: it shows what the credential permits, not who delegated the choice.
- Observed effects are claimed only with recorded runtime evidence in the correct scope.
- Missing evidence is not a pass. Not-evaluated planes stay visible on the proof frontier.
The five evidence planes are described on the methodology page and the System Constellation.
How it complements the tools you already have
Scanning, observability, model testing and governance are each independently useful. None of them, alone, answers the consequential question HAIEC asks.
Code scanning
Finds risky patterns and source-visible paths in supported code.
HAIEC relates source-visible paths to the consequential question: what can this reach, and which authority applies.
Observability
Records what running systems actually did.
Observed telemetry is one evidence plane. It cannot retroactively establish the policy or delegated authority behind an action.
Model testing
Measures model behavior, quality and failure modes.
A well-tested model can still take an unauthorized action through a connected tool. Behavior scores do not answer authority questions.
Governance tooling
Tracks inventories, policies and attestations.
Governance records what teams declared. HAIEC compares declarations against the applicable evidence where connected.
Inspect the proof
The Kestrel example is a real, frozen source-backed evaluation. Its report shows the action paths it assessed, the scoped conclusion it produced, and the explicit frontier where evidence was partial or unknown.
The example is bounded by its frozen scope. It is not a universal product benchmark, and its conclusions apply to the evaluated evidence, not to every deployment.
Better audit and change review
Screenshots and manual records still have a place as context. Source-qualified, versioned records simply make review easier to repeat and harder to silently alter.
Versioned inputs and outputs
Each evaluation records the claim, rules and evidence scope that produced it, so reviewers can compare runs instead of guessing which inputs changed.
Integrity where issued
Named evidence content and issued receipts carry integrity protection appropriate to their object class. Detection of change is not the same as proof of truth.
Retention controls
Named content and evidence-hash columns are locked against update, and dedicated audit-log tables are protected against update and delete. That is a bounded retention control, not a claim that every record in the platform is immutable.
Framework support
Scoped, versioned outputs can support framework mapping and auditor review. Mapping is not certification, and no record format guarantees acceptance.
For external grounding on repeatable documented testing and uncertainty, see the NIST AI Risk Management Framework. NIST does not endorse HAIEC.
What would you want to assess first?
Pick one AI system and one consequential choice it could make. That is enough to scope a bounded evaluation.
The MCP path requires real credentials, an authorized source, preflight and user consent. It is an integration guide, not a one-click demo.
Common questions
Does deterministic evaluation mean the AI behaves deterministically?▾
No. A model can generate different answers each run. Deterministic refers to the evaluation method: the same qualified evidence, claim version, rule version, scope and evaluation snapshot produce the same scoped conclusion. The AI system under review remains probabilistic.
Can an AI model evaluate or approve itself?▾
AI can help run the workflow: it can explain findings, call bounded tools and assemble context. But the consequential determination should come from a deterministic evaluation of qualified evidence, not from the model approving its own output.
What does a content hash actually prove?▾
A hash proves integrity, not truth. It shows that a specific recorded artifact, evidence set or receipt has not changed since it was recorded. Each object class has its own integrity owner: normalized evidence content, uploaded file bytes, issued evaluation scope, and published receipts are different things. A verified receipt does not certify that the underlying claim is safe or true.
Do all five evidence planes have to be connected before the evaluation is useful?▾
No. A meaningful source-visible action path can be established from code evidence alone. Policy authorization, effective grants and observed runtime effects are separate evidence questions. Missing planes stay visible as not evaluated rather than being treated as a pass.
Is deterministic assurance evidence automatically accepted in an audit?▾
No method guarantees auditor acceptance. Deterministic, versioned records with explicit scope and stated limitations are easier to review than unverifiable assertions, and they can support framework mapping work. Acceptance remains the auditor’s decision; mapping to a framework is not certification.
Can HAIEC observe that an AI action actually happened in production?▾
Only when the relevant runtime evidence sources are connected and qualified for that scope. Source analysis establishes what the code is capable of; it does not observe execution. When observed evidence is absent, the conclusion records that frontier instead of inferring an effect.
What happens when the code, policy or model changes?▾
A frozen evaluation keeps its original evidence, versions and snapshot. Changed inputs are assessed in a new scoped evaluation, so reviewers can compare what changed rather than silently rewriting the earlier conclusion.
Decide with evidence
Start with one consequential action path and the authority questions behind it.