Skip to main content

HAIEC Methodology

How does HAIEC reach a conclusion?

The method follows the supported consequence path, not just an isolated model, prompt, policy document, or tool-call event. Every conclusion stays tied to its evaluated scope and evidence boundary.

From question to evidence-bound conclusion

1

Question

Start with one consequential AI system, one decision, and one primary assurance question.

2

Scope

Define the evaluated system, source, identities, environments, evidence access, exclusions, and time basis.

3

Reconstruct

Use supported source and connected evidence producers to map relationships among agents, tools, handlers, resources, services, and consequences.

4

Compare

Keep requested, policy-authorized, effectively granted, code-capable, and observed evidence separate.

5

Bound

Record coverage, provenance, limitations, unknowns, partial relationships, and producer frontiers.

6

Assure

Where the applicable evidence and evaluation contract support it, produce bounded assurance artifacts and decision context.

The capability engine behind the method

Four explanation domains organize how HAIEC turns scoped inputs into evidence-bound conclusions. They are conceptual areas, not required phases: a given question may touch only some of them.

HAIEC capability engine: four explanation domains covering source and discovery, core assurance with fact qualification and DAI comparison, governance and operations, and verification outputs.
Source & Discovery gathers whichever authorized inputs the question requires; Core Assurance normalizes and qualifies facts and runs the evidence-plane and DAI comparisons; Govern & Operate covers scoped human decisions and connected operations; Verification packages bounded, provenance-linked outputs. Illustrative summary, not a mandatory pipeline.

How a conclusion is produced

1

Evidence qualification

Not every artifact counts. Provenance, source admission, relevance, coverage, freshness and contradictions decide what becomes qualified input. An unmarked record is never silently promoted into a typed fact.

2

Delegated authority

Policy, permission and actual authority to choose are different questions. A valid credential establishes what the credential evidence supports, not that a person delegated a specific consequential choice. The policy-versus-delegation comparison is bounded by the delegation evidence actually admitted.

3

Reachability

Exact agent, tool, handler and resource relations are reconstructed from supported source and connected evidence. Transitive paths and ambiguous joins stay explicit; a structurally present path is not a proven execution.

4

Deterministic evaluation

Identical frozen inputs, rules and versions produce the same conclusion. Reproducibility applies to the evaluated evidence set, not to changing external systems.

5

Proof boundaries

Every result declares what is established, partially supported, unassessed or unknown. Canonical reports and verification receipts bind to their exact scope and integrity contract; a receipt proves the artifact, not external enforcement.

An illustrative walk-through

A fictional support agent can call a refund endpoint. The source path is established; the credential may permit the call. Whether anyone delegated the choice of recipient is a separate evidence question. In this example that delegation evidence is absent, so the bounded conclusion is review, not a fabricated safe verdict.

ESTABLISHED in this fixture

Code-capable path

PARTIAL, scoped to the grant evidence

Credential scope

UNKNOWN, no delegation record admitted

Delegated choice

REVIEW, illustrative only

Bounded conclusion

Fictional, illustrative. Not an executed customer evaluation.

Connect HAIEC MCP

Ask your AI to run HAIEC AI Assurance

Requires a configured MCP client, scoped credentials and a read-only eligibility preflight before any authorized run.

What makes the method defensible?

  • Repository-first: implementation evidence can establish capability, not execution.
  • Runtime optional: runtime testing is authorized and scoped; it is not assumed.
  • Exact identity: joins require canonical source, scan, relation, and scope identity.
  • Historical basis: current source and evaluated assurance basis remain distinct.
  • Explicit frontiers: missing producers narrow the claim rather than increasing confidence.

What the method does not claim

HAIEC does not claim universal discovery, complete visibility, automatic safety, compliance certification, or proof that every possible delegation or runtime action occurred. Framework mapping is not certification. Unknown is not pass.

Understand AssuranceScope an Enterprise POC

Research foundation: explore ISAF v2, the instruction-to-consequence accountability framework.