AI Evidence Chains

Linked bundles for sequential and cumulative AI decisions: recommendations leading to actions, escalations leading to outcomes, agents that build on prior steps.

Most AI decisions are not isolated events. An AI recommendation leads to a human approval which leads to an action. An AI flag leads to an escalation which leads to a determination. An AI agent's step leads to the next step which leads to the outcome. H33 produces evidence chains: cryptographically linked bundles that document the full sequence of related decisions from origin to outcome.

Why single-decision evidence is insufficient

A typical regulated AI workflow has more than one AI decision. Each decision feeds the next. The audit demand is for the full chain. Examples: insurance underwriting (AI screens application, underwriter reviews, accept/modify/reject triggers issuance); healthcare triage (AI assigns priority, nurse confirms, escalation triggers physician review, decision triggers treatment); federal benefits determination (AI screens for eligibility, caseworker reviews, supervisor approves); trading and risk (AI flags trade, analyst reviews, manager approves, decision triggers actions); agentic workflows (each agent step depends on the prior step's output). In each case, the audit interest is the chain. Per-decision evidence is necessary but insufficient.

How H33 chains evidence

Each decision in the sequence produces an evidence bundle. Each bundle references the prior bundle's commitment via the PipelineDag EC object's prior_bundle_digest field. The chain is a hash-linked list from the originating decision to the outcome. The link is cryptographic. Modifying any bundle — or removing a bundle from the middle — breaks the chain at the modification point. The verifier detects the break and reports which bundle is the first invalid one.

The structure of a chain

An evidence chain is more than a list. Origin: the first bundle has no prior; documents the originating event. Step: each subsequent bundle references the prior bundle by digest. Branch: a decision can trigger multiple parallel sub-decisions; each branch has its own chain. Merge: a decision can depend on multiple prior decisions; the bundle references all priors whose outputs informed it. Outcome: the final bundle documents the workflow's result. The chain can be arbitrarily deep and arbitrarily branched.

Use cases

Insurance application chain. A customer submits a policy application. An AI underwriter scores (bundle 1). A human underwriter reviews and confirms (bundle 2, referencing bundle 1). The policy is issued (bundle 3). Months later, a claim is filed. The carrier traces the chain back to the issuance, the underwriting confirmation, and the original AI score. Clinical decision chain. A patient arrives at the ED. An AI triage system assigns priority. A nurse confirms. A physician orders imaging. The imaging AI flags an abnormality. The physician decides on treatment. A subsequent medical-legal review traces the chain. Federal eligibility chain. An applicant applies. An AI determines preliminary eligibility. A caseworker reviews. A supervisor approves. The determination is issued. An appeal traces the chain back to each decision's basis. Agentic workflow chain. A research-assistant agent plans a workflow. Each step retrieves, analyzes, synthesizes, and produces. A subsequent citation challenge traces the chain.

Chain verification

The verifier handles chains. Each bundle is individually verified (signatures, EC objects, optional anchor). Each bundle's prior reference is verified against the predecessor's commitment. Branching and merging structures are verified. The chain's connectivity is verified end to end. The verifier produces a per-bundle verdict and a chain-level verdict. The chain is verifiable in parallel for performance.

Common questions

What happens if a bundle in the chain is missing?
The verifier reports the missing bundle. Subsequent bundles cannot fully verify. The chain has a known break point.

Can chains span vendors?
Yes. Each bundle is self-contained. A workflow that crosses vendors produces a chain where some bundles came from Vendor A and others from Vendor B.

Can chains span time?
Yes. The chain has no time limit.

What's the storage overhead of a long chain?
Each bundle is tens of kilobytes; chains of thousands of bundles are gigabytes. For very long chains, Merkle-tree summarization can reduce storage.

Can a chain be redacted?
Selective disclosure is supported. A bundle in the chain can be redacted to a digest-only reference.

Get Started

Run the demo Download the verifier Download a bundle