Regulation is written for organizations. Software needs something a machine can check. ClearAgent explores the gap between those two sentences. The public TypeScript implementation covers concrete audit and oversight mechanisms. A separate private LangGraph service explores agent orchestration and policy retrieval.
The problem with the obligations
The EU AI Act places requirements on autonomous systems — record-keeping, human oversight, traceability. As written, they describe what an organization must be able to demonstrate. They do not describe a function signature.
"Maintain adequate records" is not a test. You cannot write an assertion against it, and you cannot tell whether last week's change made it better or worse.
Picking the three that are addressable in software
Not every obligation is a software problem. Some are procedural and belong to a compliance team. Three of them are genuinely addressable in code, and the build was sequenced around those:
- Record-keeping — what happened, in a form that survives review
- Human oversight — where a person can intervene, and whether they could have
- Traceability — why a given action was taken, reconstructable after the fact
What the public implementation checks
The public TypeScript code implements transaction verification, SHA-256 content hashes over canonical JSON, and policy-based triggers for human review. Those mechanisms can help expose altered records and route specified decisions to a person. They are not a validated compliance score or a substitute for organizational review.
Retrieval precision over breadth
In the separate private LangGraph work, policy lookup pairs vector embedding search with cross-encoder reranking. The design goal was deliberately narrow: return the one governing clause for an agent action, not a ranked list of ten plausible ones.
A list of plausible clauses puts the judgment back on the human and makes the audit trail weaker, because "the system suggested these ten and someone picked one" is not provenance.
The ledger
The public implementation records verifiable transaction data. Canonical JSON and a SHA-256 content hash make a changed record detectable when it is checked against its original hash.
The point is inspectability. Agent behavior is otherwise opaque — you have a prompt and a result, with little account of what happened between them. The recorded artifacts make a narrower, testable claim: a reviewer can inspect a transaction and detect specified changes to its content.
Scope
This is an engineering project, not a product. It has no users. The interesting work was architectural: deciding what a system can actually record, verify, and send for review.
