The regulatory reasoning, end to end.

A California WC claim is a chain of statutes, and most systems treat each link as its own module with its own data entry. Chiron treats the claim as one working record from intake to closure: eligibility and §3600 classification and MPN routing, AOE/COE causation, MTUS/ACOEM-aligned treatment, UR and the RFA under §4610, IMR under §4610.5, the P&S/MMI determination, impairment and apportionment under §§4663–4664, the QME/AME process under §4062, billing and the pay clock under §4603.2, and the defensible record through settlement or closure under §5410.

Because it is one record, the causation finding a physician wrote in month one is still there, unedited, when the impairment rating is computed in month eight — and the reasoning that produced it is still inspectable. One Digital Employee, one working record, one signature standard, across 40 backend modules running in production today.

Every specialist in the Round Table runs Chain-of-Verification on its own output: an initial answer, then a set of adversarial verification questions — did I invent anything not in the record? What would opposing counsel argue against this? Would this survive cross-examination? — and only then a revised final answer. The self-critique happens before the draft ever reaches the physician’s screen.

What the physician sees while this runs is a live, honest reasoning trace: which specialist is working, which guideline it retrieved, which sources it found — a structured account of what happened, not raw model output and not a progress bar faked to look busy.

Defend This takes the same discipline further on request: the platform argues against its own determination first, in the voice of the defense QME, the UR reviewer, or the carrier, and assembles what it can into a submit-ready defense packet — honest about what it can’t yet answer.

The must-review-before-final gate is enforced at the data layer, server-side — not a banner a browser could skip. The physician sees the visible reasoning that produced the draft, not just the draft itself, can edit any clause, and only then attests. That review-and-attest step is what AB 3030 / SB 1120 disclosure and AB 489 require, and here it is architecture, not a policy memo: there is no path from an AI draft to the chart, a payer, or an attorney that skips the physician.

The architecture also errs toward the person the process is supposed to protect: if the IMR-eligibility classifier can’t reach a confident answer, it fails open rather than closed, so a worker’s appeal window is never lost to a classification error.

At the moment of attestation, a SHA-256 hash binds the physician’s signature to the exact rendered document — the specific PDF, the specific content, at that specific moment. Four public verification endpoints, one each for IMR, QME, PTP-QME, and closure documents, let anyone holding a copy confirm it hasn’t been altered since it was signed. No login, no PHI in the check, and the endpoints are rate-limited against abuse.

That means a payer, an opposing attorney, a QME, or the DWC itself can verify authenticity independently of ChironAI — the proof doesn’t depend on trusting the vendor that produced the document.