Everything ChironAI OM does, in production.

Chiron leads every case. Issac assesses utilization-review denial risk on every case, no exception. Justine convenes when apportionment or mixed causation is in play; Eli convenes on billing disputes; Theo opens automatically the moment Chiron’s confidence drops below 70%. Runs in parallel, degrades gracefully, and is visible on screen — live, not a spinner.

A compound, five-component reasoning architecture — a classifier, a domain-tuned clinical reasoner, and frontier reasoning slots working in concert. Not a single model behind a prompt.

Every specialist self-critiques before a physician ever sees the output: initial finding, then a verification pass — did I invent anything not in the record? what would opposing counsel argue? would this survive cross-examination? — then a final answer. Three passes, one determination.

A live, step-by-step trace of what the system is actually doing: which specialist is running, which guideline it just retrieved, which source it just found. Engineered to never fake a progress bar; the chain-of-thought itself never leaks to the transcript.

Ask the record anything; the answer is grounded only in the case file and the guideline corpus, with inline citations. A hard code guardrail deletes any citation the reasoning didn’t actually retrieve, backed by an ICD-10 hallucination guard and apportionment-citation validation.

The system argues against its own determination first — standing in as the defense QME, the UR reviewer, and the carrier. It tallies what it can answer, what it can’t yet, and where the exposure sits, then assembles a submit-ready defense packet.

Physician review/override endpoints sit on every determination. Deterministic engines override the reasoning layer on the actual numbers. The IMR-eligibility classifier fails open, so a worker’s appeal window is never lost to a model error.

Arising-out-of and course-of-employment narratives under Labor Code §3600, drafted from the full record and versioned from first draft to final.

The apportionment offset for a prior award of permanent disability runs as a deterministic engine built on the Brodie/Strait three-prong test: exact, reproducible, versioned, and self-documenting.

Whole Person Impairment under the AMA Guides, 5th Edition, with DRE/ROM method selection and deterministic AMA Combined Values math, feeding a 2005 PDRS calculation — FEC rank, age adjustment, occupation adjustment, final PD. The age/occupation adjustment currently runs a simplified table pending the full DWC schedule load.

Every apportionment and impairment narrative is anchored to the controlling authority: Escobedo, Brodie, Strait, Hikida, Almaraz/Guzman.

Instant, deterministic dollar-impact modeling on the case’s own AI-derived findings plus the Labor Code constants that actually govern the number — $870 per permanent-disability point, the SB 863 §4660.1 1.4 adjustment factor. Math over facts already in the record, not a live-AI simulation.

Every statement in a causation analysis traces back to the specific field in the record that produced it. The reasoning path is reconstructable years later for deposition.

Every treatment plan is checked against MTUS (California) and ACOEM (national) guidelines at the point of care — retrieved from two dozen authentic California DWC guideline documents, not a summary. Deviation surfaces with the citation and the rationale.

Before a request for authorization goes out, the system scores its own denial probability and the downstream IMR-appeal risk, and names the evidence that needs to be attached for the request to survive utilization review the first time.

Every request for authorization tracked from submission through UR determination, with the timeline visible end to end.

Flags requests that qualify for automatic authorization under the statute rather than routing into a full utilization-review cycle.

The specific gaps behind a UR Survival Score — a missing exam finding, a guideline citation not yet on file, treatment outside the MTUS window — surfaced inline, not buried in a report.

Every UR denial is classified against Dubon II and Bodam before it’s routed, because that classification is what gates IMR eligibility versus a straight appeal.

The §4610.5 independent medical review packet assembled from the case record, timed against the 30-plus-5-day mail clock.

Every IMR and appeal deadline computed against the California business-day and holiday calendar. Day-count specifics on QME-adjacent deadlines and record-retention windows remain under counsel review, so the system flags rather than finalizes those.

The statute’s twelve required sections drafted and checked against a compliance checklist before the report leaves the practice.

The qualified-medical-evaluator and agreed-medical-evaluator panel process runs as a state machine, including the Vázquez good-cause generator, which produces a byte-stable filing exhibit when a party has good cause to request a new panel.

The treating physician’s determination is reconciled line by line against the QME or AME findings, with disagreement surfaced rather than silently overwritten.

Doctor’s First Report of Occupational Injury or Illness, drafted against the 5-business-day statutory deadline.

Auto-drafted on the 45-day cadence or the moment a material change triggers one; the physician edits every section before signature.

Permanent-and-stationary determinations triggered by the clinical milestone, not by elapsed time, deadline-tracked and physician-attested.

Every statutory clock in the claim encoded and holiday-aware against the California calendar. A daily 07:00 cron surfaces every approaching and overdue report across the practice.

Every AI-drafted report carries the generative-AI disclosure required by AB 3030 — non-dismissible by design, and it persists through the signed PDF.

Official Medical Fee Schedule coding applied to every billable encounter.

Page-count tiering for medical-legal reports computed under ML204 rather than left to a biller’s judgment call.

The 45-business-day payment rule tracked against every bill; penalty and interest computed automatically the moment the payer misses it.

Second-review and Independent Bill Review appeal packets assembled from the same billing record that generated the original claim.

Evaluation-and-management coding leveled against the CMS 2021 Medical Decision-Making rules, not the retired time-based table.

Every case classified — industrial, denied, delayed, first-aid, non-industrial — at intake, versioned as the record develops.

§4616 medical-provider-network verification, refreshed weekly from the California DWC’s public open-data feed, so a referral never routes to a network that lapsed.

Flags cases whose recovery is lagging the trajectory the treatment plan predicted, before the file goes stale.

An overnight-worked worklist waiting at the start of the day: what the system read, flagged, and drafted while the practice was closed.

The moment a new document lands in the file — a QME report, a UR determination, a lab result — the case is re-read and the relevant determinations refresh.

Case closure drafted with the five-year reopening window computed and tracked, so a case never closes blind to its own reopen date.

Four public, no-login, zero-PHI, rate-limited endpoints — IMR, QME, PTP-QME, closure. A SHA-256 hash binds the physician’s attestation to the exact rendered PDF, so anyone holding a document can confirm it hasn’t changed since signature.

Each clinic runs on its own separately provisioned database and its own tenant Blob storage containers, with Postgres row-level security enforced across 40 tables as a second layer, not the only one.

Every PHI access recorded automatically, by decorator, not by developer discipline — tamper-evident by construction.

Failed-authentication bursts, cross-tenant access attempts, and mass-export patterns are monitored and mapped to HIPAA’s breach-notification triggers.

The rate at which the signing physician agrees with Chiron’s determination is tracked over time as an operational trust metric for the practice — visible, not hidden from the people signing the reports.