An AI-tutored subrogation training program that teaches claims representatives through short lessons, knowledge checks, and realistic accept, negotiate, or deny simulations, graded to the same standard as a real file.
Progression is enforced in code, not left to the model: only one function in the codebase can mark a level complete.
Each level pairs its lessons with a knowledge check drawn from a rotating set of three curated question sets, then a hands-on simulation. A level unlocks only after the one before it is actually completed.
What subrogation is, inbound vs. outbound direction, reading an inbound demand package, deductible reconciliation, liability and comparative negligence basics, and the accept, negotiate, or deny decision. Closes with a knowledge check and an inbound-demand simulation.
Building a liability percentage from the specific duties breached, one lesson per common fact pattern (failure to yield, improper turns, following too closely, proper lookout), plus deductible reimbursement. Closes with a knowledge check and a mock claims-system simulation.
The actionable responses to a demand, supporting a position with evidence, and qualifying related claims. Closes with a knowledge check and a demand-exchange negotiation simulation: accept, counter, deny, or request-for-info.
What a repairable or total-loss settlement must prove, salvage and deductible recovery, and assembling a demand packet correctly. Closes with a knowledge check and two demand-center simulations: repairable and total-loss.
Each level closes with a simulation built around the same kind of decision a representative makes on a real file, then grades the actual output, not just a selected answer.
A synthetic inbound subrogation demand arrives with a full review package.
An accept, negotiate, or deny decision, with a written rationale.Graded against a versioned rubric; low grader confidence routes to human review instead of an automatic pass.A mock claims-system workspace: worklist, loss details, policy, financials, contacts, notes, and documents.
A graded file note reflecting the assigned claim.Same rubric-graded, human-review-on-uncertainty standard as every other simulator.A synthetic demand arrives through a demand-exchange-style negotiation interface.
Accept, counter, deny, or request-for-info, backed by supporting evidence.Graded on the response chosen and the evidence cited for it.Assemble an outbound demand packet: supporting documents, damages worksheet, and salvage credit.
A published demand packet plus a diary follow-up entry.Two full runs: one repairable settlement, one total-loss settlement.Two Gemini models run on Vertex AI, one tutors, one grades, with the grader held at a temperature of 0.0 so grading is as repeatable as an LLM call can be. Neither one decides whether a trainee has met a prerequisite or completed a level, only one function in the codebase is allowed to change that state, and prerequisite gating is enforced as a database constraint with no admin bypass. If the grader's own confidence is low, the attempt is routed to human review instead of an automatic pass.
| Activity type | Passing threshold | Attempts |
|---|---|---|
| Knowledge check | 80% | 3 max |
| Simulation | 75% | Unlimited |
Every knowledge check and simulation is graded against a corpus of state statutes and industry reference manuals, indexed for retrieval. A failed attempt produces a cited explanation of what was missed and why, and acknowledging that remediation is what unlocks the retry.
The program runs on the same infrastructure discipline as the rest of the suite, kept entirely separate from live claim work.
Every scenario, demand letter, and claim file is fabricated for training. Every generated document carries the literal watermark SYNTHETIC TRAINING DOCUMENT — NOT A REAL CLAIM.
The program never sends real communications to a carrier and never makes a production claim decision. It is a training system, not a claim-handling system.
Trainees sign in through their own isolated Firebase Identity Platform tenant, behind an invitation allowlist that fails closed, with email verification required.
Every attempt is append-only, never overwritten, and every session is logged as an audit event.
One lesson in Level 1 is titled “What Varies by Your Own Carrier”, and names the topics it deliberately leaves out, because they belong to your own written procedures, not a generic course. We take that as the spec: as part of onboarding, we work with your team to author those topics into your own edition of the program.
Tell us about your team, and we'll follow up to scope a carrier edition.
Deterministic progression, grounded grading, and synthetic-only content, with a curriculum that can be adapted to your own carrier's rules.