LogicPearl for utilization management

AI reads the packet. Policy owns the decision.

Messy prior auth packets become deterministic UM dispositions, with reviewed evidence and source receipts attached.

How it works

Observers read the packet. Pearls own the decision.

Loading pearl metadata...
01 Packet arrives

A provider sends a normal prior auth packet with clinical notes.

02 Evidence is extracted

Reviewed extractors emit evidence findings; LLM review is advisory only when deterministic checks do not match.

03 Policy decides

Compiled BCBSMA policy returns the disposition. Evidence the system cannot prove routes to Physician Review — it never guesses a denial.

04 Review improves extraction

A physician reviews a judgment once; it becomes a deterministic rule that scales across every future packet.

Why this matters

Typical AI agent Tops out near 95% — and can't tell you which 5% it got wrong
LogicPearl-controlled agent Same input, same verdict, every time. Uncertainty routes to a physician — never to a denial

Potential CMS / BCBSMA rule differences

Possible coverage-rule gaps, queued for reviewer confirmation.

Loading rule-difference summary...

Point-in-time replay

Replay any decision under the policy snapshot in force when it was made.

Loading BCBSMA source diff...
Input case Loading case
Packet rewrite Messiness 0 / 10, 0 factors
Raw packet JSON
Optional: edit raw JSON, then apply to rerun.

Packet rewrite

Customize the LLM rewrite

Adjust how messy the clinical packet becomes before the deterministic policy run.

0 / 10
Rewrite factors Slider selects defaults. Click any factor to override.

Choose a case, set messiness, then rewrite the live input.
Case facts

System verdict

Ready Pick a case to begin

Choose a synthetic case on the left, or press Play the golden path on the Dashboard.

HL7 FHIR Da Vinci DTR Standard FHIR vector in, deterministic policy out

Evidence review

See what matched, what did not, and what needs review.

LLMs can recommend whether messy text should count. Reviewers approve or reject before a dataset row changes the deterministic extractor.

Trust receipt

A source-backed disposition, not an AI opinion.

Loading rule bundle...

Controlled improvement

When the case needs review, the system learns cleanly.

Two learning lanes, both gated by a physician, neither touching BCBSMA policy text. Lane 1 teaches the evidence extractor which clinical wording counts. Lane 2 separately teaches the case adjudicator which decision to record.

Lane 1 · Evidence wording Teach the extractor which wording counts as evidence
Lane 2 · Case adjudication Teach the case adjudicator which decision to record

Audit trail

Defend the decision as it happened.

When someone asks why a case was denied six months ago, answer with the rules in force at the time — not the rules you wrote last week. Start with the receipt: disposition, source snapshot, artifact, packet, bitmask, and replay behavior. Tool traces stay below as supporting evidence.

Execution summary

What facts entered the pearl, and what rule fired.

Execution trace Conversation, tool calls, bitmask board
Conversation Real agent run
Agent tools Waiting
    Workflow action return_um_disposition(packet)
    Policy pearl Submission readiness
    Waiting No packet evaluated yet
    Evidence internals Inspect raw observer output and emitted pearl inputs
    Review loop Show packet, pearl, and physician update path

    BCBSMA rule library

    Every coverage rule, compiled from BCBSMA policy.

    Human-readable rules across the full BCBSMA policy set — each one traceable to its source PDF. Search below, or expand any rule to see the compiled predicate behind it.

    Source corpus Synthetic test cases and BCBSMA source PDFs

    Corpus

    BCBSMA policy corpus

    Loading corpus...
    Test cases 0 cases
    Source PDFs 0 PDFs