Revenue Recovery · Governed Decision Layer

The governance layer for your AI-assisted recovery stack.

Every decision (lane fit, readiness, next action, missing evidence) comes from a signed rule artifact, not a prompt. The same input produces the same answer every run, and every change is previewable against your current inventory.

Runs beside your existing AI. Does not replace the workbench.
Rust + Wasm artifacts. Same binary runs in browser, server, or edge.
Every call is source-grounded, versioned, and signed.
ACCOUNTA-108422
Ready to Act
Lane
Payment Variance
St. Anne Specialty Hospital · EverPeak Health
High priority
Estimated recovery
$7,360
Balance at risk
$22,640
Recommended action
Submit underpayment appealtrace · payment_variance.v2026.04
Checks passed
✓Remittance present
✓Contract terms present
✓Expected allowed modeled
✓Variance threshold met
rulevariance_threshold_met
Built
from your data
Recovery lanes, yours

We compile each pack from your labeled historical decisions, so no SME hours are required on your side. The demo carries five representative lanes (denials, payment variance, NSA, VA/TRICARE, workers’ comp / MVA) to show the shape. Production lanes are scoped against your actual recovery workflow, and knowledge transfer (including an onsite if you want one) ships with the artifact.

~0.3
ms / eval
Wasm decision latency

Compiled from Rust. Runs locally in the browser for the demo, on the server in production. Same artifact, same answer.

100%
traceable
Every call is signed

Every decision returns a rule trace, the exact fired bits, and the source-grounded reason. No black-box recommendations.
02 · System at a glance
Documents flow through your existing AI. LogicPearl verdicts gate the action. Your workbench stays the workbench.
Wasm · signed · versioned
SOURCEAccount packet835 · notes · portals · call logsunchanged intake pipelineingestYOUR AIExtraction · RAG · LLMtyped features + confidencesmodels stay exactly where they arefeaturesLOGICPEARL · DECISION LAYERpearl.wasm + rulesallow · fired bits · counterfactualsdeterministic · ~0.3 ms / evalversioned · signed · diffableverdict + traceWORKBENCHYour UIaction + governance_tracesame specialist workflowaudit replay: same artifact + features ⇒ same verdict
03 · What changes

Your reality could look like this.

A governed decision layer does not replace your team. It handles the repetitive readiness check so their judgment lands on accounts that can actually move.

Work the right inventory

Accounts are routed into the right recovery lane before specialist time is spent. No queue triage, no duplicate review.

Catch blockers early

Missing contract support, notification proof, carrier details, or benchmark evidence are surfaced immediately, not mid-appeal.

Standardize specialist judgment

The same account gets the same readiness decision every time, with a clear reason and a rule trace you can audit.
04 · How it works

Four steps from account packet to next action.

Everything LogicPearl returns is tied to a rule you can inspect and a source document it read.

01

Read the account packet

LogicPearl ingests normalized facts from your current systems, documents, and workflow tools. No ETL project.

02

Select the right lane

It determines whether the account belongs in denials, payment variance, NSA, VA/TRICARE, or workers’ comp / MVA.

03

Check readiness

It evaluates whether the evidence and timing support action now, or whether the account is blocked on specific support.

04

Return the decision

Your team gets status, missing items, next action, and a traceable reason, back in the workbench they already use.

Sample resultaccount · REV-66776 · VA Community Care
Status
Needs More Support
Missing item
72-hour notification log
+ confirmation ID
Next action
Resolve notification proof before filing
05 · Live demo · real LogicPearl Wasm

Click an account. Watch the decision change.

Every account below is evaluated through the actual compiled LogicPearl rule artifact for its recovery lane, running as WebAssembly in your browser. Flip the evidence and the answer really changes.

logicpearl.ops / recovery / inventory
Inventory0 accts
Loading inventory…
Loading…
Loading compiled lane artifact…
Live Wasm eval · —trace · —
Engine I/Othe bytes your code would actually send and receive
Select an account to inspect the engine I/O.
Integration · import { loadArtifact } from '@logicpearl/browser'; const a = await loadArtifact(url); a.evaluate(features)same ABI · Wasm / native / REST

Each lane artifact (pearl.wasm + metadata) is fetched once then evaluated locally. The same signed artifact ships into production via the Rust native binary, server-side Wasm, or REST. The decision does not drift.
06 · Recovery lanes

One decision format. Any recovery lane.

The five lanes below are representative shapes we use in the demo. Each asks a different business question and trips on a different blocker. Your production lanes are compiled the same way, against your actual workflow, and always return the answer in this shape.

Payment Variance
$500
minimum variance worth pursuing (demo default; set per customer)
Business question

Is this underpayment worth working, and do we have enough contract support to act?

Common blocker
!Current payer amendment missing
Output

Work / do-not-work signal, blocker, readiness trace

Denials
Deadline
payer-specific appeal filing window gated per account
Business question

Does the clinical and coding support an appeal inside the timely filing window?

Common blocker
!Missing admission H&P or progress notes
Output

Ready / needs-support status, missing evidence list, appeal path

NSA
30-day
CMS open-negotiation window (regulatory)
Business question

Are we inside the open-negotiation window, and is the benchmark evidence strong enough?

Common blocker
!Open-negotiation window expired
Output

Window status, QPA delta, do-not-work when closed

VA / TRICARE
72-hr
VA emergency-care notification rule (regulatory)
Business question

Can this account move now, or will missing 72-hour notification proof block payment?

Common blocker
!Notification proof missing
Output

Ready / blocked status, missing support item, compliance trace

Workers' Comp / MVA
Per-state
statute of limitations varies per jurisdiction
Business question

Is liability confirmed, and does the state fee schedule support recovery?

Common blocker
!Police report / liability confirmation missing
Output

Recovery signal, liability gap, fee-schedule delta
07 · Analyst experience

What your specialist actually sees.

Operational first. The readiness call, the missing item, the recommended action, all visible before any rule-engine detail. The machinery is one click away, not in the way.

  • Lane, status, priority: visible at a glance
  • Missing item called out by name, not by rule ID
  • Plain-English explanation, expandable to rule trace
  • Source documents linked for every fired check
workbench / account viewREV-66776
VA / TRICARE
VA case · Capital Regional Hospital
Needs More Support
Priority
High
Days remaining
2
Balance
$13,880
Missing item
72-hour notification log
Required: confirmation ID for VA Community Care submission.
Recommended action
Resolve notification proof before filing
×
Check failed: 72-hour notification proof
VA Community Care requires documented notification within 72 hours of admission with a carrier-confirmation ID. No confirmation ID was found in intake notes or the notification log.
rule · notification__within_72hv2026.04source: intake.pdf p.2
08 · Governance

You already run AI in recovery. LogicPearl is the deterministic layer beside it.

Modern recovery teams run RAG, document extraction, and LLM summarization across their inventory. That gets you more context per account. It does not get you a defensible, repeatable readiness decision. LogicPearl runs alongside those AI pipelines and contributes one governed, auditable answer per account.

Your recovery stack, with LogicPearl
LogicPearl decision layer
ready / blocked · missing items · next action · rule trace
governed
Your AI pipelines
RAG · document extraction · LLM summarization · embeddings
Your workbench & systems of record
queues · PAS · contract repo · 835 ingest · document store
AI-assisted todayRAG · extraction · summarization

You’ve invested in the models. The gap is a deterministic layer beside them.

  • Context, not decision
    LLM summarizes the account; the specialist still has to decide ready / blocked themselves.
  • Prompt drift between teams
    Two analysts ask the same model the same question and get two different framings of the answer.
  • No audit trail on the call
    You can see which chunks were retrieved, not which policy clause made the account un-workable today.
  • Payer / contract updates don’t propagate
    A new amendment lands; RAG picks it up eventually, but no one knows which accounts just changed status.
  • Hallucination risk on the recommendation
    The model will cheerfully suggest an appeal path the timely-filing window already closed on.
With LogicPearl alongsideGoverned decision layer

Same AI pipelines. One auditable, deterministic answer alongside them.

  • Decision, not just context
    Returns ready / needs-support / blocked with a named next action, not a paragraph the analyst has to interpret.
  • Deterministic policy checks
    Evidence gates and timing windows are rules, not prompts. Same account → same answer, every run.
  • Source-grounded rule trace
    Every call cites the clause, the document page, and the extracted field the model pulled it from.
  • Payer updates propagate on rollout
    Edit a rule, preview impact across current inventory, ship with a signed version bump.
  • Hallucination-bounded
    LLMs read and extract. Rules decide. The model cannot suggest an action outside a governed playbook.
Lane artifact changerevenue_ready_va_tricare
earlier → currentlive diff · signed
Loading real diff bundle…
source · CMS transmittal 20260218owner · ops.rules@logicpearlpreview on inventory · +412 / −0
08c · Architecture

Keep your AI stack. Add a governed decision partner beside it.

Today, an LLM reads policy chunks and guesses readiness. With LogicPearl working alongside your AI, the LLM does what it's best at (pulling typed fields out of documents) and the rule engine decides. Your extraction models keep their job; a deterministic partner handles the verdict. No retrieval hallucinations in the decision path.

01
Account docs
PDFs · 835s · notes
extract
02
LLM extractor
typed fields only
facts
03
LogicPearl rules
governed · versioned · signed
decision
04
Ready / blocked + trace
one answer · source-grounded
LLM extraction contract (step 02)ts
type Extracted = {
  notification_confirmation_id: string | null;
  notification_timestamp: string | null; // ISO-8601
  contract_amendment_date: string | null;
  denial_code: string;
  clinical_documents_present: boolean[];
};
Compiled rule + evaluate (step 03)wasm
// A learned predicate, compiled from labeled decisions.
// This is one of 11 rules in the VA / TRICARE lane artifact.
const rule_008 = {
  id: "rule_008",
  when: { feature: "hours_until_72h_deadline", op: "<", value: 0.5 },
  action: "route_to_late_notification_fallback_review",
};

// One call at the decision boundary.
const verdict = lane.evaluate(facts);
// -> verdict.allow === false
// -> verdict.firedRules[0] === rule_008
// trace: va_tricare_notification_readiness v2026.04
What this buys you. The LLM never sees the policy. It can’t hallucinate a rule because it isn’t being asked to reason about rules. When CMS publishes a new transmittal, you update one rule version (not a vector index), and every account re-decides identically.
LLM stays in its lane
It reads documents and emits typed fields. No policy interpretation, no appeal recommendations, no free-text judgment.
Decision lookups are deterministic
When the rule engine needs a contract, payer, or CMS fact, it queries by key, not by embedding similarity. Your RAG stack is free to keep serving context to humans; the decision path reads exact records.
Swap models without re-validating
Replace GPT-4 with Claude, Gemini, or an in-house model. Rules don’t change. Decisions stay identical for the same extracted fields.
08b · Counterfactuals

Ask “what would have happened if …?” and get an answer.

Because every decision is a rule trace, not a prompt, LogicPearl can replay any account under different evidence, thresholds, or policy versions. No re-running the model. No guessing.

Account · REV-66776 · VA / TRICAREcounterfactual · live wasm

Flip the evidence. The engine re-decides.

Every toggle changes a real feature value that the compiled va_tricare_notification_readiness artifact reads. No simulation, no heuristic.

×72-hour notification confirmation present
✓Emergency clinical record present
✓Veteran identity fields complete
✓Other health insurance captured or N/A
Decision under this evidence
Needs More Support
Loading…
Gates passing
0 / 0
Est. recovery
$0
trace · …— mslive wasm
Before you ship a rule
Preview the impact on current inventory. See exactly which accounts change status before anyone touches them.
After an adverse outcome
Replay the account under the evidence you had at the time. Distinguish model error from missing evidence.
When a payer pushes back
Hand them a rule trace with counterfactuals: "if this 835 line had been reported correctly, the account would have been workable."
09 · Pilot

What would you want a pilot to prove?

There is no single pilot shape we push. Tell us what you would want to see demonstrated, whether that is a specific lane, a specific outcome, a specific payer, or a specific jurisdiction, and we will scope the narrowest test that would actually answer it.

Artifact readylowest data volume · deadline-sensitive

VA / TRICARE: 72-hour notification

Compliance-heavy lane. Value shows up on the first batch where notification proof is missing, and the 72-hour window makes misses expensive.

Pilot signal
Ready / blocked verdict per account, named missing support, recovered vs. lost $ comparison against your team’s prior action.
Paid engagement. De-identified data. You keep the artifact whether or not you continue.
Artifact readysmall-volume · high-dollar · deadline-driven

NSA: 30-day open-negotiation window

Small volume, high dollar. LogicPearl flags accounts inside the 30-day window with benchmark support ready, and do-not-work signals for windows already closed.

Pilot signal
Window status per account, QPA delta, benchmark readiness, benchmarked against historical outcomes.
Paid engagement. De-identified data. You keep the artifact whether or not you continue.
01

Scoping call

30 minutes. You walk us through one recovery lane that hurts today. We learn what data lives where.

02

Data-fit review

We look at a sample of de-identified accounts together and tell you honestly what we can and can’t decide from the fields you have.

03

Shadow run

LogicPearl runs silently alongside your team on real inventory. Readiness calls are logged, not acted on.

04

Compare & decide

Side-by-side: our readiness call vs. your team’s action vs. the actual recovery outcome. You decide whether to keep going.

What each lane needs from your data
We’ll confirm fit against your actual schema in the scoping call. These are the fields the rules read.
de-identified sample, read-only
Payment Variance
Fields we read
Remit lines, contract reference, and amendment effective dates.
Volume
We benchmark against whatever history you have: last month or last year, whichever you trust.
Denials
Fields we read
Denial codes, clinical document inventory, timely-filing dates, appeal outcome.
Volume
We’re looking for signal across outcomes, not a fixed account count.
NSA
Fields we read
QPA disclosure, negotiation-window timestamps, benchmark references.
Volume
Smaller volume lane; even a few dozen accounts can show whether windows are being missed.
VA / TRICARE
Fields we read
Referral auth, notification logs, confirmation IDs, payment disposition.
Volume
Compliance-heavy; value shows up on the first batch where notification proof is missing.
Workers' Comp / MVA
Fields we read
Acceptance letters, adjuster records, state fee schedule references, incident reports.
Volume
State-by-state, so we scope the pilot to states where you have volume.
10 · FAQ

FAQ

Beside it, not over it. Your AI keeps ingesting packets, running its extraction models, and recommending actions. LogicPearl takes each recommendation at the decision boundary and replaces the ambiguous judgment (“is this account ready to act?”) with a deterministic verdict (same input, same answer, every run) and emits the signed governance trace alongside it. Your existing investment keeps compounding; the decision itself becomes defensible.