# Claims-Denial Practice Case — Annotated Strong Answer

This is one defensible reasoning path through the fictional practice case, not
anonymous ground truth. Alternative answers are acceptable when they preserve
the evidence classes, authority, limitations, and decision thresholds in the
source pack.

Complete your own first pass before using this calibration aid.

## 1. Field view

Strong claim:

> In two observed NorthLake cases, an authenticated specialist inspected source
> evidence, changed or completed the proposed correction, approved submission,
> and relied on claims-system readback. The two cases do not establish frequency.

Why it is strong: it states occurrence, scope, and limitation. It does not turn
Maya's “every modifier correction” statement into measured frequency.

Next field move: assign the policy owner to confirm revision 8 applies to the
pilot service boundary and observe or query additional NorthLake modifier cases.
Do not begin Summit design from a NorthLake exact-build test.

## 2. Material reframe

Inherited claim: the pilot automatically resubmits eligible corrected claims for
NorthLake and Summit.

Contesting evidence:

- S4 requires claim- and revision-scoped specialist approval and readback.
- S3 shows specialist judgment changed both reviewed corrections.
- S5 shows a material NorthLake modifier defect and duplicate behavior.

Proposed reframe:

> Phase one is limited to the evidence-supported NorthLake cohort. The service may
> normalize, retrieve cited evidence, and recommend a correction. An authenticated
> specialist retains approval and submission in the workbench. Summit and automatic
> submission remain excluded until separately evidenced and authorized.

Safe fallback: continue existing specialist correction and submission; use the
service in observe/recommend mode only after duplicate recommendation is fixed.

Disposition: **continue discovery** until policy applicability, metric verification,
label authority, and the duplicate failure are resolved. A later **bounded kickoff**
may authorize the recommendation-only slice. Product Security cannot accept the
business outcome; the Revenue Cycle VP cannot waive the technical effect boundary
alone.

## 3. Outcome and economics

- Decision event: next-phase funding recommendation on 20 Aug.
- Eligible workflow: NorthLake inpatient medical-necessity denials meeting the
  documented July query and having permitted source access.
- Exclusions: Summit, automatic submission, missing required evidence without
  manual routing, unsupported payer behavior.
- Baseline: 42-minute median receipt-to-reviewed-correction for the July cohort.
- Sold acceptance target: 30 percent average handling-time improvement under S1,
  without an attached baseline or metric contract. Proposed interim four-week
  continuation floor: at least 15 percent reduction in median
  receipt-to-reviewed-correction time and at least 70 percent adoption. The measures
  are not interchangeable, and the interim floor does not amend the sold target. Both use
  zero unauthorized submissions, no error increase under independently adjudicated
  labels, and full cost below CAD 18 per eligible case.
- Counterfactual: existing specialist-led correction plus deterministic repair of
  the NorthLake adapter. The AI proposal must beat that smaller intervention.

Blocking gates:

- **Data:** correction-label authority and adjudication are unknown; 18 records
  lack linked notes.
- **Process:** specialist workflow is observed, but representative frequency and
  Summit workflow are unknown.
- **Governance:** policy applies on its face, but the policy owner must confirm the
  pilot service boundary.
- **Accountable capability:** release and operating owners exist; independent metric
  verification is unassigned.

Why a generic “data readiness score” would be weak: the pack contains data, but the
decision-bound labels, missing-note behavior, and source authority are unresolved.

## 4. Smallest mechanism

- Deterministic rules: denial-code normalization and policy invariants.
- Scoped retrieval: encounter and payer-policy passages with revision and purpose.
- Foundation-model call: bounded correction recommendation and explanation.
- Human review: correction approval and submission.
- No bounded agent is justified; the route does not require autonomous planning or
  tool selection.

Invalidation triggers: deterministic coverage becomes sufficient, citation or
correction quality falls below threshold, review capacity saturates, cost exceeds
the ceiling, policy changes, or adoption remains below the decision threshold.

## 5. Data, state, and dependency design

Canonical records: case revision, evidence-bundle revision, recommendation revision,
specialist disposition, submission approval, and claims-system readback. Model output
is a proposal, not canonical approval.

Critical typed relationships:

| From | Relationship | To | Consequence |
| --- | --- | --- | --- |
| S4 policy r8 | constrains | submission capability | Recheck at effect boundary |
| S5 build test | contests | NorthLake adapter assumption | Route remediation and regression review |
| Evidence bundle rN | supports | recommendation rN | Stale evidence invalidates only linked recommendation |
| Reframe r1 | affects | scope, architecture, eval, acceptance | Preserve unrelated artifacts |

Data/context contract: purpose-bound NorthLake sources, named source and label
authority, preparation lineage, revision/freshness checks, restricted-field handling,
retention, and manual routing when required context is missing or contested.

## 6. Evaluation argument

`rec-0.4.2` does not support pilot release yet. Protected failures cannot be
averaged into the strong citation, security, latency, or estimated-cost results.

Required before reconsideration:

1. Fix duplicate recommendation behavior and prove idempotency.
2. Restore the readback simulator and exercise timeout recovery.
3. Exercise disable and queued-work behavior.
4. Assign independent label authority and adjudicate disputed cases.
5. Bind the production tenant policy to the exact candidate.
6. Add missing-note behavior and relevant non-English adversarial coverage where
   the target population requires it.

False-green check: intentionally break an evidence citation, replay the same case
revision, stale the policy binding, and fail readback; the suite must reject the
candidate or prevent a completion claim.

## 7. Trusted action boundary

The candidate has no submission capability, which is the strongest current safety
boundary. Recommendation writes still need tenant and case scope, bounded inputs,
an idempotency key, concurrency preconditions, audit fields, and workbench readback.
If submission is ever proposed, it requires a separate contract and fresh specialist
approval scoped to claim, correction revision, actor, facility, and policy revision.

## 8. Readiness decision

Decision for `rec-0.4.2`: **hold**. The duplicate write, inconclusive recovery path,
unexercised rollback/disable behavior, unassigned label authority, and unbound
production policy are release-relevant gaps. A corrected exact candidate requires
new bound evidence; prior passing results cannot silently transfer.

After revealing S8, review `rec-0.4.3` as a distinct candidate. Its bound evidence
closes the protected failures above, and the Revenue Cycle VP, Product Security
Director, and Platform Lead each record only the disposition within their scope.
Decision for `rec-0.4.3`: **constrained launch** for the recommendation-only
NorthLake canary. It does not authorize Summit, automatic submission, or claim
acceptance against the sold 30 percent outcome.

## 9. Service review

Four-week decision for `rec-0.4.3`: **constrain**. Do not scale volume. Citation
acceptance is promising, but the small cohort does not establish a correction-error
trend. Adoption is 46 percent, duplicate entry consumes most of the time gain, and
the unresolved support incident keeps full economics incomplete. Fix the evidence
handoff, include the incident cost, and rerun adoption, correction-quality, and
cycle-time measurement. The 14.3 percent median improvement also misses the
predefined 15 percent continuation floor and remains well below the sold 30 percent
acceptance target.

## 10. Transfer and learning

For the constrained `rec-0.4.3` service, the receiving team has exercised support,
a bounded change, stale-policy response, disable, staging rollback, and queue
reconciliation. Transfer is not complete:
retirement/data disposition remains tabletop, and temporary delivery access is
still active. Exit should remain conditional on those exercises or an explicitly
accepted residual support boundary.

Sanitized reusable learning: payer-specific inherited adapters need versioned policy
exceptions and regression evidence. Do not productize the customer, payer, case, or
operator details.

## Common weak answers

- Treating the SOW, sponsor statement, observed cases, policy, and test report as
  equivalent truth.
- Claiming two cases prove the normal frequency of specialist intervention.
- Fixing the adapter and assuming the automatic-submission promise is then allowed.
- Scaling because citation quality is high while adoption and workflow cost miss
  their thresholds.
- Calling a simulation, document handoff, or model-generated status customer or
  production authority.