Build a claims preparation assistant, from source records to release evidence.

Part 2 of 10 · Business baseline

Measure today’s work before anything changes.

In the current workflow, an examiner opens assignment, claim, policy and document systems plus the procedure library, resolves the parties and applicable dates, assembles a packet, and corrects it after review. Missing or contradictory facts go to the source owner. Waiting and correction count as part of the work.

Baseline method: compare accepted work, including exceptions.

For a real baseline, log every eligible team A preparation over two ordinary weeks, stratified by complexity. Start the clock when preparation is requested; stop when the examiner accepts a packet with supported facts, the applicable procedure, citations and explicit unresolved items. Include waiting, corrections and manual fallback. Count each effort minute once; the exception bucket includes fallback effort not counted elsewhere. Record exclusions and reasons before comparison.

The completed teaching record assumes 40 eligible cases in 5–16 October 2026. A separate timed replay uses the same case facts, balanced examiner order and independent review for the target workflow. It includes four exception cases completed manually. These invented values show the calculation; they are not observations or forecasts. The timed replay is separate from the repeat-trial release tests in part 8.

Preparation is faster in this invented replay; the decision outcome is unchanged. No better coverage, payment or claimant outcome is established. Compare case mix, queue age and reviewer staffing in the monitored rollout before drawing a benefit conclusion.

Part 4 of 10 · Prerequisites

What must exist before the first line of code.

This dependency register is part of the release boundary. Each row needs an owner, an accepted commitment and failure evidence. All states below are synthetic assumptions.

The 10-minute age limit is an invented design choice, not an insurance standard. A recent timestamp alone cannot establish current truth; the current-version check and source change handling must work too.

Part 6 of 10 · Implementation path

From request to cited summary.

1. Verify the caller. Identity and assignment are checked before anything is retrieved.
2. Resolve the context. Claim, roles, governing date and procedure version are fixed in a request record. Missing or contradictory values stop the request or ask the examiner to clarify.
3. Retrieve within entitlement. Sources are filtered by the examiner’s permissions before the model sees them.
4. Prepare. A fixed sequence of approved read tools assembles status, events within this case, parties and procedure. The assistant cites each source with its as-of time. Cross-case search and note tools are unavailable in runtime-v2.
5. Present for review. The examiner sees the summary, the citations and anything withheld, with the reason.
6. Record. The request record keeps the requester, sources, versions, policy decisions and result for later reconstruction.

The model sees a narrow context view. Credentials and the full protected record stay in the execution environment.

Part 8 of 10 · Acceptance evidence

The evidence that releases part of the scope and holds the rest.

This is a completed synthetic evidence record, not a test report from a running service. Thresholds were chosen for this example before its invented results. They are not universal release thresholds.

Candidate 1.0.0 fails: one of three trials on a same-employer fixture returns an unrelated employee’s claim. The first failing component is cross-case party resolution; the severity is critical. Hold the entire candidate. Excluding the feature in prose cannot make that candidate releasable.

For candidate 1.1.0, the technical owner removes the prior-claims route, credentials, index and cached outputs from the serving path. Independent review checks that direct, indirect and cached cross-case requests fail closed. Only then does a fresh held-out evaluation support the narrower decision.

evaluations/claims-context-v2

The synthetic retest uses 40 new, examiner-adjudicated fixtures: 30 permitted and 10 boundary/failure cases, each tried three times (120 trials). The boundary set extends F-01–F-04 with cross-case route, cache and dependency-outage checks; F-05 is the permitted pattern. These fixtures are distinct from 20 development examples and the timed replay. Expected outcomes belong to independent examiners, not the generating path. Repeat trials expose variation; this small constructed set cannot estimate rare-event safety.

Evidence index

evidence/claims-release-v2 binds the version record, BASE-01 timing and effort ledger, DEP-01 dependency acceptance, REQ-C-1042-01 allowed trace, REQ-F01–F04 clarification, correction, denial and stop records, the evaluation above, the failed 1.0.0 trial, isolation retest and DRILL-02. These names refer to the teaching records on this page; no hidden logs, executed test files or real signatures are implied.

Part 10 of 10 · Economics and next commitment

What it costs, what it returns, and what you decide next.

Next commitment: fund the bounded observation period, not team B or prior-claims history. At the review, proceed only with observed quality, manageable effort and a credible service benefit; reshape if exceptions consume the capacity; stop if the operating case fails. Internal-note drafting or writing requires a separate scope and decision.

Read next:

• The 90-Day First Governed Capability
• Context Before Query
• Agent Onboarding
• AI Token Economics