AI walkthroughs

An AI request usually takes several turns. Eve chooses a tool, receives its result, and uses that result to choose the next step. A search returns a location to inspect. An exact read supplies the source passage. Eve can repeat this loop before writing an answer.

These condensed examples use synthetic questions and current tool names. They illustrate observable actions, not private model reasoning or a fixed script. Copilot and trial examples follow recorded synthetic runs. The workup example shows a common source-reading pattern. The treatment example includes a recorded failure so the missing step is visible.

Copilot: answer a question about a report

What did the breast biopsy show?

This question starts in a case's Copilot thread. The synthetic source contains a final pathology report; a separately confirmed chart summary is not required to answer what that report says.

grep({
  query: "breast biopsy pathology diagnosis histology receptor",
  paths: "source",
  limit: 8,
});
 
// The returned locator points to page 1 of the synthetic report.
read_file({
  path: "source/breast-biopsy.pdf",
  pages: 1,
  includeLayout: false,
});

The read supplies the report's text. In this fixture, it identifies invasive lobular carcinoma of the left breast. The answer reports that finding and its source; it does not change the chart. In an application run, the exact read can also return a citation alias for the answer. The standalone fixture trace did not record an alias, so none is invented here.

General Copilot can also answer without an attached case. Authorized patient and Library lookups use the app's MCP tools, including list_workspace and read_document. Their returned references determine what can be opened. See the evidence paths.

Workup: identify a missing fact that changes a decision

For this advanced bladder cancer case, which missing workup items could change treatment readiness?

The patient record establishes the documented disease setting. Eve then checks the relevant guideline branch and its original pages. A prior diagnosis draft can help with orientation, but its claims still need source support.

A missing test result remains unknown. The draft should explain which treatment decision depends on it without turning every unknown into a reason to delay all care. Page numbers come from the installed source release and the tool result; they are not fixed in this example.

Treatment: connect alternatives to their source

After this colon-cancer resection, which adjuvant options should I review, and what missing information could change the choice?

The synthetic evaluation uses fictional Cedar and Maple courses. Its one-page guideline lists both, leaves baseline neuropathy unresolved, and supplies no dose or schedule. These names describe test data, not clinical regimens.

grep({
  query: "resected stage III colon postoperative adjuvant treatment",
  paths: "guidelines",
  limit: 8,
});
 
// Page 1 belongs to the checked-in synthetic guideline fixture.
read_file({
  path: "guidelines/nccn/colon-eval/canonical.pdf",
  pages: 1,
  includeLayout: false,
});

The intended result keeps both source-listed alternatives, the neuropathy unknown, and the absence of dosing instructions. When a real source supports a specific course, a matching treatment template can clarify its administration. An exact product-label read can resolve a remaining label question. Neither lookup creates an alternative that the guideline did not support.

What the recorded run missed

In the synthetic run checked on September 10, 2026, Eve read the patient record but used later-page cursors on new guideline searches. The searches reported a match count but returned no item on those pages. Eve never opened the guideline PDF and omitted both fictional alternatives. The evaluation passed 9 of 11 checks and failed the exact-page and alternative-coverage checks.

This trace predates the pagination guard in grep.ts: a continuation past the available results now returns an error directing the caller to retry with cursor: null and offset: 1. Use those values for a new query. Continue only with a cursor returned for the same query and paths, then inspect the exact source before using its claims. The recorded run produced a draft without the guideline alternatives; it does not demonstrate a successful treatment flow or validate the later fix.

Trials: compare recorded criteria with patient facts

Which internal trials might fit this later-line kidney-cancer case, and which eligibility facts still need confirmation?

This synthetic case has metastatic clear cell renal cell carcinoma after several therapies, no actionable molecular alteration, and a stated wish to avoid hospitalization. Two pinned institutional trial documents supply the trial criteria.

grep({
  paths: "trials",
  query: "metastatic clear cell renal cell carcinoma later line",
  limit: 8,
});
 
// These paths belong to the checked-in synthetic trial fixture.
read_file({
  path: "trials/aaaaaaaa-1111-4111-8111-aaaaaaaa1111/v1/about.md",
});
read_file({
  path: "trials/aaaaaaaa-1111-4111-8111-aaaaaaaa1111/v1/uh-rcc-301.md",
});

The recorded run read both trials and revisited narrow patient-source line ranges. It identified the later-line RCC study as potentially relevant, kept unrecorded screening facts unknown, and excluded the biomarker study because its required actionable alteration was absent. The six deterministic checks and four clinical criteria passed.

The saved trial draft supports a criterion review with Supported, Mismatch, Unknown, or Not applicable rows. Recruitment is the status recorded in that document on its effective date. A clinician confirms eligibility information and decides whether to pursue a trial; the workflow does not enroll or contact anyone.

Follow an example in the code

Run the two bounded treatment and trial examples with:

bun run eval -- treatment-planning/retrieval/development/cases trial-matching/development/cases/0000 --max-concurrency 1 --skip-report

These evaluations exercise drafting against synthetic files. Application tests separately cover access checks, evidence registration, durable artifacts, and clinician actions. How AI requests run describes those application steps.