October workshops are openBuild a Search AI Agent$99 early bird

Enroll now
Skip to content

PATHWAY PROJECT · CAPSTONE PROJECT · 3 SMALL IDEAS + A GUIDED LAB

Biomedical evidence review agent

A promising laboratory result can inspire a question without establishing a benefit for patients. Build the assistant that keeps that distinction visible.

You will learn to: Review a biomedical draft claim by claim, preserve measurement and version limits, and prepare a traceable evidence preview for a researcher.

Preparing your lesson and this browser’s progress…
Read the complete lessonAll the ideas in one place · works without the editor

Use these notes to review a concept or read at your own pace. The interactive workspace above adds predictions, editable code, and actual run results.

IDEA 1

Read the measurement before the headline

An evidence review connects each sentence to what a source actually measured. A laboratory signal, a questionnaire response, and a patient outcome describe different observations. Before accepting a number, read its measurement label, units, and sample count. A shared topic or an identical value does not make those observations interchangeable.

Our BENCH-1 card is completely invented: six fictional assay samples produce an authored signal value of 42 arbitrary units. It is a coding fixture, not a published experiment or evidence about a treatment. You can verify that a draft accurately repeats this record. You cannot use it to conclude that patients recovered, or that any biological intervention works.

Read the measurement before the headline

  1. Study record. Read the explicit measurement, unit, sample count, and limitation.
  2. Draft statement. Inspect what the proposed claim says was measured.
  3. Evidence boundary. A missing outcome remains unmeasured even when a number happens to match.

Invented biomedical records and scripted decisions. This diagram explains the checking workflow, not a real scientific result.

Inspect the review rule

if (claim.measurement !== study.measurement || claim.unit !== study.unit) {
  return { status: "needs_review", reason: "measurement_not_recorded" };
}

This rule compares declared fields in an invented dataset. It makes a specific mismatch visible; it does not read arbitrary biomedical papers or establish scientific truth.

Think it through: The card measures an assay signal. A draft describes recovered patients using the same number. What should the review say?

IDEA 2

A candidate claim must survive an independent check

A model-assisted research workflow might propose a summary and then use other tools to check it. That proposal is a candidate, not evidence. In this browser lab the candidate claims are authored input and the checker reads a captured copy of the original records. Changing a returned study object cannot rewrite what that checker uses.

The review keeps one result per claim. Missing sources, changed versions, and mismatched observations receive distinct reasons. A causal claim always goes to methods review in this deliberately limited exercise. Study design matters for causal conclusions, but our few equality checks cannot appraise design quality, bias, alternative explanations, or whether a real finding is reproducible.

A candidate claim must survive an independent check

  1. Read. Locate the requested study and keep its source version.
  2. Verify. Check the candidate against the captured original fields.
  3. Retain gaps. Return a review row even when a claim cannot be verified.

Invented biomedical records and scripted decisions. This diagram explains the checking workflow, not a real scientific result.

Inspect the review rule

const study = tools.readStudy(claim.sourceId);
const checked = tools.checkClaim(claim.id);
rows.push(checked);

The read makes source information observable in the trace. The separate check tests the original claim rather than trusting the candidate's self-description.

Think it through: The draft labels a claim causal, but all of its numbers match. What can this limited checker establish?

IDEA 3

Review a specific evidence package

A source can change after a draft cites it. Keep the cited source version in the claim and compare it with the fetched record before preparing a review. If they differ, expose the difference instead of silently substituting the newer value. A researcher needs to know which observation a statement relied on and what has changed.

The final tool makes an in-memory preview only when every row verifies and approval names the same draft and source versions. It independently rechecks the evidence rows. An accurate but unapproved table remains a useful draft. Even an approved preview is only an authored exercise artifact: there is no journal submission, clinical decision, or claim that this is a complete biomedical review system.

Review a specific evidence package

  1. Source versions. Ensure each claim cites the actual record being checked.
  2. Reviewed draft. Approval belongs to one specific version of the review.
  3. Local preview. Recheck the complete rows before returning a review artifact.

Invented biomedical records and scripted decisions. This diagram explains the checking workflow, not a real scientific result.

Inspect the review rule

if (!input.approval || input.approval.draftVersion !== input.draftVersion) {
  return { status: "draft", ...report };
}

A preview is the final local output, not permission to publish or advise anyone. An evidence gap and an approval gap remain different outcomes.

Think it through: A researcher reviewed draft 1; your new claims are in draft 2. What should happen?

Put it into practice

Produce one evidence row per invented claim and a reviewable draft; require the exact evidence and approval before a local preview.

  1. Read BENCH-1 and identify the measurement, units, sample count, and explicit limitation.
  2. Run the starter on the unmeasured-outcome case. Locate the place it trusts a claim without checking it.
  3. Insert the independent-check snippet, then compare the returned row and actual tool trace.
  4. Run every case, including the revised draft and missing source. Keep the different failure reasons visible.
  5. Change one field in the experiment input, predict the review outcome, and run your code again.

Your next experiment: The numbers still match, but the draft now claims causation. Predict what this checker can establish, then change the claim kind back and compare the traces.

An evidence assistant earns trust by showing exactly what was checked, what was missing, and which version a person reviewed.

Key terms

Assay
A procedure that measures a specified property; this exercise uses invented assay fields only.
Evidence review
Examining whether a source supports a particular statement and retaining its limitations.
Source version
The identified edition of a record used to support and review a claim.

Sources and scope

Original Stemtiq teaching, reviewed 2026-09-14. The named researchers and organizations do not endorse this course. Classroom cases are authored exercises, not published findings.

NIH: Understanding Clinical Studies

Reviewed by NIH 26 June 2025 · primary indexed text checked 13 September 2026

Observational associations and randomized intervention designs support different kinds of conclusions.

This explains study design; it does not provide evidence for the fictional study cards in this lab. Direct page access was blocked during review; its primary indexed text was available.

Anthropic: Writing effective tools for agents — with agents

11 September 2025

Design distinct tools with clear parameters, relevant returned information, and evaluations of how the agent actually uses them.

A description or schema does not guarantee the right action. A live tool can return different data for the same arguments as its environment changes.

Anthropic: Demystifying evals for AI agents

9 January 2026

Define tasks, trials, and graders; inspect both execution records and final outcomes; repeat trials when model behavior varies.

A score depends on its cases and grading rules. Repeating a deterministic classroom case does not measure the variability of a live model.

OpenAI: Guardrails and human review

Reviewed 13 September 2026 · undated documentation

Distinguish automatic checks from approval decisions, pause sensitive tool requests, retain state, and resume after an application approves or rejects them.

Model-generated approval text is not authorization. Resume examples that automatically approve a request do not establish that a person reviewed it.

Your JavaScript really runs. The model decisions and school data are authored simulations, so you can learn without an API key. Every workspace also includes a separate real SDK example to explore next. Passing the lab’s cases is practice, not proof that an agent is ready for real-world use.