Lab

From a question to relevant work.

The model selects relevant records instead of writing new career claims. It has no file access or action tools. If the API fails, keyword retrieval keeps the public work searchable.

What happens when things go wrong?

Test an unsupported claim, challenge a project attribution, or simulate a service failure. The failure exercise skips the model and demonstrates the fallback.

Recorded guide checks.

gpt-4.1-mini, 2026-09-21T03:01:44.141Z. These specific prompts were tested; they are not a guarantee of every possible response.

Same task. Clearer instructions.

Two model runs add a dismiss action to the same React fixture. The second also receives explicit reuse, accessibility and styling guidance.

Recorded 2026-09-21T01:59:22.905Z with gpt-4.1-mini. Temperature 0, one run per condition.

import './notice.css';
import { Button } from './Button';

export function Notice({message,onDismiss}:{message:string;onDismiss:()=>void}) {
  return <section className="notice">
    <p>{message}</p>
    <Button className="notice-dismiss" onClick={onDismiss} aria-label="Dismiss notice">×</Button>
  </section>;
}

import './notice.css';
import { Button } from './Button';

export function Notice({message,onDismiss}:{message:string;onDismiss:()=>void}) {
  return <section className="notice">
    <p>{message}</p>
    <Button type="button" aria-label="Dismiss notice" className="notice-dismiss" onClick={onDismiss}>×</Button>
  </section>;
}

One paired fixture, one run per condition, same model and temperature 0. Structural checks do not establish general skill effectiveness, complete behavior or accessibility conformance. Model output can vary. Guide checks test these exact prompts; they are not a guarantee against all attacks. API usage tokens are observed; dollar cost was not measured.