Source note: Local copy of a Summer Into AI 2026 submission originally published on Substack. Screenshots, video, and audio found in the post body are mirrored locally when publicly accessible; profile and avatar images are intentionally not copied. Read the canonical post.
Week 4 of Summer into AI 2026 hosted by @advisoryhour — theme: Built on Yesterday.
FOIA Ghost compiled a classified dossier on your ZIP code from real EPA, Census, and FEMA data. Sworn Testimony puts that dossier on the witness stand. Cross-examine it like opposing counsel — and while it answers, a second AI audits every claim against the raw federal data and stamps it SUPPORTED, UNVERIFIED, or CONTRADICTED. The witness has been instructed to occasionally embellish. Catch it lying and your score goes up.
Worth noting alongside this: Eric's Arcadia (Week 4) rebuilt nearly fifty past competition projects into a 3D arcade you walk through. Sworn Testimony goes the other direction: it takes one past project and interrogates it under oath
Built on yesterday — what changed
The original: FOIA Ghost (Week 3) — a passive dossier generator — you read what it wrote
The dossier is now a character: an evasive, bureaucratic witness you cross-examine in a streaming deposition
A second AI — the Verifier — audits every answer against the raw EPA/Census/FEMA JSON and renders SUPPORTED / UNVERIFIED / CONTRADICTED verdicts with cited evidence
The witness is secretly instructed to fabricate details roughly one answer in three — catching a lie scores a point, so cross-examination strategy matters
Same real federal data pipeline as the original, reused line for line
How the AI works
The dossier synthesis — real EPA Toxic Release Inventory facilities, Census ACS income and poverty, and FEMA disaster declarations are fetched live for any ZIP, then Claude writes the case file entered into the record.
The witness — Claude roleplays the case file itself as a sworn, reluctant witness, streaming answers that cite the real figures. A hidden server-side coin flip tells it to embellish about a third of the time; the browser is never told which answers.
The Verifier — AI auditing AI — an independent forced tool call compares each answer against the raw data and returns a structured verdict with evidence. Green, amber, or red badge on every answer. When the badge goes red, you caught it.
How to play
Enter any US ZIP code and open the case file.
Read the dossier, then cross-examine: ask about pollution counts, poverty rates, disaster history — or use the suggested questions.
Watch each answer get audited. A red CONTRADICTED badge means the witness lied and you exposed it — the score at the top counts your catches.
Press it on specifics. Vague questions get vague (safe) answers; precise questions force precise claims — which are easier to catch.
Where to play
Demo: week-4-demo-5-sworn-testimony.vercel.app
Code & README: github.com/GlimmerForge/summer-into-ai → demo-05-sworn-testimony/README.md
Summer into AI 2026 · Theme 4: Built on Yesterday
Original source
Canonical Substack URL: https://jakestrait5.substack.com/p/summer-into-ai-2026-sworn-testimony.