How a video correction reaches the platform
Every step below is described in the project's own documentation and code. Numbers are a snapshot taken on 2 November 2026 from the pipeline's state file.
1. Text correction (first layer)
- A model drafts a written correction for each question. Its instructions say: solve the question yourself before looking at the official answer key, then compare.
- If it disagrees with the official key, it still writes the correction for the official key, but marks the disagreement, sets its confidence to low and explains the clinical reason and the letter it would defend. Disagreement is flagged, never silently resolved.
- The draft goes through automatic gates and then a clinical audit by Claude; a draft never goes straight to the site. All 23 demo questions have a text correction marked as audited by Claude.
2. Video script
- A Claude agent writes the narrated video script from the question and the audited text correction (all 23 demo videos were scripted this way).
- Before the clinical audit, automatic gates run: the script is applied to a template, linted, validated and rendered as a preview. A script that fails twice is marked as failed and the queue moves on.
3. Clinical audit (second layer, separate agent)
- A separate Claude agent, acting as a residency preceptor, audits the script before it may use the GPU. It checks, in order: the defended letter is the official key and matches the text correction; clinical facts, doses, cut-off values, time limits and conduct match the text correction and current Brazilian guidelines; units are complete; each on-screen label describes what the photo or drawing really shows; no contradiction between narration, screen and alternative.
- Verdicts: approved (nothing to fix), corrected (the auditor fixes a specific line, label or drawing) or rejected (the script goes back to the producer with the reason).
- The rendering queue only takes scripts marked as audited. A test (with a negative control) proves that an unaudited script is refused.
4. Checks before and after rendering
- Before the GPU: every number spoken with a unit must be written on screen, and photos cannot be enlarged more than 1.5 times. Failures go back to the producer without spending GPU time.
- After rendering, each video gets a metrics report. Every check is paired with a negative control: a deliberately broken input that must make the check fail. Examples from one demo video (PSU MG 2025, question 1):
| Check | Negative control |
|---|
| Speech recognition of the final audio: global ≥ 0.97, worst line ≥ 0.92 | fails as expected |
| Every narration line is present in the final audio | one line silenced on purpose: coverage drops from 100% to 0% |
| Video length ≤ 7 minutes and audio = video = timeline | fails as expected |
| No repeated image (sha256) | fails as expected |
| Script validator: 0 errors | 12 planted defects are caught |
| Layout geometry test | 3 broken layouts (highlight, label, overlap) are caught |
5. Fingerprints
- For theme lessons, a script only goes to the GPU if the audit record holds the same fingerprint (sha) as the script. If the narration changes, the audit has to be redone.
- After upload, published videos are checked online, including their sha256, before they are listed on the site.
6. What the numbers look like (snapshot, 2 November 2026)
- 701 video scripts approved and rendered; 654 of them were written by Claude agents.
- Of the 701: 509 approved by the auditor with no change, 192 approved only after the auditor corrected something.
- 104 scripts were rejected by the auditor at least once and sent back.
- A real test of the clinical audit on one question: about 45 seconds and US$ 0.23 per question; it found 2 real defects in a pilot video that had already been approved (method steps out of sync with the drawing, and a misspelled word inside a drawing).
7. Example: clinical audit of 9 October 2026
- Scope: 11 new mini-lessons and 43 finished question corrections.
- Answer key against narration: no disagreement in the 43 corrections.
- Corrections made in the mini-lessons: acute hemorrhage reduces HbA1c (the script said it did not); adolescent confidentiality now states the 14-year threshold (below it, sexual activity is presumed violence and must be reported); a guideline-change year not found in the sources was removed.
- Points the model could not settle (for example, which amenorrhea age criterion to teach) were listed for the physician to decide, not decided by the model.
How the 23 demo questions were chosen
- PSU-MG first-year (R1) exams 2023 to 2026, up to 5 per area.
- Video approved and audited; script written by a Claude agent; official key present (no annulled question) and equal to the letter defended in the script.
- No third-party commentary existed for the question at any stage of generation; no image in the question statement; not in the list of questions whose answer changed after the exam.
- Every photo in the video has a recorded open license (or public domain), credited on the question page.