# EdAI Systems

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [John Ferguson](https://cerebralvalley.ai/u/DrJCF)
- **GitHub:** https://github.com/drjcf/cogpath
- **Demo video:** https://youtu.be/S8-Z36mcNro
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1091

A model asked for an exam question that requires analysis and synthesis writes one whose answer the article already states. Specialty boards write items to a blueprint of cognitive levels, and every generated item that fails that standard costs a reviewer's time or the exam's validity.

cogpath treats the cognitive path as a program: recall(4) > analyze(causal) > evaluate(efficacy, safety) > synthesize(management plan). Each operator has parameters, a strict JSON schema for its artifact and dependencies on earlier steps, and Muse Spark executes it as a separate call on the Meta Model API with effort set per step. Before any item, the source becomes claims as verbatim spans located in the PDF, audited and certified by a person. The item is assembled from the artifacts: the key is the terminal artifact, each distractor is the option a learner who failed at one step would choose, and the stem is written to a case the source does not cover.

The verifier then uses the model against itself. It asks for the answer with the notes of steps 1 to k only, with the source withheld and outside knowledge forbidden, and requires the key to become determinable only at the last step; a judge in code re-checks every quoted piece of evidence against the notes. It corrupts each step the way a learner who fails there would, re-runs the rest, and requires the key to be unreachable. A single-shot control on the same focus takes the checks both arms share.

Across ten builds on six sources (50 items), 40 keys were not stated by the source, 37 needed the full path, 15 passed all six checks; single-shot controls, 15 of 50 keys not stated. A reviewer certifies or rejects each item with a reason, and the bank exports with the path, the checks and the decision.

It runs in production at cogpath.org: a React console on Firebase Hosting, an Express API on Cloud Run, a Cloud Run Job per ingest or build with live progress, Firestore and Cloud Storage for the records, Firebase Auth with a reviewer allowlist, the model key in Secret Manager, and a test suite that runs the pipeline, the store and the API end to end.

## More from Meta Global AI Developer Hackathon

- [jobseekers](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1088)
- [CRYPTOR.AGENCY](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1089)
- [First Pass](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1090)
- [PackMatch](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1092)
- [AnaliZac](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1093)
- [Koanic](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1094)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1091. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt