ASG Scientific Audit Workbench

A human-supervised AI research system that turns domain expertise into executable, reproducible, and evidence-gated software.

Comment

Inspiration

What it does

Inspiration

Scientific projects built with AI can become difficult to review when speculative hypotheses, tuned models, hidden targets, and incomplete provenance are mixed together. ASG was built to make that process more honest and checkable.

The atomic/nuclear descriptor model is the case study. The submitted Build Week product is the audit workbench around it: a browser-based verifier, claim boundary, report generator, and release-checking workflow.

What it does

ASG Scientific Audit Workbench verifies evidence packages, SHA hashes, manifests, metrics, and claim boundaries. It classifies results as PASS, FAIL, HOLD, or BLOCKED so that a reviewer can see what is ready, what is partial, and what must not be claimed yet.

The current case study includes frozen mass and scoped nuclear charge-radius residual evidence, with explicit limits. Electronic/PIM shell physics and pure DNK node mechanics remain gated.

ASG is not submitted as a completed new physics theory. It is submitted as a reproducible scientific audit workflow for high-risk research claims.

How I built it

ChatGPT helped develop and refine the research reasoning. GPT-5.6 / ChatGPT Work helped convert research questions into bounded plans and critique weak claims. Codex made the work executable: it inspected files, wrote and ran tests, diagnosed failures, protected accepted artifacts, built the browser workbench, and returned evidence-gated outcomes.

The Build Week layer includes:

  • browser demo route;
  • browser-side SHA-256 verification;
  • claim-boundary display;
  • PASS / FAIL / HOLD / BLOCKED audit output;
  • downloadable JSON audit report;
  • presenter mode;
  • local launcher scripts;
  • pre-send verification checks;
  • public GitHub repository and demo video package.

Challenges

The main challenge was separating real evidence from hypothesis. The system had to preserve accepted model artifacts, avoid sealed truth, avoid leakage, and clearly mark gated areas.

Another challenge was presentation: the project needed to be understandable to judges and physicists without overstating the science.

What I learned

The most useful pattern was treating Codex as an engineering audit partner, not just a code generator. It helped turn a complex scientific archive into a reproducible, bounded, reviewable workflow.

What's next

The next step is an optional OpenAI-powered review assistant layer that guides a physicist or engineer through the evidence package, explains PASS/HOLD/BLOCKED states, and suggests safe next-review actions without modifying protected model artifacts or accessing sealed truth.

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for ASG Scientific Audit Workbench

Built With

Share this project:

Updates