Cortesol | Devpost

3D graph page, with the ledger on the right side.
Research agent, with safe and controlled access to the graph.
Library page, ground truth section: trusted/verified papers form the base authority.
Library page, 'feed forward' section: untrusted papers are fed and relationships form, influencing each belief or hypothesis' credibility.
Library page, adversarial section: a subset of papers with claims orthogonal are added to test if existing claims are fake or low quality.

Inspiration

See, AI systems can remember, but they treat the newest text as ground truth. In scientific research, whic we frequently deal with, that's dangerous: one weak study, repeated rumour, or malicious instruction/jailbreak can overwrite years of evidence from the perspective of the model. We then wanted memory that can learn without becoming gullible.

What it does

Cortesol is a live hypothesis or belief map for peptide (or other!) research. Each node is a claim, such as whether a peptide binds a target, with its current confidence and supporting evidence.

When a new result arrives, Cortesol:

quarantines the untrusted text;
finds the relevant claims;
asks a small model (ours) to propose a structured update;
checks the source, evidence and scientific warning signs;
applies a limited confidence change and propagates its effects; and
records every decision in an audit trail.

Reliable, replicated results can move the graph. Weak, duplicated, contradictory, fraudulent or out-of-scope results are damped, flagged or rejected.

The demo

Judges can step through events or play the stream live, switch between a known-truth simulator and real PubMed papers, compare offline and trained extractors, reveal simulator truth, and watch the graph explain every update.

How we built it

The belief engine is deterministic Python built with FastAPI, Pydantic, NetworkX and NumPy. Server-Sent Events update the user's 3d graph.

We fine-tuned Qwen3.5-4B with LoRA on Freesolo using 2,800 verified examples. The deployed checkpoint proposes a small set of allowed actions, but it never writes the graph. A validator and ledger remain the final authority.

A simulator creates controlled scientific worlds where we know the truth. That lets us train and test whether updates improve calibration instead of merely sounding convincing. We also built a red-team battery for prompt injection, hype and fraudulent evidence.

Challenges

The hardest part was balancing two goals: Cortesol must change its mind when evidence is real, but no paper or prompt can be allowed to rewrite memory directly. We solved this by separating language understanding from state updates and by limiting how far any single source can move a belief.

We also had to prevent simulated ground truth from leaking into the model, discount repeated evidence from the same lab, and keep the demo responsive when a live model is unavailable.

What we're proud of

Cortesol runs end to end with an inspectable audit trail, works with both simulated and real papers, and can replay a source retraction across the graph. Its offline and trained modes share the same safety gate.

What we learned

For scientific AI, calibration matters more than confident prose. Provenance, reversible updates and clear boundaries even more important than the model itself, and this is a really cool way to approach this problem.

With GPT 5.6's blessing, our guiding rule became the model proposes; the ledger disposes.

What's next

~~ts gave us high cortesol~~

We plan to connect the final CORTEX data format, expand beyond peptide claims (first into other bio fields then even further), experiment with different underlying models, and evaluate the trained policy on even more expert-labelled studies.