Cortesol

Company: Cortesol

Self-reported basis: The description is entirely self-reported and unverified; it originates from a Devpost submission for the OpenAI 2026 hackathon. No external corroboration, revenue data, or customer evidence is available.

What it appears to be: A prototype system designed to support scientific research agents by maintaining a belief graph that updates in a calibrated and auditable way, using log-odds arithmetic and domain-general critical appraisal of research papers.

What changed: The project was submitted as part of a hackathon; no evidence of prior development or commercial activity is provided.

Single most important open question: Is there any evidence that the belief graph system described can be scaled to handle real-world scientific workflows, or whether it remains a proof-of-concept?

The description states that Cortesol is a system for maintaining a live belief graph of claims, initially focused on peptides but designed to support any field. It processes research papers through a pipeline that includes:

• Quarantining text as data
• Retrieving related claims
• Performing domain-general critical appraisal (study design, statistics, red flags)
• Screening for fraud, hype, p-hacking, and prompt-injection
• Validating against a closed operation vocabulary
• Updating belief using bounded, source-capped, provenance-carrying arithmetic (ℓ' = ℓ + logΛ, capped by log(τ/φ))

The system is built with:

• Belief engine: Python, deterministic validator + log-odds engine over NetworkX graph
• Model stack: Freesolo / Flash (OpenAI-SDK-compatible), with FastAPI and Server-Sent Events for serving
• Frontend: 3D-force-graph visualization

Inference: The system is described as a prototype, not a production-ready product. It is not clear whether it has been tested in real-world scientific workflows or deployed beyond the hackathon.

The description states that Cortesol aims to be a "rough first step towards a scientist AI" according to Yoshua Bengio, and is described as a hypothesis synthesizer & evaluator. It positions itself as an agent memory system that updates belief in a way that mirrors careful scientific practice — resistant to manipulation and fully auditable.

It claims to:

• Update belief in proportion to evidence quality
• Be resistant to social pressure and sensationalism
• Provide a deterministic, capped, and auditable belief update mechanism
• Support cross-domain critical appraisal without subject knowledge

Inference: The positioning is aspirational — it is not clear whether the system has been validated or tested beyond the hackathon. The claim of being "a rough first step" suggests it is early-stage.

The description does not explicitly state a target customer or ideal customer profile (ICP). It implies that Cortesol is intended for scientific research agents, particularly those working in domains where belief updates must be calibrated and auditable. The system is designed to support domain-general critical appraisal, suggesting it may appeal to researchers, AI agents, or institutions involved in scientific validation.

Inference: The ICP is not clearly defined. It is implied that the target is scientific research teams or AI agents working in domains like biology, medicine, materials science, etc., but no explicit customer segment is stated.

There is no evidence of a business model or pricing structure in the description. The project is described as a hackathon submission and does not mention any monetization strategy, licensing, or customer acquisition plans.

Inference: No commercial model is evident from the provided information.

The system is built using:

• Backend: Python, FastAPI, NetworkX, Pydantic
• Model stack: Freesolo / Flash (OpenAI-compatible), OpenAI API key by default
• Frontend: 3D-force-graph via 3d-force-graph library
• Data flow: Text is quarantined and processed through a deterministic belief engine; no text ever writes state directly to the belief graph

Inference: The system uses a hybrid of LLMs for appraisal and deterministic logic for belief updates. It is described as having a capped, bounded arithmetic for belief updates, which suggests an attempt at robustness against model compromise.

The project was submitted to the OpenAI 2026 hackathon, indicating it is a prototype or proof-of-concept. There is no evidence of:

• Revenue
• Customers
• Product usage
• Deployment
• Iteration beyond the hackathon

Inference: The system has not demonstrated any traction or maturity beyond a hackathon submission.

The description does not mention any competitors. It positions itself as a scientific agent memory, with a focus on belief calibration and resistance to fraud or hype. It is not clear whether there are existing tools in this space, nor how Cortesol would differentiate from them.

Inference: No competitive landscape is described; the project appears to be self-contained within its own conceptual framework.

• Unproven scalability: The system is described as a prototype and has not been tested in real-world scientific workflows.
• No evidence of validation or testing: There is no mention of experiments, benchmarks, or real-world usage.
• Unclear commercial viability: No business model or pricing structure is evident.
• Limited team size (4 members): May indicate limited capacity for development or scaling beyond a prototype.
• Highly technical and abstract claims: The belief update mechanism is described in mathematical terms but not validated.

Inference: The project is early-stage, with no evidence of real-world application or commercial viability.

1. What specific scientific domains have you tested the system on, if any?
2. How does the system handle edge cases in belief updates (e.g., conflicting evidence)?
3. Has the system been validated against real research papers or only synthetic data?
4. What is the current status of the belief engine — is it fully deterministic, or are there still model-based components that could be compromised?
5. Are there any plans to integrate with existing scientific databases or repositories (e.g., PubMed)?
6. How would you scale this system for use by multiple researchers or institutions?