Exocortex R&D: a lab where evidence decides about AI

The problem

Companies deploy AI after a good presentation. Success criteria are set after the results are seen, failed attempts disappear from the report, and later nobody can explain why a solution was chosen. Audit and risk have nothing to check. There is a second risk too: publishing results makes it easy to carry client data out of the organisation. The lab addresses both problems at once.

How it works

• Criterion first, measurement second. Every experiment starts with a card stating the question, the measure of success and the threshold. It is frozen, and its checksum is published before anything is calculated. Changing one's mind means a new version of the card, and the previous one stays visible.
• A person makes the decisions. An idea goes through a quick test, then a pilot with blind assessment, and finally gets a rollout plan. After each stage a person decides whether to go on. Most ideas end at the quick test.
• The machine counts, the person judges. The machine collects signals, calculates and publishes. The person writes and approves the card, assesses samples blind and makes the decisions.
• A steady flow of ideas. Every week I review new publications, models and datasets. In the first full week 198 papers went through the radar.
• Negative results are published too, because they show where not to invest.

Security

No file reaches the public repository without 5 automated checks, including personal data and similarity to private material. A file stopped by a check waits for a person's decision. I test the gate itself too. Every night it receives a set of files with deliberately planted leaks and must stop every one of them.

The lab is cut off from my private data. Every night a test from inside the lab tries to connect to the private database, and every such attempt must fail. Client materials are not and will not be a data source for the lab. Every file also records whether a model or a person wrote it and whether a person checked it, so an auditor can get an answer about what belongs to whom.

What it gives an organisation

A decision to deploy rests on a threshold set in advance, so the criterion cannot be fitted to the result. An outsider can reconstruct from the record of the work why a solution was chosen. Pilots end instead of dragging on, and publishing results does not carry client data out.

Status as of 30 September 2026

Running: the gate with nightly tests, the public repository (since 29 September), the lab.exocortex.zone site, nightly isolation checks, the experiment queue and the idea radar. The first experiment is in blind assessment. No hypothesis has a result yet. I say this plainly, because the whole project is about not promising more than the evidence shows.

My role

I am the project owner: I set direction and scope, approve hypothesis cards, assess blind and make the gate decisions. I run the execution through agents (Claude Code, Cowork), following the division of labour described in the lab's documentation.

What this project proves

That I can design and run a way of working in which an organisation makes decisions about AI based on evidence, defends them before an audit and does not leak client data along the way.