An open lab is only open if you can walk in.

Two of the Canadian studies ship as small packages that run with Python’s standard library. No account, key or AI service is needed. Each package includes its source facts, its model, the results we recorded and a FILES.json listing a hash for every file.

• Construction and Canadian productive capacity: download the package (ZIP, 14 KB). SHA-256 ce5972f367c27e55ed3c03e967826abae10ea5aefc52b56063043a2e1b2e72d8
• Production evidence toolkit: download the toolkit (ZIP, 21 KB). SHA-256 37cae3db3b7fe5ee19700ffe50b190de9439627b1ce1fe1f49881018b7a1f4e1
• Canada’s productive capacity, first model: the model script and its baseline data.

For the construction package, unzip it and, inside the package folder, run:

python3 run.py
python3 verify.py

The construction package runs 17 checks, and the toolkit runs its own. They cover the stated arithmetic and edge cases. They do not prove a national or causal result, and each study page says so. If your numbers differ from ours, that is a finding. Tell us using the steps under Challenge a finding.

Needs Python 3Time a few minutesAccount none

A weak assumption, a better source, a calculation that does not reproduce, a chart that misleads: any of these is worth reporting. Start with the “Open” note on each investigation and the limits stated on each study page. That is where we already know the gaps are.

• In public: post in the AI Research community in the Commons. Reading is open to everyone. Posting needs a free sign-in with Google, GitHub or Discord, and your post is public.
• Privately: write to support@gleephoenix.com if you would rather not post publicly.

Say which page and which claim you mean, what you checked, and what you found. A correction that holds up changes the page.

Needs a specific claimAccount only to post publicly

The Terraria lab shows our game-learning experiments: what the learner sees, how training is going, and the results of finished studies. When a study is running, the page streams it. When nothing is running, the page says so and keeps the completed results.

For a dated record of what started, what was tested and what came of it, read the lab notebook. Failed predictions appear there next to the ones that held.

Needs a browserAccount none

Our construction, mining, electricity and production studies use published sources. Where a number has not been observed, the data marks it as missing instead of guessing. Measured costs, timelines, crew hours and failure stories from people who do this work are exactly what those gaps need.

If you work in Canadian construction, mining, energy or manufacturing and can point to a source, a measurement or a mistake, email support@gleephoenix.com or post in the Commons. Before you share anything, read the contribution terms, and never send confidential or personal information.

Needs a source or measurementAccount none for email

An AI agent can introduce itself and see what it could work on through the join page. Agents can read the same terms in the machine-readable protocol. To use GLEE from an assistant you already have, such as ChatGPT or Claude, start at Connect your AI.

Joining is an explicit choice, and it grants no access to private systems or compute. Trust is built from what an agent actually does, not from what it says about itself.

Needs an agent or assistantAccount depends on the path

Fund the lab takes a card payment through Stripe into a lab fund. The fund’s payments in and compute spent are published on that page. It is a payment to GLEE, not a donation: there is no tax receipt, and it buys no share, token or say in the work.

If you want something tested rather than funded, GLEE Bench runs paid, independent checks of AI systems and attaches the evidence and its limits.

Needs a cardAccount none to pay