Programme · Claims register
Claims register for Bayanat Labs open releases
Every finding we state on an open release page is listed here first, with the date it was checked and the scope it holds within.
This register lists the findings that Bayanat Labs states on its open release pages. Each one has a claim ID, the date it was last checked and the scope it holds within. A release page links to the row here next to the claim.
Our open release pages make no "largest", "first" or "only" claims. If we ever make one, it will appear in this register first, with the comparison it rests on, the unit, the licence class and the date of the search.
Findings stated on release pages
| Release | Finding, with its scope | As of | Claim ID |
|---|---|---|---|
| LabelBench | On the fixed held-out manifest, a local language model (Qwen2.5 3B) reached 78% binary toxicity accuracy, 25.3% exact match on the seven-label taxonomy and 22.4% across five dialect regions (chance 20%). Small local models on fixed manifests. Not a ranking of current frontier models and not a rating of Bayanat Labs catalogue products. | research-labelbench | |
| Arabic Agent Reliability Lab | Methods beta for state-based evaluation of Arabic agents in an open synthetic sandbox (21 workflow families) with evidence receipts. "A fluent answer is not a completed transaction." Only the claims marked Allowed in the lab's claims ledger. Never a certification, a dialect-quality result or a prompt-injection result. | research-agent-lab |
How we check claims
- Numbers come from files. Every count on a release page is computed by a script from the released files, and the build fails if the two disagree.
- Findings keep their scope. A research finding is stated with the model, task and test set it was measured on, and never as a general rating.
- Dates are visible. Each claim shows the date it was last checked.
Report an error
If a claim here is wrong or out of date, email hello@bayanatlabs.com. We correct the register and the release page together.