Praxagent

Independent AI research & applied systems. Mechanistic interpretability, controlled evaluation, and reproducible engineering — with public artifacts, code, and receipts.

praxagent.ai · Blog: praxagent.ai/blog · Code: github.com/praxagent

Praxagent LLC is an independent, self-funded AI research and applied-systems studio. We work in the open: model outputs, prompts, hashes, and data are published for inspection, and every claim ships next to the control that earns it.

How we work

• Open artifacts. Weights, code, prompts, hashes, and receipts are public — a result you can't inspect isn't a result.
• Controls before claims. Matched baselines, null tests, and decisions frozen before the outcome. A number without its control is rhetoric.
• Research that has to run. Theory grounded in real systems on frontier-scale models, with the failures reported plainly.

What's here on the Hub

• Jacobian lens for Qwen3.5-397B-A17B — a fitted Jacobian lens for a frontier-scale open-weights model (Apache-2.0), released with fit code, eval receipts, integrity hashes, and the full audit trail.

More lenses, evaluation artifacts, and replication data as the work lands.

What we research

• Mechanistic interpretability — Jacobian lenses, sparse-autoencoder feature mapping, internal-readout validation
• Evaluation & reproducibility — controlled comparisons, leakage checks, immutable artifacts
• AI safety & security — prompt injection, deceptive-behavior testing, guardrail analysis
• Agents & human collaboration — shared environments with tools, memory, and explicit oversight

Independent, self-funded, and honest about what reproduced and what didn't.