Governing Untrusted AI — the Agent Harness study
How to turn an untrusted, weak model into governed, evidenced work with zero patches to the core — a reproducible study with a full evidence trail you can re-run.
- The architecture: four ownership layers, and why the model gets zero effect authority
- The methodology: evidence classes, six veto gates as property tests, pre-committed thresholds
- Crash-proof exactly-once effects and a live external veto — one complete sample chapter
- Results across all six gates, the weak-model production playbook, and cell economics
- The clean-room reproduction (94/95), re-runnable from a fresh checkout
Weak or untrusted models can still do real work if the work is governed by construction: every effect is declared, policy-checked, and captured as signed, replayable evidence, while the model itself holds no authority to act. Across 31 reviewed tasks, six formal veto gates, and a 94/95 clean-room reproduction, this study documents the harness, the experiments, and a reproducible evidence trail you can re-run.
Read before you buy.
The opening pages and the method, in full — no email, no gate.
The work behind it is inspectable.
This isn't a claim on a slide. The research runs on CHP — every step is a signed, hash-chained record you can open and replay.
- execution_startedb4704fe7be…8784bf#2688
- llm_requestb519112daa…e89fc3#2689
- llm_response6472dcf99d…25ea67#2690
- execution_completed✓ successb0fcca823f…62aaaf#2691
Not a paper — a running system.
Auxo Agent Harness is the system this research comes from — The runtime behind the study above: untrusted models do real work under policy, with signed, replayable evidence and no authority to act on their own.