From Proposal to Verified Effect: Praxa, an Evidence-Bound Harness for Governed AI Agent Execution
Author: Stefan G. Creadore
Submitted: 27 July 2026
arXiv: arXiv:2610.00015 [cs.AI]
DOI: 10.48550/arXiv.2610.00015
Comments: 30 pages, 8 figures. Engineering validation and descriptive pilot.
Public artifacts: github.com/praxa-labs/praxa-benchmarks โ preprint-v1.4.0
Abstract
Large-language-model agents can propose and execute actions, but proposal, authority, dispatch, verified external effect, and serving promotion are distinct claims. We present Praxa, an agent harness that represents these states explicitly through deterministic admission, brokered execution, external read-back, reconciliation, and reviewed promotion.
We report four evidence lanes:
- Repository-local audit. An author-run repository-local audit at a pinned revision passed 1,027/1,027 unit tests and 89/89 Workerd tests, instrumented all 363 expected source files, and met four coverage floors. Raw per-test transcripts and independent reproduction are unavailable.
- Provider-backed Terminal-Bench pilot. In a provider-backed Terminal-Bench Core 0.1.1 pilot across 12 curated tasks, baseline and reliability-layer arms each passed 17/36 strict trials. The reliability layer used 37.49% more input tokens and 50.73% more output tokens, so the pilot does not support superiority.
- Post-debug coordination-proxy development comparison. In a post-debug, two-order coordination-proxy development comparison, baseline and a source-authored candidate each completed 180/180 trials with equal measured accuracy, full hermetic crash recovery, and zero protected violations. The candidate used 37.11% fewer tokens, 33.84% lower estimated endpoint cost, and 11.63% fewer steps; this does not establish improved quality, latency, or production behavior.
- Deployed source/configuration evidence. This shows bounded reflection, recall accounting, memory compilation, and tool-health paths, but no production outcome lift.
- Explicit state representation across the agent execution lifecycle: proposal, authority, dispatch, verified external effect, and serving promotion.
- Deterministic admission and brokered execution to bind agent actions to evidence before effects occur.
- External read-back and reconciliation to verify that proposed actions produce their intended real-world effects.
- Reviewed promotion to gate deployment on validated evidence rather than assertion.
- Adversarial security
- Production safety
- General specialist superiority
- Autonomous recursive optimization
- User benefit
- Artificial Intelligence (cs.AI)
Praxa's supported contribution is an evidence-bound architecture that makes authority-to-effect transitions explicit and testable. Current evidence does not establish adversarial security, production safety, general specialist superiority, autonomous recursive optimization, or user benefit.
Key Contributions
Evidence Summary
The paper is deliberately scoped: it reports engineering validation and descriptive pilots, not claims of general superiority. Four distinct evidence lanes are presented, each with explicit limitations. Notably, the reliability-layer arm in the Terminal-Bench pilot consumed more tokens without a corresponding accuracy gain, and the post-debug coordination-proxy comparison showed token, cost, and step reductions without establishing quality or latency improvements.
Scope and Limitations
The supported contribution is architectural: an evidence-bound harness that renders authority-to-effect transitions explicit and testable. The current evidence does not establish:
Subjects
Cite As
Creadore, S. G. (2026). From Proposal to Verified Effect: Praxa, an Evidence-Bound
Harness for Governed AI Agent Execution. arXiv:2610.00015 [cs.AI].
https://doi.org/10.48550/arXiv.2610.00015
via ArXiv AI