This forum post introduces the paper "From Proposal to Verified Effect: Praxa, an Evidence-Bound Harness" by Stefan G. Creadore, published on arXiv (2610.00015) on 2026-10-05.
Overview
Large-language-model agents can propose and execute actions, but proposal, authority, dispatch, verified external effect, and serving promotion are different claims. Praxa is an agent harness that represents these states explicitly through deterministic admission, brokered execution, external read-back, reconciliation, and reviewed promotion.
Four Evidence Lanes
1. Repository-local audit (author-run): At a pinned revision, 1,027/1,027 unit tests and 89/89 Workerd tests passed; all 363 expected source files were instrumented and four coverage floors were met. However, raw per-test transcripts and independent reproduction are unavailable.
2. Terminal-Bench Core 0.1.1 pilot (provider-backed, 12 curated tasks): Baseline and reliability-layer arms each passed 17/36 strict trials. The reliability layer consumed 37.49% more input tokens and 50.73% more output tokens. The pilot does not support superiority conclusions.
3. Two-agent development comparison (debugged): Baseline and source-candidate arms each completed 180/180 trials with equal accuracy; full sealed crash recovery and zero protected violations. The candidate reduced tokens by 37.11%, estimated cost by 33.84%, and steps by 11.63%, but this does not establish quality or latency improvement.
4. Deployment evidence: Shows bounded reflection, recall accounting, memory compilation, and tool health paths, but no production effect improvement.
Key Takeaway
Praxa's reliable contribution is its evidence-bound architecture, making the transition from authority to effect explicit and testable. The current evidence does not establish adversarial safety, production safety, general expert superiority, autonomous recursive optimization, or user benefit.
Paper link: arXiv:2610.00015