Run Undes Through Claude Code and Codex CLI
How Undes Pro can route selected roles through locally installed Claude Code and Codex CLI sessions while preserving the evidence and trust boundary around the result.
Read the setup note
Research and engineering notes
Undes Blog
Research, practical verification methods, and sanitized engineering case studies about making AI-generated work reviewable before merge.
Why Cheaper AI Code Generation Does Not Necessarily Reduce Engineering Cost
A quantitative comparison of model routing, sufficient-context retrieval, multi-model deliberation, automated checks, and evidence-based verification.
Read the analysisWhy AI-Generated Code Makes Review the Bottleneck
Generation gets faster immediately. Engineering confidence does not. The constrained resource moves into understanding, checking, and merge decisions.
Read the analysisHow to Verify AI Engineering Claims Against Repository Evidence
A practical method for separating observed facts, inference, assumptions, and recommendations while preserving the evidence chain.
Read the guideFrom Prompt to Reviewable Trust Artifact
How context preparation, critique, evidence expansion, reconciliation, and explicit failure semantics produce a reviewable artifact.
Read the architecture noteFinding the Real Lock Boundary in an OIDC Callback
A mutex existed, but the concurrency window opened earlier during token exchange and session creation.
Read the case studyWhen Evidence Disappears Between Pipeline Stages
The planner selected the right target, but the final answer never received it. The failure was in cross-stage state ownership, not model reasoning.
Read the case studyReviewing a Risky Pull Request Before Merge
A green unit suite does not cover a changed concurrent retry boundary. The review turns that gap into a bounded decision and reproducible open check.
Read the case study