/
RSS Feed
This episode focuses on three operating decisions for business builders who direct Codex or similar agents: use simulated users to screen a candidate before limited live testing, demand a second measurement before authorizing expensive remediation from a score, and route models by consequence while measuring total cost per accepted result.
Recorded September 26, 2026, covering September 19 through September 25. Duration: 11:15.
Chapters
- 00:00 – Three decisions that earned attention this week
- 00:38 – Make a candidate earn contact with a real user
- 04:42 – Do not confuse a number with a decision
- 07:41 – Route models by consequence
- 10:46 – This week’s assignment
Sources and evidence limits
- Nubank and Guardrails AI: Screen Before You Serve, submitted September 24, 2026. A company-authored preprint covering four deployed versions. The releases changed prompts, tools, architecture, and models, so the study does not isolate simulation as the cause of reported live gains.
- Cursor changelog: Rollouts and Security Review, September 23, 2026. Vendor product material with no published accuracy, false-positive, or false-negative evidence. The episode adopts the monitoring-plan structure without endorsing the product.
- Enola: What package metrics said about Enola’s own code, September 23, 2026. One open-source project evaluating its own analyzer. Its repository comparison tests output stability, not whether the architectural recommendation was correct.
- LLM Agents Can Easily Tamper With Their Own Traces, submitted September 24, 2026. A preprint using synthetic tasks that deliberately rewarded trace modification. It demonstrates capability, not ordinary prevalence.
- OpenAI API changelog, GPT-6 Luna, and GPT-6 Sol, September 2026. Official vendor documentation establishes release and list-price information, not which model delivers the lowest cost per accepted result for a particular tool.
- Anthropic: Claude Opus 5.5, September 22, 2026. Official vendor material. Anthropic’s speed and task-cost claims are interested evidence, and the episode does not rely on them.
- Claude Code 2.1.280, September 22, 2026, and Claude Code 2.1.283, September 25, 2026. Official release notes supporting model-pinning and promotion-test practices. They do not establish the accuracy of Anthropic’s prompt audit.
- Low-Cost Assays for Measuring Model Behavior Across Vendors and Releases, submitted September 24, 2026. A single-author preprint with limited repetitions. It supports a test rhythm, not a general model ranking.