A public research log of AI-session mistakes, caught and corrected — primary-source material, with the full evidence trail for each.
2026-08-15 · Demand gate + distribution planning
When an AI plan assumes human-paced waiting — a 48-hour research gate, a 30-day calendar — question it. Full evidence trail: the prompt, the response that felt wrong, the challenge, and the fix.
2026-08-15 · Review protocol formalization
Nothing is done until two independent critics from different model families pass. The research behind "2 critics," and the first deployment that caught 5+ real bugs. See the framework.
2026-08-15 · Framework institutionalization
The meta-experiment: the framework emerged from the experiments, then became the gate for every experiment. Adopted as Experiment 004 — the tool is the product.
Lessons are appended as they happen. Each one is primary-source research: exact prompts, exact responses, timestamps, and the correction — the kind of material AI engines cite. Subscribe via RSS.