01 / Blogue

Notes du terrain.

Taxonomies de défaillances, motifs défensifs, et les décisions d'architecture qui empêchent les systèmes LLM de s'effondrer à 2h du matin.

À LA UNE

Meta-Prompting in Production: One File the Agent Can Touch, One Verifier It Can't

One AI writing prompts for another sounds like leverage until you watch it fail: an 80-node workflow that repeated every mistake, re-paid every failed run, and trusted every generated instruction. Tuning the meta-prompt never fixed it. A three-layer contract did: a verifier the generator can't touch, domain facts it can't ignore, generation left free in the middle. Two production failures in its first 24 hours stress-tested exactly that.

La couche méta15 juillet 202610 min de lecture

Filtrer par pilier

Taxonomie des défaillances en production Cadre STATE appliqué Architecture défensive La couche méta IA réglementée et Loi 25 Teardowns

TeardownCadre STATE appliqué21 juillet 202610 min de lecture

Morgan Stanley AI @ Work Assistant STATE Teardown: Evals Are Not Traces

Morgan Stanley runs one of the most disciplined eval programs in enterprise AI: pre-deployment evals, human-graded tests, a daily regression suite. The STATE score is still 4/10, because nothing in the public record can reconstruct what the assistant told one advisor about one client yesterday. This teardown dissects the gap between aggregate evals and per-interaction traces, and shows the single table that closes it.

TeardownCadre STATE appliqué6 juillet 20269 min de lecture

Ramp Financial Automation Agent STATE Teardown: Graduated Autonomy Proves the Decision, Not the Workflow

Ramp runs one of the most disciplined AI agent rollouts in production finance — graduated autonomy, human escalation, 99% policy accuracy. STATE score: 6/10. Across three public sources, there is no evidence the approval workflow survives a crash mid-flight.

Taxonomie des défaillances en production8 mai 20266 min de lecture

The AI Didn’t Hallucinate. It Told the Truth. That Was the Problem.

A technically solid LLM email pipeline — 10,000 emails, 36 classification buckets, six N8N workflows — reproduced institutional confusion at scale and called it procedures. This is what Knowledge State Failure looks like, and why no model upgrade fixes it.