Most frontier models read a document beautifully. Ask one to transcribe a private equity capital call notice and it gets the fund, the currency, the amounts, the dates, almost always right. The word “almost” is the whole problem. A probabilistic reader gives you a number that is usually correct and never certified, and in a back office that reconciles other people’s money, “usually” is not a state you can sign off on.
By March, my agentic development setup had hit a wall. I was not vibe coding: I had built my own set of custom Claude Code skills, with conventions, templates, the whole apparatus. It held up fine on short tasks. On long ones, two problems kept coming back.
Over the dozen mobile missions I worked on, the main caveats of codebases I faced were that the Model is unclear and disseminated, and data paths and behaviors are hardly readable.
This article digs into these problems and offers concrete directions on the architectural level, as well as some advices on the implementation level.
Un micro-post pour vous indiquer que le code C++ de calcul d’ICM est disponible sur mon Github sous licence CC NC SA (pas d’utilisation commerciale et propagation de la license obligatoire).
Rappelons-nous une des conclusion du dernier épisode :
(…) le calcul prend 27ms pour une moyenne sur 15500 échantillons. C’est très raisonnable en soi, mais si j’envisage de simuler quelques milliers de tournois comportant des milliers de mains elles-mêmes incluant plusieurs calculs d’ICM, ce ne sera pas suffisant.