Field Note · 2026-07-01
How Hermon training follows evaluation failures
Hermon adapters improve through domain contract failures, targeted data, probes, promotion gates, and live demo checks.
Operators and builders who want to understand why MapleAI trains domain models instead of looping blindly.
01
The eval is the curriculum
Hermon training starts from concrete failures: missing JSON fields, weak refusal quality, poor domain language, missing rollback, or missing audit controls. Those failures become data-generation targets for the next adapter.
02
Promotion is not automatic
A new adapter can be trained and probe-passed without immediately becoming the public demo. MapleAI separates latest trained, latest promoted, and latest public snapshot so deployment remains auditable.
03
Why this matters
A small lab cannot outspend frontier model companies. It can win by creating sharper domain contracts, safer evals, transparent progress, and useful vertical models under Maple authority.