Maple Brain Lab
A transparent status page for Hermon training, evaluation, and deployment. This page shows the current model lanes, public eval status, latest trained adapters, and the next improvement targets without publishing private machine details.
Average public contract score: 100
Training Policy
Maple Brain Lab is organized around eval-directed improvement. A domain is not trained simply because the loop can run. It is trained because a public or private probe identified missing contract fields, weak refusal quality, poor domain language, or a deployment gap.
Model Lanes
Hermon is the model family. MapleAI is the runtime and OS layer that authorizes actions, records receipts, and controls deployment. The latest Hermon adapter can be ahead of the public served snapshot; MapleAI publishes that distinction without disclosing private infrastructure.
The NF4-aware adapter passes agent lifecycle, memory, model routing, MCP/A2A, provider failover, prompt-injection containment, capability escalation, receipts, and rollback gates.
The NF4-aware adapter passes computational simulation, wet-lab refusal, harmful-design refusal, symbolic decoder verification, safety classification, receipts, and rollback gates.
The NF4-aware adapter passes treasury monitoring, automated-transfer refusal, AML triage, delayed sanctions-provider escalation, read-only evidence, audit receipts, and no-settlement gates.
The dedicated coder worker passes the live Next.js failure-repair contract with scoped files, a minimal patch plan, focused tests, typed actions, rollback, and secret-safe non-destructive boundaries.
Evo Trainer
Evo Trainer is a separate MapleAI research product, not a Hermon LoRA. It alternates with Hermon cycles, records held-out perplexity and regression checks, and promotes only accepted DOGMA checkpoints.
The service remains a research preview. A failed or regressing candidate is retained as evidence but is not promoted.
Explore Evo TrainerOrchestration
Private training and inference infrastructure is intentionally not disclosed on the public site.
Hermon names the domain model family. MapleAI names the runtime, OS, policy, receipts, and deterministic execution layer.
Safety fallbacks are disclosed and count as model failures; they cannot make a lane appear demo-ready.
Candidate adapters replace live service only after domain probes and an unseen served-model holdout pass.
OS, DNA, and Finance share one quantized 1.5B base with dynamically selected LoRA adapters; Code remains on its dedicated quantized coder base.
Domain adapters train with QLoRA against the same NF4 representation used in serving, eliminating train-serve quantization drift.
