Coming soon
CMVL
Current language models learn from raw text. They see words in context
and absorb statistical patterns. They can't explain what they learned,
and they can't distinguish correct reasoning from confident guessing.
CMVL is a different approach to model training. The engine produces
structured proofs for each ruling it makes. A model trained
on these proofs learns the answer and the reasoning behind it. Every example carries token-level attribution that no existing training corpus provides.
The idea
Proofs as training data
How models learn today
Billions of parameters are trained on trillions of tokens, with "predict the next word" as the only training signal. The model learns patterns, but the patterns are opaque. When it is wrong, the fix is to retrain on more data and hope; the feedback loop is statistical, slow, and expensive.
How CMVL trains
The FR-OS engine tests inputs and produces structured proofs. Each proof records what was checked, what evidence supported the ruling, and what was ruled out. A model trained on these proofs learns the reasoning process, not just the answer. When the model disagrees with the engine, the disagreement becomes the next training example.
The loop
Engine teaches. Model learns. Engine verifies.
Engine evaluates
FR-OS processes inputs and produces determinations with complete proof records. Each record is a structured training example with full attribution.
Model predicts
A small, efficient model learns to predict what the engine would determine. It runs in milliseconds where the engine runs in seconds. It extends the engine's reach to inputs the engine hasn't seen.
Engine verifies
The engine spot-checks the model's predictions. Agreements confirm the model is learning correctly. Disagreements become new training examples. The model improves continuously without human annotation.
Early results
Certificates beat rewards
Binary reward
Standard approach: the model gets "right" or "wrong" as feedback. In controlled experiments, models trained with binary reward achieved 0% accuracy on held-out evaluation tasks. The signal is too sparse to learn from.
Structured certificates
CMVL approach: the model gets a full proof record showing what was evaluated and why. In the same controlled experiments, models trained with certificates achieved 62% accuracy. The structured signal enables learning where binary reward cannot.
This is not just faster learning. It is qualitatively different. The certificate carries the reasoning structure. The model learns more than how to produce the right answer. It learns the pattern of evidence weighing that produces right answers across domains.
Vision
Domain-agnostic reasoning
The engine already operates across domains: natural language, medical reasoning, policy enforcement. A model trained on the engine's proofs across all domains would learn a general pattern of evidence weighing that transfers to new domains without retraining.
The engine stays authoritative. The model stays fast. Together, they produce a system that reasons verifiably at the speed of a neural network.
Early access
Follow CMVL development
CMVL is in active development. Pilots are operator-led right now. Tell us about your use case and we'll set up a call.