Evidence and Enforcement Engine

Turn optimization experiments into controlled deployment policy.

Give engineering teams immutable system identities, experiment lineage, exact finite-sample risk reporting, certificate lifecycle management, failure attribution, audit history, metering, quarantine, recertification, and safe baseline fallback. Prevent a result from one model, workload, backend, or hardware configuration from silently authorizing a different deployment.

Control plane · Django

Make every approval auditable.

Identities, experiments, immutable evidence snapshots, exact one-sided risk bounds, certificates, risk attribution, and append-only audit history. Worker-based asynchronous execution supports the lifecycle.

Runtime · OpenAI-compatible FastAPI

Keep live execution inside its scope.

Certificate lookup, optimized execution, metering, and request attribution. Quarantine, revocation, rollback, and recertification govern deployment. A scope mismatch or expired evidence triggers fail-closed full-baseline fallback.

The same infrastructure can govern a fixed recipe, an adaptive router, a training selector, or a future policy. Each policy needs its own measured evidence. Integration of the foundation does not establish hosted availability for every optimization.

Start with a single prompt.

The public demo accepts {"prompt":"Explain KV caching."} at POST /v1/demo/inference. It returns a model answer, token counts, and gateway response time through the configured inference runtime.

Try the executable Python and JavaScript snippets on Home, or copy the cURL request. This request demonstrates the API; certifying savings requires a separate workload comparison.

Run the inference example

Explore the research training integration

Research integration example

Explore the training integration.

This engineering example records an exact model revision, workload, hashes, counts, timing, and scalar metrics. The model and dataset stay in your environment. A complete customer-facing production training service is not confirmed; this integration does not establish training savings.

The Training & research dashboard includes this example, authenticated FastAPI instructions, and the saved ML-T6R experiment. ML-T6R is a separate study from the ML-T6 result on Products; it does not certify reduced training for new jobs.

Hugging Face training SDKRun an exact-scope training job

FastAPI training API

Python · Hugging Face TrainerRequests an exact-scope plan. Without approved evidence, training stays on the full baseline.

import os
from reduction_states import ReductionConfig, ReductionTrainer

trainer = ReductionTrainer(
    model=model,
    args=training_args,
    train_dataset=train_data,
    eval_dataset=eval_data,
    reduction=ReductionConfig(
        api_base_url="https://api.reductionofstates.com",
        api_key=os.environ["REDUCTION_API_KEY"],
        model_id=os.environ["MODEL_ID"],
        model_revision=os.environ["MODEL_REVISION"],
        dataset_sha256=os.environ["DATASET_SHA256"],
        hardware_target=os.environ["HARDWARE_TARGET"],
        quality_metric="eval_loss",
    ),
)
trainer.train()

https://api.reductionofstates.com/api/v1/training/jobs

Before running: make this repository’s packages directory available on PYTHONPATH and install PyTorch, Hugging Face Transformers, and its Trainer dependencies in your training environment. Initialize model, training_args, train_data, and eval_data locally. Set REDUCTION_API_KEY to a product key with training:write; set MODEL_ID and MODEL_REVISION to the exact imported revision loaded by your model, DATASET_SHA256 to the SHA-256 of a frozen dataset manifest, and HARDWARE_TARGET to your actual hardware. Use the gateway root as api_base_url; the SDK adds the training route.

Put your workload to the test.

Discuss a scoped assessment and find out which configurations are worth deploying.

Discuss a private pilot