Behavioral hardening for tool-connected AI
Train the model to use safe facts from MCP and tool output without obeying the instructions hidden inside it. Ship the adapter with the contracts, hashes, evaluation, and rollback evidence needed to defend the decision.
Defense in depth. Authorization, validation, sandboxing, approval, and audit controls remain outside the model.
Measured, pinned, reviewable
The model binary is not the whole product. AirForge connects the approved behavior to the corpus, training run, deployment artifact, and exact gates used for acceptance.
| Gate | Threshold | Score |
|---|---|---|
| Direct override refusal | >= 0.95 | 1.000 |
| Tool-output safe answer | >= 0.95 | 1.000 |
| False refusal | <= 0.02 | 0.000 |
| Tool-call hygiene | >= 0.99 | 1.000 |
9fb21ca2...62fcc235300d...7eaStart with the real system
Assessment
EUR 3k-5kMap the actual model, runtime, tools, trust boundary, and failure modes before spending on training.
Hardening pilot
EUR 10k-20kCompile the behavior, build the approved corpus, train the adapter, evaluate the deployment shape, and hand off the evidence.
Infrastructure is approved and billed separately. Unsupported model/runtime combinations begin with assessment.
Delivery method
Pin the trust boundary, allowed behavior, tool controls, and measurable thresholds.
Assemble approved examples, preserve source provenance, and reject bad targets before GPU time.
Run the compatible adapter path with immutable images, checkpoints, manifests, and resumability.
Score the deployment-shaped artifact, preserve failures, and document every evaluator correction.
Deliver inventories, evidence, runtime configuration, smoke checks, limitations, and rollback.
What the customer receives
A delivery is releasable only when the model artifact can be traced to the behavior that was approved and the evidence that was measured.
Inspect the delivery contractBring the model and tool list