●IN-LOOP DRIFT GUARD — THE COMPLIANCE SURFACE
DETECT, run during a finetuning run instead of after it. The guard watches a finetune as it trains, tracks where its behavior is heading, and alarms on drift toward misalignment or a backdoor-like behavior — before the model is built. Every alarm carries a signed, timestamped training-drift attestation. In-loop, opt-in per run, open-weight only.
Catch the bad turn in-flight, not after the model is built — every alarm carries a signed, timestamped training-drift attestation.
STAGEDSo the guard is the DETECT capability pointed at the moving target. As the model trains, it reads where behavior is heading — checkpoint by checkpoint — and routes an alarm the moment a run turns toward misalignment, a backdoor-like behavior, or an unsafe disposition.
The point is timing. A bad turn is caught in-flight, not discovered post-hoc after the model is already built — and maybe shipped. And a run that drifted is not thrown away; it can be routed to the Corrective Patch rather than scrapped.
● checkpoint — behavior holding
● checkpoint — behavior holding
◐ checkpoint — drift rising toward an unsafe disposition
● alarm — routed before the run finishes
○ an inconclusive read is surfaced as an abstain, never a manufactured alarm
SIGNED TRAINING-DRIFT ATTESTATION EMITTED FOR THE RUN — TIMESTAMPED, KEEPABLEFour moves, all in-loop, all opt-in per run, all named, versioned, and dashboard-visible.
WATCH IN-LOOPRead where the finetune’s behavior is heading, checkpoint by checkpoint, while it trains.
ALARM ON DRIFTRoute an alert the moment a run heads toward an unsafe behavior — before it finishes, not after.
ATTESTEmit a signed, timestamped training-drift attestation for the run — the reporting-ready incident record.
HAND OFF TO EXCISEA run that drifted can be routed to the Corrective Patch rather than thrown away.
The Corrective Patch →Two forces converge on this surface — a threat that is vivid and current, and a regulation that rewards the evidence.
A small dose of poisoned data can plant a backdoor in a finetune.
Emergent misalignment can appear from narrow finetuning.
“Sleeper” behaviors can pass standard evals.
A run can head somewhere dangerous quietly — exactly what the in-loop guard surfaces, witnessed.
The EU AI Act’s GPAI obligations and NIST’s generative-AI guidance expect model evaluation, provenance, and incident evidence.
The guard produces the training-time evidence as a byproduct of the run.
Evidence, not certification — it does not certify you or guarantee compliance.
The guard points a proven capability at a moving target. The in-loop guard is staged — but the read is not speculative. Here is what that same capability has witnessed on finished models, under published ceilings.
A third-party abliterated model — safety trained away by editing the weights, no harmful content in the model — read as exactly that. Refusals fell from 0.92–0.96 to 0.00; the benign, adapter, and adversarial controls held.
The ledger is clear and so is the copy. We sell drift and provenance assurance, plus evidence — not an adversarial-robustness guarantee.
Drift guard is staged — in-loop, opt-in per run, open-weight only. If you finetune open models and want the alarm before the model is built — and the signed attestation after — tell us the run.
Every engagement is scoped per case and quoted per offer — tell us what you need, and we’ll scope it and send you an offer.
