Inline middleware for LLM and agent outputs. Five-verdict drift detection, sub-millisecond evaluation, HMAC receipt chain on every call. Built for AI infrastructure teams whose customers cannot afford a model that quietly stops behaving.
The actual SDM verdict logic — Jaccard trigram drift, Shannon length entropy, character-frequency cosine — runs in your browser below. Submit 30 baseline responses, then push it off-task.
Every LLM completion or agent output passes through SDM. A spectral fingerprint is computed against a baseline, classified into one of five verdicts, and stamped into an HMAC-linked receipt chain. The control plane sees the verdict. The auditor gets the trace.
Spectral fingerprint matches the calibrated baseline within tolerance. Output passes through with a signed receipt. Default verdict for stable production traffic.
First N calls after deployment or model swap. SDM is building the baseline. Outputs pass with a WARMUP-tagged receipt so the downstream system can downgrade trust automatically.
Spectral signature has shifted past the configured threshold. The output still flows but the receipt flags it. Drift Review Board dashboard surfaces the incident with the 12-call live transition trace.
Output entropy has collapsed below the floor — the model is repeating itself, looping, or has degenerated. Receipts mark the call for immediate human review and the operator can wire a hard-block policy.
The spectral fingerprint has moved into a region the baseline never observed. This is the verdict that catches jailbreaks, prompt-injection success, and silent model swaps. Hard-fail by default.
SDM was designed for the layer of the stack that has no good answer today — the moment between the model returning a token stream and the downstream system acting on it. The 12-call live transition trace, sub-millisecond p99, and HMAC chain are the three things every serious AI infra team has asked for and no incumbent ships.
Foundation-model providers and inference platforms that need to detect when a fine-tune, RLHF pass, or quantization quietly broke a production behavior. Drop SDM into the serving path and get the receipt chain for free.
Multi-step agents fail silently. A drifted step five turns ago corrupts everything downstream. SDM stamps each step so the failure is locatable and the bad branch is replayable from receipts.
Retrieval drift, prompt-injection success, and out-of-distribution responses all look like normal completions until they don't. SDM detects the spectral shift in the output itself, not the input pattern.
SDM was designed against named buyer profiles — AI infrastructure teams, agent platforms, and developer-tools companies running inference at scale.
Design-partner program open for AI infrastructure teams and agent platforms. Includes deployment in your serving path, baseline calibration on your traffic, dashboard onboarding, and a live transition trace from your own production calls.