01HOW IT WORKS

Sits beside your telephony.
Never in the call path.

01 · INGEST

Media-stream tap

Authenticated tap (Twilio today; SIPREC-style patterns in design). Audio stays tenant-scoped, under your retention policy.

02 · SIGNALS

Condition context

13 deterministic quality features per segment — the condition context for calibration.

03 · DETECTION

Three families

Three independent detector families score every 4-second window.

04 · FUSION

One calibrated verdict

Versioned fusion: score, band, confidence, plain-language reason codes — per window and per call.

05 · ADVISORY

A human acts

Your policy maps verdicts to recommendations — step-up verification, supervisor review, fraud routing — as signed webhooks. Every override is audited.

DEPLOYMENT

Drop-in observer

No changes to call routing.

FAILURE MODE

Fail open

If detection degrades, your calls keep flowing — verdicts carry documented degraded-mode labels.

POSTURE

Advisory

Norvin recommends; it never terminates or blocks a call on its own.

WHY THREE FAMILIES

Under real telephony conditioning, our two strongest models fail in opposite directions — one stays clean on humans but misses synthetics, the other catches everything but over-flags humans. A third adjudicates. The ensemble isn't marketing; on the evidence, no family is dispensable.

02THE EVIDENCE PACK

The evidence, not just the verdict.

Every verdict ships with per-family scores, fusion arithmetic, model versions, window bounds, and plain-language reason codes. Evidence exports are content-hashed JSON — reproducible artifacts for auditors, disputes, and regulators. An analyst reviewing an alert sees why, not just what.

SPECTROGRAM · CALLER LEGILLUSTRATIVE
0:000:16 ◂ flag0:36 ▸ flag end0:60
Embedded synthetic span · high band · windows 04–14
EVIDENCE EXPORTILLUSTRATIVE
{
  "call_id":      "HARNESS-8f3c",
  "call_verdict": "high",
  "confidence":   0.91,
  "windows":      20,
  "families": {
    "clean_path":  { "score": 0.31, "v": "1.4.0" },
    "high_recall": { "score": 0.88, "v": "1.4.0" },
    "adjudicator": { "score": 0.79, "v": "1.2.1" }
  },
  "fusion":       "calibrated/2026.02",
  "reason_codes": ["cadence-regularity",
                   "spectral-artifact"],
  "condition":    "telco/8k",
  "content_hash": "sha256:7e3a…c81",
  "posture":      "advisory"
}
Tamper-evident

Content-hashed exports, verifiable offline.

Your audio stays yours

You set retention; metadata-only mode stores no audio at all. We never train on it.

No transcripts, no embeddings

Detectors output scalar scores and versions. Nothing to leak.

Audit-ready

Overrides, policy changes, and audio access are all audit events.

03DEPLOYMENT & INTEGRATIONS

Runs where the audio lives.

Single-host package — app, workers, storage, three containerized detectors — proven to run from images alone. CPU-only, no GPU. Deploy in your environment so audio never has to leave it, or vendor-managed if you prefer. Tenant-isolated at every persisted row.

Twilio Media Streams
SUPPORTED TODAY
Generic WebSocket streams
SUPPORTED TODAY
Signed webhooks out
SUPPORTED TODAY

HMAC-verified advisory recommendations into your alerting and workflow tools.

SIPREC-style tap
IN DESIGN