SELF-HOSTED · NO EGRESS · ONE CONTAINER

Catch the answer that's confidently wrong — before it ships

A wrong AI answer that looks right is the expensive one: it passes every output check, and you find out months later when a decision built on it surfaces. The drift sensor watches the shape of your model's confidence distribution per task class and warns you days before that happens. It runs inside your network. Your data never leaves.

See the live demo → Get a license

The failure it's built for

Most monitoring reads the output: was the answer right? By the time the output looks wrong, the bad answers already shipped. A model under drift doesn't go loud — it goes quiet. It hedges. Its per-field confidence spreads out. The mean still looks fine, so accuracy dashboards stay green while the edge-case tail rots.

The signal you're missing: the distribution flattens before the answer breaks. A sharp, on-domain confidence distribution means the system has an opinion. A flattening one means it has started guessing — while still answering. That gap is where confident-but-wrong lives.

What it is

A small service you deploy next to your pipeline. You pin a baseline per task class, hand it the confidence scores you already compute (or the output text), and it tracks the shape of each class's distribution over a rolling window. Three states per class: stable, watch, degrading.

No egress, by design. The sensor makes no outbound calls to contrarianAI — none. It runs entirely in your VPC. Nothing a carrier or compliance review can object to. (Optional: it can POST a one-line status alert to your own Slack — metadata only, never your data, off unless you turn it on.)

How it deploys

  1. Deploy one container.
    docker run -d -p 8787:8787 -v /your/data:/data contrarianai-sensor
  2. Pin a baseline per task class, from a sample of known-good results.
    POST /pin  { "taskType": "acord-commercial", "referenceText": "..." }
  3. One line in your pipeline. After each inference, hand it the per-field confidence scores you already have.
    POST /observe-scores  { "taskType": "acord-commercial", "scores": [0.97, 0.92, 0.41, 0.88] }
  4. Watch it. Open /dashboard in a browser, ask Claude via the built-in MCP server, or get an optional Slack alert the moment a class crosses into degrading.
    Drop-in adapters for LangChain.js and an MCP server are included; a language-agnostic HTTP sidecar covers everything else.

Want to see the dashboard first? Open the live demo — no install needed.

Get a license

Flat per-deployment license. No metering, no telemetry — because nothing leaves your network. Try it free in EVAL mode first; a key unlocks production use.

Pilot
30 days

One deployment, full features. Prove it on one workload.

Start a pilot
Annual
from $6k
per deployment, billed yearly

One running sensor, one year, all adapters and updates.

Get your license →
Perpetual / Site
one-time

Buy it outright, or license a whole team. Let's scope it.

Talk to us

After checkout you'll receive a signed license key by email. Paste it into SENSOR_LICENSE_KEY and EVAL flips to LICENSED. The key is verified offline — it never phones home.

FAQ

Does any of our data leave our network?
No. The sensor runs in your VPC and makes no outbound calls to us. License keys are verified offline. The only optional outbound is a status alert to your own Slack, which you turn on and which carries metadata only.
What does it actually read?
Per-task-class confidence scores you already compute, or the output/retrieved text. It measures the distribution's shape (alignment + spread) over time — not the content of any single answer.
How hard is integration?
One container and one POST per inference. There are drop-in adapters for LangChain.js and MCP; everything else uses the HTTP sidecar. Most teams are live in an afternoon.
How does the license work?
A signed key you paste into an env var, verified locally against an embedded public key. No license server. Without a key the sensor runs in EVAL mode so you can evaluate it fully before buying.
Is the underlying analyzer open source?
Yes — the core is the MIT-licensed Context Inspector. The sensor is the continuous, stateful production monitor built on top of it.