Latent Contact

Score every output.
No judge model.
No added latency.

Backed by Y Combinator

Latent reads your model’s internal activations while it generates and scores whether the answer is about to be wrong. High risk gets interrupted, escalated, or replaced. For teams that hold their own weights or run their own inference.

Real calibration run on public filings. No signup, nothing to install. Security and deployment detail →

token-level risk Llama-3.1-8B · layer 29 generating
prompt

Replay of a real held-out generation from the shipped trace set, shaded by the per-token risk the probe assigned. Drag the threshold control below to change which tokens escalate.

at 0.62 · 0 of 0 flagged
95%of what it surfaces at a 10% review budget is a real failure. Measured on thousands of expert-labeled outputs
2.7×more failures caught at a 10% review budget than reviewing a random 10%, on the same outputs
0 msof measurable added latency, with every output scored
0 bytesleave your environment in the serving path. No weight custody, no inbound connections, no telemetry
The gap

A fluent wrong answer leaves no trace in the text.

Output-level checking asks a second model, a search index, or a person to re-read what was already written. That works when the mistake is visible on the page. It fails on the case that costs you money: an answer that reads correctly, asserts nothing obviously false, and is wrong.

It also fails structurally when the field needed to verify an answer is the field you are required to strip before sending it anywhere. Once the patient age or the account identifier is redacted, the checker no longer holds the thing that decides whether the answer was right. Nobody notices, because the checker still returns a confident pass.

The model’s own internal state carries that signal while it writes the answer. Latent reads it where it already exists, inside your serving process.

How it works

One line into vLLM, then a control you can actually staff.

01

Install

Registers through vLLM’s standard general_plugins entry point. No fork, no patched engine, no change to how you serve. A read-only forward hook on the scored layer.

02

Calibrate

Fits on your model and your traffic, on your hardware. Selects the layer, the probe, and a threshold matched to the review budget you can actually staff. The artifact stays local. With no labels at all, a probe trained on synthetic examples already reached 85% precision at a 10% review budget on real traffic.

03

Operate

Risk-ranked escalation queue, token-level traces, policy tied to review capacity, and an append-only audit record. Reviewer decisions feed the next recalibration.

Probes are per model and per domain. A fine-tune, a version bump, or a shift in traffic changes what the correct escalation threshold is.

The console

What your team operates.

The escalation queue, ranked by risk. These are real rows from the calibration run. Each flag points at the sentence most likely to be wrong.

Latent · reliability console open the live console ↗
Needs reviewhighest risk first

    Scoring held-out traffic…

    Where it runs

    Inside the forward pass, inside your environment.

    Output-level checking reads the finished text, which is why it goes blind the moment the field it needs is redacted. Latent reads the layer while the answer is still being written, and routes on what it finds. Sending only the riskiest 10% to a frontier judge costs a tenth as much as judging everything.

    policy on flag
    prompt

    forward pass
    probe · L31
    route
    ship
    flag
    output

     

    Runs in your VPC or on your own hardware
    No weight custody, ever
    Read-only hook, generation unmodified by default
    No inbound connections, no ingress rule
    Calibration on your hardware, artifact stays local
    Prompts and outputs recorded as hashes by default
    Who this is for

    Anyone who holds the weights.

    Agents that run long without a human in the loop

    A wrong step early is compounded by every step after it. Scoring each generation as it happens gives the loop something to halt or retry on before the error propagates.

    Teams shipping open-weight models into production

    Llama, Qwen, Mistral and the rest, fine-tuned or served as they are. Hallucination is the failure everyone names first, and output-level checks cannot see a confident wrong answer that reads perfectly well.

    Models that keep learning after deployment

    Continual fine-tuning moves behaviour between evals. A probe recalibrated against recent traffic tracks the model you are actually running, not the one you last benchmarked.

    AI vendors deploying into customer environments

    The model sits inside the customer's infrastructure with no telemetry coming back, so quality has to be measurable from inside the boundary.

    Regulated enterprises running their own models

    Banks, insurers and health systems, where a reviewer queue is a headcount line and an audit record is a supervisory expectation.

    Air-gapped and classified deployments

    Environments where a managed API is not an option at all, and every control has to run inside the boundary.

    Start with a scoped pilot.

    One model, one review workflow, calibrated on your representative traffic. Runs on a sandbox model and never touches production, which is scoped to clear as an evaluation rather than a full third-party risk assessment.