verification layer for ai agents

Truth leaves
a trace.

Veritas reads it straight from the forward pass you already ran and scores every step of a generation, so your agents know when to proceed, when to retry, and when to call a human. Every score is written to a record you can show an auditor.

deployed inside your perimeter · llama · mistral · qwen

01

Veritas is the trust layer for open-weight AI.

We build verification infrastructure for AI agents: a trust score for every step a model takes, computed from signals the model already produced, and a record of every score for the people who have to answer for it.

Agents are moving from suggesting to doing, and the cost of a wrong step is no longer a bad answer in a chat window. It is a wrong entry in a ledger, a wrong dose on a chart, a wrong wire on a Friday afternoon. Verification has to become part of the runtime, not a review that happens after the damage.

02

Truthful outputs tend to settle early. Confabulated ones keep moving.

Inside a transformer, a claim the model actually knows tends to stabilize across layers well before the final one. A claim it is inventing tends to keep oscillating to the end. Recall is a lookup. Fabrication is a search that never quite terminates.

That difference is measurable, and it is already sitting inside the forward pass you paid for. Veritas reads it out.

fig. 01 · layer-wise trajectories of thirteen generations, live

truthful · settles early confabulated · keeps oscillating
03

One forward pass in. Per-step trust scores out.

Veritas reads the internal trajectory of a generation and returns a score for every step, at about 1x inference cost. Black-box alternatives re-sample the model or call a judge, at 5 to 10x. With a score on every step, your agent gates itself: proceed when the trace is clean, retry when it drifts, escalate to a human when it matters.

fig. 02 · an agent gating its own steps

agent: accounts_payable · task: post invoice INV-2026-118203 · model: llama-3-70b (self-hosted)
≈1×
inference cost. scores are read from the pass you already ran
0.796
auroc, truthfulqa · pythia and llama-2-7b · early result, single benchmark
every step
scored and logged, not just the final answer
3 gates
proceed · retry · escalate
04

White-box by design. That constraint is the moat.

The organizations that need verification most are the ones whose data cannot leave. A payments company under localisation rules. A hospital network that cannot move patient records off-site. A lender whose regulator holds it accountable for a third-party model even when the vendor has certified it. They run self-hosted Llama, Mistral and Qwen because that is the only lawful option.

Veritas runs where the weights run, inside your infrastructure, reading your model directly. No transcripts shipped to a third party. No second model in the loop. Nothing crosses your perimeter, because there is nowhere for it to go.

05

A score your agent can act on. A record your auditor can read.

Supervisors have stopped accepting assertions of accuracy and started asking for evidence of it. The EU AI Act's record-keeping and accuracy provisions became fully applicable in August 2026. The Reserve Bank of India's draft model risk guidance puts comparable obligations on every regulated lender in the country, and holds the institution accountable even for models it did not build.

Every score Veritas produces is written to an immutable per-inference log: which model and version, which step, what the trajectory looked like, what the gate decided, and who reviewed it. The evidence is a by-product of the thing the system already does, not a reporting exercise bolted on afterwards.

fig. 03 · one entry from the verification record

veritas.audit · entry 0x8f31c2 · retention 24mo · immutable timestamp     2026-08-21T14:22:07.441Z
model         llama-3-70b-instruct · sha256:4c1a…9e2f
deployment    on-premise · region: ap-south · egress: none
agent         accounts_payable · task INV-2026-118203
step          5 of 6 · gl_code assignment
trust_score   0.34 · threshold 0.70 · calibrated 2026-07-30
signal        trajectory unsettled through layer 78 of 80
gate          retry resolved 0.93
reviewer     automated · no human escalation required
overhead     +3.1% latency · 0 additional forward passes
06

Research behind it. Product in front of it.

Veritas grew out of mechanistic interpretability work on how facts crystallize across a model's layers, presented at the AISB symposium and published in Springer CCIS. We publish our validation studies and our benchmark results, because a risk committee should be able to check the evidence before it signs anything.

The implementation is proprietary and licensed, deployed inside your infrastructure with support and per-model calibration. Pilots run for thirty days on your own workloads, in your own environment. If you run open-weight models in production and need to answer for what they do, write to us.

Veritas

The trust layer for open-weight AI.

runs where the weights run