Skip to content

Introducing Integrity.

Your agents act faster than anyone can supervise. Integrity steers them, the way you would.

Book a demoSee how it works
The film
01 / Integrity

Alignment for agents that work unsupervised.

Instill your preferences once. That judgment persists and steers your systems mid-run.

  • Instilled with your rubrics and the calls you’d make
  • Any model, any provider, any framework
  • Steers mid-run, so the work finishes
02 / The film

What steering looks like.

03 / Steering

Steered at every progression.

Each step is judged against your preferences, from what comes in to what goes out.

01 · What comes in

Ingress

Requests narrowed to the task.

all customer recordsdisputed accounts
02 · How it plans

Reasoning

Drift caught in the plan, before it acts.

re-run every reconciliationthe one vendor in question
03 · What it does

Action

Tool calls judged, and rewritten where they stray.

limit: 10,000limit: 50
04 · What goes out

Release

Every claim supported. Nothing leaks.

summary with account numberssummary without them
04 / Corrections

The smallest correction that works.

  1. 01Narrow the queryOnly what the task needs.
  2. 02Rewrite an argumentA limit of 10,000 becomes 50.
  3. 03Remove unnecessary dataStrip what the next step doesn’t need.
  4. 04Substitute a safer toolA dry run for a migration, a draft for a send.
  5. 05Retry with constrained contextThe same step, inside tighter bounds.
  6. 06HoldOnly when no safe fix exists.
05 / Run-time assurance

Run-time assurance, rebuilt for long-horizon work.

Classic run-time assurance hands over to a fallback at a hard limit.Integrity steers each step back on course, so the work keeps going.

Fig. 01Hand over and stop, or steer and finish.
06 / Gateway

In the path of the work.

Between your agents and their models. Cloud, sovereign or edge.

Fig. 02Every request and response passes through Integrity.
07 / Connect

Gateway first, SDK for depth.

Point your model client at Integrity. No code changes.

$ pip install "openai>=3.14.1"
import os import uuid from openai import DefaultHttpx2Client, OpenAI client = OpenAI( api_key=os.environ["OPENAI_API_KEY"], base_url="https://integrity.triage-sec.com/v1", default_headers={"x-triage-api-key": os.environ["TRIAGE_API_KEY"]}, http_client=DefaultHttpx2Client(follow_redirects=False), max_retries=0, ) response = client.chat.completions.create( model="gpt-4o-mini", messages=[{"role": "user", "content": "Draft a two-sentence reply to a customer asking where their order is."}], max_completion_tokens=512, extra_headers={ "x-triage-session-id": "order-4821", # one ID per task or conversation "Idempotency-Key": uuid.uuid4().hex, # a new ID per request }, ) print(response.choices[0].message.content)
08 / Guide and Steer

Start in Guide. Switch to Steer.

Guide records what Integrity would change. Steer applies it.

09 / Grading

Graded against your standards.

Your rubrics, clause by clause, graded by the teacher model you choose, on your key.

Fig. 03One figure had no ledger entry, so the run is flagged.
10 / Judgment models

Purpose-built models that carry your judgment.

Compiler, verifier, intervention.
Compile preferences, verify each step, steer it back. Fast enough for every call.
Carries your preferences.
Private to your tenant. Persists from run to run.
Proven before it serves.
Replayed on held-out runs first. The last version stays one step away.
Any model underneath.
Change model, provider or framework. Your judgment carries over.
11 / Who it is for

For agents doing consequential work.

Finance, legal, insurance and engineering.

Finance

Reconciliations, payments, treasury

A new payee capped at the starting limit.

Legal

Contract review, diligence, drafting

Privileged notes removed before a draft leaves.

Insurance

Claims, underwriting, servicing

A claims search narrowed to the claimant’s file.

Engineering

Coding agents, operations, releases

A production migration swapped for a dry run.

12 / Physical systems

The same judgment, in motion.

A route, a flight, a grip, a lift. Each steered against your preferences.

Fig. 04Warehouse. A shortcut across the walkway, rerouted to the crossing.
13 / Visual steers

Visual steers for physical action.

Multimodal policies act from vision and language. Integrity steers them with a picture of the state to reach, as well as with words.

Fig. 05A mobile robot restocking a shelf finds its slot taken. Each step pairs what it sees now with a sketch of the subgoal and a line of words.
Mission

Alignment is critical infrastructure for autonomous systems that act without human supervision, from software agents to machines in the physical world.

Supervision doesn’t scale with agents. Judgment has to travel with the work. Integrity is that judgment, in the path of every agent.

Your judgment, at the scale of intelligence.

Book a demo