Triage has raised $1.5 million in pre-seed funding led by BoxGroup, with participation from Precursor Ventures and angels including Zach Lloyd, Michael Fertik, Bill Shope, Niklas de la Motte, and Cory Levy.
We’re building an applied AI lab around a simple conviction. Institutions should retain authority over the intelligence they put to work.
Every organization has its own judgment about how work should be done. What constitutes sufficient evidence. Which information a decision should consider. When an exception is warranted. Where discretion ends and human approval becomes necessary. As AI takes on more consequential work, those judgments need to remain in force across the decisions it makes.
We call this institutional alignment.
A model’s general capabilities are only the starting point. Deployment places those capabilities inside an organization with particular obligations, incentives, and standards. A financial institution, a law firm, and an engineering organization can use the same underlying model while requiring materially different behavior. Each needs a way to make its own judgment operational and preserve it as models and tasks change.
Triage’s work in downstream alignment addresses that problem. We study how behavior diverges during execution and how to correct it while preserving the work already accomplished. Our approach, defensive interpretability, uses evidence from a model’s context, reasoning, actions, and outputs to inform intervention.
Integrity brings that research into production. An organization defines its constitution, the policies and standards its AI should follow. Integrity evaluates behavior against it across frontier, open-source, and bespoke models. When a run materially diverges, Integrity steers it back toward that constitution so useful work can continue. When safe continuation is unavailable, it can stop or escalate the run.
That judgment needs to extend across the task. A research agent can pursue an increasingly weak hypothesis through individually reasonable experiments. A coding agent can produce a working implementation while gradually departing from the specification. Understanding the accumulated evidence makes it possible to recognize when locally plausible decisions are taking the work off course.
Getting started is straightforward. Teams connect their existing model calls through Integrity and define the parameters of their judgment. Feedback on its assessments and interventions makes that judgment more specific to the organization over time. Integrity works with existing routers and agent harnesses, allowing teams to retain the systems they already use.
Our work on Secure Agents complements this by limiting the tools and data an agent can reach to what its task requires.
The economic stakes follow directly. Drift consumes inference, creates rework, and demands human supervision. Correcting it during execution can preserve useful progress and make longer periods of autonomy practical. Our measure of success is useful work completed per dollar of inference and supervision.
This also gives institutions greater freedom in model selection. An organization’s constitution should remain durable as it adopts a more capable frontier model, an economical open-source model, or a model trained for its own domain. Triage’s ambition is to make that continuity possible.
The funding supports our research in divergence detection, adaptive steering, and deployment-specific judgment, alongside the infrastructure required to put that work in customers’ hands.
We’re building toward safe and economical autonomy that remains faithful to the institution it serves.
To work with us, reach out at info@triage-sec.com.