Capability · Mission-specific evaluation and release gates

GAVEL

Prove your AI agents perform, and keep them improving.

GAVEL builds mission-specific evaluation suites the program owns: tasks, rubrics, automated checks and release gates drawn from real workflows and expert judgment. Compare models, vendors and releases on your own work, and catch regressions before deployment.

Talk to us about a pilot All capabilities

How GAVEL works

Each letter is a step. The green one is a person.

GGround

Builds tasks from your real workflows.

AAuthor

Writes rubrics and automated checks.

VValidate

Your experts confirm the ground truth.

EEvaluate

Grades every model, vendor and release.

LLearn

Gates the release and feeds what failed back in.

Where the same engine re-points

Benchmark: model and vendor comparison
Benchmark: acceptance testing
Benchmark: configuration sweeps
Benchmark: repeat-reliability measurement
Benchmark: pre-release regression gates
Production: scoring real traces
Production: drift detection
Production: routing uncertain cases for review
Production: adding confirmed incidents to the suite

What it does for you

Measured on your work, not on a demo.

You are swapping models and approving vendor systems faster than you can prove they work. GAVEL turns your real workflows and your experts’ judgment into an eval suite the program owns: tasks, rubrics, automated checks and release gates. Every candidate runs through the same suite, regressions surface before deployment instead of after, and the next swap takes an hour instead of an afternoon. Your experts set the bar once and spend an hour or two a week keeping it current.

What the suite reports

% PASSMission successTasks clearing the SME-defined mission bar

PASS^kRepeat reliabilityRight every time, across every repeated trial

# / 100Critical failure rateHigh-severity failures, counted separately, never averaged in

$ / PASSCost per successWhat one correct outcome costs, not one run

What the engagement asks of you

20 to 30 hrsSME time up frontAcross the team, to set the mission bar

1 to 2 hrs/wkSME time to sustainReviewing flagged cases

60 daysPilotOne use case, held to kappa 0.7 on held-out data

The platform in production today

Not a prototype. The same engine, already fielded.

Hundreds of thousandsLive calls

100+Businesses on the platform

Several hundredLocations

99.95%Control-plane availability

The platform behind Northmark runs in production today, across several hundred locations for more than 100 businesses. It has handled hundreds of thousands of live calls, with 99.95% control-plane availability.

Talk to us about a pilot sized to one job, accepted against your PWS.

Available through teaming partners on OASIS+ and GSA MAS.

Talk to us about a pilot