Introducing Resolve,
our fast decision-making model.

The decision engine for the enterprise. Evidence-based decisions with measurable confidence.

Request an evaluation
Wherever the work is.
Hosted Private cloud On-premise Air-gapped

Illustration of how Resolve answers. Timing reflects measured median speed.1

Every enterprise runs on decisions.

Millions of them, every year. Each one waits for someone to read the whole file. Resolve reads it for them.

Pay or hold.

Approve or escalate.

Renew or let it lapse.

Who signs off.

0.37s

To read a case file and make the call.1

55 pages
Read in a single pass. Contracts with every amendment. Tickets with their whole history.2
1,000,000
Decisions a day, on a single GPU.2
35×
Faster, at least, than the leading reasoning models, which took 13 to 58 seconds on the same cases.7

It knows when it knows.

0decided by Resolve
0sent to your team

Illustration. Move the threshold to see the split change.

Every answer arrives with a probability. You decide where the bar sits. Resolve takes every case that clears it. Your people take the rest. Automation you set, measure and stand behind.

2 in 3
Decisions taken on its own at about 94% accuracy, calibrated on just 50 of your own past decisions.3
98%
Correct, across the half of its answers where Resolve was most confident.3

It reads like your best people.

Real case files disagree with themselves. The right answer depends on knowing which document counts. Resolve finds it.

The version that governs

Three drafts of the same contract. Resolve works from the one that was signed.

The approval that counts

An approval in an email thread, overridden on page 31. Resolve sees the override.

The fact that changes everything

Change one deciding fact and the answer changes with it.

84%
Fewer errors finding the evidence that decides a case.4
92%
Of the time, it follows a single changed fact to the right answer. Its base model managed 40%.4

Your files never leave the building.

Resolve runs inside your own environment. No calls to outside services. Nothing sent anywhere.

One GPU.
A single server-class GPU runs either edition. No cluster required.
No internet.
Designed to run fully disconnected, inside your network and your controls.
No customer data.
Built entirely on synthetic case files. Your data trains nothing unless you ask us to.

Two editions. One engine.

Both run on the same base, on the same GPU. We test both on your cases and recommend the fit.

Resolve Workflow

For the decisions your business makes all day.

  • Invoices and expenses
  • Contracts and renewals
  • IT operations and access
  • Support tickets and routing

Resolve Evidence

For the files where the documents disagree.

  • Which version governs
  • Which approval counts
  • Which override applies
  • When the answer cannot yet be known

What we built.

Synthetic data generators

Training cases with answers computed by code.

  • Seven kinds of evidence operation across twelve business settings
  • Nine data-quality gates on every batch
  • Screened against our evaluation data, so tests stay clean

Evaluation machinery

Every model, tested the same way.

  • Same items, one pass, each model through its own published code or API
  • Rules and metrics written down before the final runs
  • A locked test set, read once per released model

Post-training

New skills added to an open decision model.

  • Built on Kev-27B, which is based on Qwen3.8-27B
  • Trained on synthetic case files only, with no customer data
  • New decision types learned from about 2,800 examples in about two hours6

Serving system

Fast serving, with accuracy checked against the model we tested.

  • Fused kernels on one H100 GPU: about 0.37 seconds per case file1
  • Served answers compared with the evaluated model: on the fast path 7 of 536 differ, 86.8% against 87.1% (see the comparison)
  • Runs inside your own environment

How Resolve works and how we tested

Measured in the open.

We put Resolve beside leading decision models from some of the world’s largest AI labs and cloud providers. Same case files. Same questions. Every result published. Resolve matches the most accurate of them, and approves the fewest cases that should have been held.

8
Decision models, tested side by side.
536
Enterprise decisions across contracts, finance, IT and support.5
87%
Accuracy, level with the best in the field.5

Read the technical announcement Compare decision models Try the demos

And it learns your rules.

Teach it a new kind of decision with a few thousand examples. Two hours later, it decides like it always knew.

On a new type of decision, after two hours of training on one GPU.6

Resolve

Decide at the speed of your business.

Start with your own cases.

We agree what success looks like with you, run Resolve on a sample of your decisions, and show you every result before you commit.

I’m interested in

We use these details to reply to your request. See our privacy notice.

Notes

  1. Median 0.36 to 0.37 seconds per case file for both editions, served on one NVIDIA H100 GPU, one request at a time, on Dscvry’s enterprise decision benchmark (EDB). The hero animation is an illustration; the case file, question and probabilities shown are examples.
  2. Long-document testing up to about 56,000 tokens, roughly 55 to 60 pages, in a single pass. Throughput of 12.5 to 13 decisions per second on one H100 GPU, equivalent to about one million decisions in 24 hours.
  3. Simulated on Dscvry’s locked test set of 228 case files, with an automation threshold set from 50 labelled questions per workflow. The 98% figure covers the half of answers with the highest confidence (98.4% to 98.5% correct).
  4. Resolve Evidence compared with its base model, Kev-27B, on 644 questions from Dscvry’s own case generator, which was also used in training (in different business settings): 34 errors against 215. On the 102 questions where one governing fact changes, 92.2% against 40.2%.
  5. EDB development set: 536 questions on 90 synthetic case files, written by Claude Opus and checked by Claude Sonnet. Each model answered through its own published code or API with no per-model tuning. Resolve Workflow 87.1%, statistically level with the most accurate models tested. Full method and results in the technical announcement; every model side by side in the comparison.
  6. Resolve Evidence, trained on about 2,800 generated examples in about two hours on one GPU, evaluated on business settings it had not seen, produced by the same case generator.
  7. Median times per case file for general reasoning models answering the same case files through their hosted APIs, against Resolve’s median of 0.37 seconds. Reasoning models answered nearly every case correctly; measured on different hardware and networks.

Resolve is built on Kev-27B by Jared Palmer and Qwen3.8-27B by the Qwen team, both released under the Apache-2.0 licence. Resolve is offered under evaluation terms, with a commercial licence for production, agreed in a contract with each client under our enterprise licensing terms. Dscvry is a trading name of Discovery AI Ltd (United Kingdom) and Redscvry Technology Pvt Ltd (India). All other product names are trademarks of their respective owners.