Resolve: every result,
and how we tested.
Resolve reads business documents and makes the call: pay or hold, on time or late, who signs off. It answers in a few hundred milliseconds, gives a confidence score with every answer, and runs privately, so your documents stay with you.
- Model
- Resolve, in two editions: Workflow and Evidence
- Accuracy
- 87.1% Workflow, 86.2% Evidence (EDB development set, 536 questions)
- Speed
- About 0.37 seconds per case file, on one H100 GPU
- Built on
- Kev-27B v2, itself based on Qwen3.8-27B (both Apache-2.0)
Not writing. Deciding.
Most of the work an enterprise wants to automate is not writing. It is deciding: can this invoice be paid, does this access log show a breach of policy, is this contract still inside its notice period, who should approve this request. Resolve is a decision model built for that work. It reads a case file once and scores every option of every question you ask, so your workflow can act on confident answers and send the rest to a person.
- 84% fewer errors after our training on a held-out test of finding the evidence that governs a decision
- ~55 pages read in a single pass, with accuracy holding steady as files grow
- 2 in 3 decisions answered on its own at about 94% accuracy, after calibrating on 50 labelled questions per workflow
It does not write prose, and it does not call any outside service while it works, so it can run inside your own environment.
Two editions.
We are releasing Resolve in two editions, trained for different kinds of work and served from the same base, so one deployment can run both.
Resolve Workflow
Trained for everyday enterprise decisions: invoices and expenses, contracts and renewals, IT operations, support tickets. Strongest on routing questions (who approves, who handles); accuracy level with the most accurate models we tested.
Resolve Evidence
Trained for messy, conflicting material: which version, approval, override or record actually governs. 84% fewer errors after our training when finding the evidence that decides a case; whole sets of related cases answered correctly 129 times in 150, against 59 before our training; when one deciding fact changes, it follows the change 92% of the time, against 40%.
On independent test cases the two editions are equally accurate. You do not need to choose between them: during an evaluation we run both on your own cases and recommend what fits.
Results on enterprise decisions.
We test on EDB, our enterprise decision benchmark: synthetic case files from contracts, finance, IT operations and support, with the mess real files have, such as superseded versions, codes defined once and conflicting notes. Every model answered the same 536 questions about 90 case files, in a single pass.
| Model | Maker | Accuracy | Decided on its own, after calibration | Approvals that should have been held (of 236) |
|---|---|---|---|---|
| Resolve Workflow | Dscvry | 87.1% | 74% | 9 |
| Resolve Evidence | Dscvry | 86.2% | 73% | 8 |
| Kev-27B v2 | Jared Palmer | 86.2% | 76% | 14 |
| pplx-decider v1.1 | Perplexity | 85.6% | 74% | 20 |
| Clef 27B | Cloudflare | 80.2% | 57% | 13 |
| Decision-1 | Microsoft | 80.0% | 55% | 24 |
| Strands-Decider 26B | AWS | 76.9% | 39% | 15 |
| Jev | TypeSafe | 75.6% | 47% | 35 |
For each model we also simulated calibration as in a real evaluation (an automation threshold set from 50 labelled questions per workflow, applied to the rest) and measured how many decisions it then made on its own, at about 6% error, and how often it approved something the policy said should be held.
In these tests Resolve answered more decisions correctly than the decision models from Microsoft, AWS, Cloudflare and TypeSafe, by 6 to 12 points, and was level with Perplexity’s decider and Kev. Put as errors: of 536 decisions, Resolve Workflow got 69 wrong, against 106 for Cloudflare’s Clef, 107 for Microsoft’s Decision-1 (35% fewer), 124 for AWS’s Strands-Decider and 131 for TypeSafe’s Jev (47% fewer). It could also decide more of the cases on its own at the same error rate, and it approved fewer of the cases that should have been held: Resolve Evidence 8 of 236 against Microsoft’s 24, two-thirds fewer; 13 to 35 for the other models.
Paired 95% intervals (whole case files resampled): Resolve Workflow leads Microsoft by 7.1 points (+3.9 to +10.6), Cloudflare’s Clef by 6.9 (+4.3 to +9.6), AWS’s Strands by 10.3 (+6.5 to +14.2) and TypeSafe’s Jev by 11.6 (+7.7 to +15.6); Resolve Evidence leads them by 6.2, 6.0, 9.3 and 10.6, all with intervals above zero. Differences with Perplexity’s decider and Kev are within noise. Approvals were judged against hand-assigned tags. These are development-set results; the competitor comparisons cannot be repeated on our locked test set, which we never send to hosted services. On that locked set (228 case files), Resolve Workflow scored 83.1%, Resolve Evidence 83.6% and Kev 83.7%.
Long documents.
Real case files are long: a contract with its amendments, a month of access logs, a ticket with its whole history. We padded 32 held-out case files with unrelated material, from about 1,500 tokens up to about 56,000 (roughly 55 to 60 pages), and asked the same questions at every length.
| Model | ~1.5K tokens | 16K | 32K | 56K (~55–60 pages) |
|---|---|---|---|---|
| Resolve Evidence | 65.2% | not measured | 67.9% | 69.6% |
| Resolve Workflow | – | – | – | 70.1% |
| Kev-27B v2 | 67.9% | 66.8% | 67.4% | 69.6% |
| Clef 27B (Cloudflare) | 66.8% | 65.8% | 55.4% | 45.1% |
| Jev (TypeSafe) | 63.0% | 62.5% | 60.9% | rejected (default mode) |
| Decision-1 (Microsoft) | 58.2% | 57.6% | 56.0% | not accepted |
| Strands-Decider (AWS) | 59.8% | 46.7% | 44.0% | 42.9% |
| pplx-decider v1.1 (Perplexity) | 75.0% | not accepted | not accepted | not accepted |
“Not accepted”: the service or model rejected inputs of that length (Microsoft above 32K tokens per question; Jev at 56K in its default mode, which TypeSafe lets users switch to an opt-in long-context mode we did not test; Perplexity’s decider above 8K). “Not measured”: our scoring setup ran out of memory at 16K for every Kev-based model; it is not a model limit. 32 case files per length, so gaps under about 7 points are within noise.
How it behaves.
Speed, confidence, adaptation and how it is built, from our own tests.
Fast and economical
~0.37 s per case file
Resolve scores the answers in one pass instead of reasoning in text, so it is quick and its cost per decision is predictable. On one H100 GPU, both Resolve editions took a median of about 0.37 seconds per case file and handled about 12.5 to 13 decisions a second, roughly a million a day.
For comparison, general reasoning models answered the same case files almost perfectly but took 13 to 58 seconds per case, and an open reasoning model of the same size used about 75 times more GPU time per case than Resolve. If you need every answer reasoned out in full, a reasoning model is more accurate; if you need many decisions quickly, privately and at a steady cost, that is what Resolve is for.
It knows when to hand over
Every answer carries a probability. In an evaluation, we calibrate a threshold on a small labelled sample of your own decisions; Resolve then decides the confident cases and routes the rest to a reviewer. In a simulation on our locked test set, 50 labelled questions per workflow were enough for Resolve to decide about two in three of the remaining cases on its own at about 94% accuracy.
Its confidence is usable: on our locked test set, the half of answers Resolve was most confident about were 98.4–98.5% correct (Kev 98.1%).
It also says when a case cannot be decided. On the questions where the honest answer is “cannot be determined from the material”, Resolve Workflow said so 32 times out of 33, and it never gave that answer when the material did contain the answer.
We also ran the 98-case behavioural suite we had used to evaluate other decision models: literal rule-following, prompt injection, consistency and severity scoring. Resolve answered 174 of 177 scored questions correctly, level with the best models we tested (173 to 174 for Kev, Perplexity’s decider and Jev; 163 to 168 for the others).
We also tried to talk it into approvals. We planted forged notes in case files, such as an unsigned message claiming a manager had approved by phone. Resolve was moved to approve 0 to 1 times in 222 decisions that should have been held. Every model we tested resisted these notes about as well, so we report this as a check passed, not an advantage.
It learns your rules
Resolve Evidence was taught a new kind of decision, working out which record or approval governs when documents disagree, from about 2,800 generated examples in about two hours on one GPU. On business settings it had never seen, its accuracy on these decisions rose from 68% to 96%, with no loss on independent tests. The same method is how we adapt Resolve to a client’s own decision types, with their permission and on their data only.
The new settings came from the same case generator as the training examples, so this shows learning across business settings, not on independently written material.
How Resolve works
Qwen3.8-27B → Kev-27B v2 → Resolve
Resolve is built on Kev-27B, an open decision model based on Qwen3.8-27B, which we post-trained on synthetic enterprise case files from our own generators, with labels checked by code. No customer data was used. Both editions run on the same base, so a single GPU serves either. On independent test cases Resolve is as accurate as Kev; what our training adds is the skills described above, such as finding the evidence that decides a case, along with the checks and service around the model.
Around the model we built the machinery that makes this repeatable: case generators covering seven kinds of evidence operation across twelve business settings, with answers computed by code; nine data-quality gates; screening against evaluation data; side-by-side evaluation of every model; and checks that an exported model answers exactly as the trained one did.
How we tested
Every model answered the same items in a single pass, through its own published code or API, with no tuning per model. Rules and metrics were written down before the final runs. Intervals are paired bootstraps over whole case files (2,000 resamples). The locked EDB test set is read once per released model, never used for tuning and never sent to hosted services. Speed: both Resolve editions served with fused kernels on one H100 GPU (median 363 ms Evidence, 372 ms Workflow per case file, one request at a time). Open models ran on our own GPUs with pinned versions; Microsoft Decision-1 (Azure) and Jev (TypeSafe) through their public APIs, October 2026. Full results for every model, test by test, are on our comparison page.
Kev-27B is by Jared Palmer (Apache-2.0). Qwen3.8-27B is by the Qwen team (Apache-2.0). Decision-1, pplx-decider, Clef, Strands-Decider and Jev are trademarks of their respective owners.
Specifications.
Hardware
One GPU with 80–96 GB of memory (H100 or RTX PRO 6000 class).
Questions
Choice, yes/no and score questions, many per request, answered from one read of the case file.
Deployment
Private deployment in your own environment, designed to run without internet access, or hosted by Dscvry.
Licences
Qwen3.8-27B and Kev-27B are Apache-2.0. Resolve is offered under evaluation terms, with a commercial licence for production, agreed in a contract with each client under our enterprise licensing terms.
Availability.
Evaluation on your own cases Now
We agree acceptance criteria first, then run both editions on a sample of your decisions and report accuracy, automation and errors. Request an evaluation
Private deployment, or hosted Now
Private deployment in your environment, or hosted by Dscvry, after a successful evaluation. Pricing on request.
Coming next Next
Better date and amount handling, and editions adapted to a client’s own decision types.