Services · Evals and tuned models
Test first. Then a model that has to pass the test.
Most model projects start by training. We start by measuring. We build a test from your own past decisions, grade the models you could use today against what actually happened, and show you the gap. You can stop there. If a tuned model is worth building, it has to win that test first, and the weights are yours.
What it is
An eval and, when it earns its place, a model tuned to one decision your organization makes again and again. The eval comes first: a test set built from your records and graded against outcomes wherever they land in public data (volumes against weather, prices, rates, hiring), with the same machinery that grades Palanor's own forecasts in public. Every example carries the world as it was known that day: first-reported figures and publication delays, with no hindsight leaking in. A model trained with hindsight looks brilliant in testing and fails live.
Who it's for
Organizations with a judgment they make often and a record of making it: forecasting a business quantity such as volume, demand, claims or churn; sorting incoming documents with past decisions as the labels; first drafts that must pass a facts check against the evidence; reviewing assumptions against experience.
Format
Three stages. Each ends with a scorecard you keep, and you decide at the end of each whether to go on.
Stages
01
Readiness
Map the decision, build the test set from your records, and grade the models you could use today against it. You get a gap report: what a tuned model would have to beat.
02
Build
Curate the examples, tune an open model, and test it against the eval and against what happened. Delivered to your own cloud or to an endpoint in your account.
03
Care
A monthly regrade against what happened and a watch for drift. A retrained model replaces the current one only if it wins on data it has not seen.
Outcomes
- A scorecard from your own decisions: how the models you could use today do, and what a tuned model has to beat.
- Grades from what actually happened, not only from reviewers, wherever the outcome lands in public data.
- Point-in-time context: each example sees only what was known on its day.
- You own the weights: an open model, running in your cloud or your account.
Your data
How your data is used is agreed in writing before any work starts.