BackContext is live: one context layer for every agent you run.
Explore BackContextPublic benchmarks tell you which model is strongest in general. They cannot tell you which model gets your order desk's abbreviations right. For that you need evals built from your own verified outcomes, and a gate that blocks any release that gets worse on them.
Leaderboards are useful for picking a starting point. They test reasoning, coding and general knowledge on public data. Your tasks are narrower and stranger: a seniority convention one staffing client uses, a lot-code format a seed distributor adopted years ago, an abbreviation everyone on the order desk knows and nobody wrote down. No public benchmark contains those, so no public score predicts how a model will do on them.
The best label is what actually happened. The candidate was placed. The order shipped with these line items. The claim was approved. BackContext links each recorded task to its outcome, and only verified outcomes enter an eval set. We do not use a model to grade another model, and we do not ask annotators to guess what the right answer would have been.
Every candidate model runs side by side against Claude Opus 5.5, Claude Fable 5.1, GPT-6 Astra and GPT-6.1 Sol on the same held-out set. The report tracks task accuracy, cost per 1,000 tasks, cost per 1,000 successful tasks, median latency and fallback rate, with a breakdown of errors by slice so you can see where each model fails, not just how often.
Averaged across our production customers, fine-tuned models score 97.4% on held-out tasks, against 93.1% for Fable 5.1, 92.6% for GPT-6 Astra, 91.8% for Opus 5.5 and 88.9% for GPT-6.1 Sol.
An eval set is only useful if it can stop a release. Each new model version has to match or beat the current production model on every tracked slice before it ships. If version three is better overall but worse on PDF orders from one region, it does not go out until that slice is fixed. The gate runs automatically, and its result is part of the release record.
On Hytech Seeds' held-out order set, Opus 5.5 scored 91.8%. Its errors clustered in three places: lot codes, regional abbreviations, and variety names that differ by a single character. hytech-orders-v2, trained on 206K order records verified against what shipped, scored 96.7%. A public benchmark would have ranked Opus 5.5 far ahead. On the work that pays Hytech's bills, it was behind.
Smaller than most people expect. At around 95% accuracy, a slice of 1,000 examples gives a margin of roughly ±1.4 points at 95% confidence, which is enough to separate a fine-tuned model from the frontier models in our results. A few hundred examples per slice will catch large regressions. We size each slice to the decision it has to support.
Slices matter more than the headline number. We split each test set by input type, source and difficulty, such as email versus PDF orders, or junior versus senior requisitions. A model that averages 97% can still fail badly on one slice that matters to the business, and only a sliced report shows it.
We build the eval set first, before any training. Sometimes the answer is that a frontier model already does well enough on a task at a cost you can live with. You should know that before paying for a fine-tune, and an honest eval built from your own outcomes is the only way to find out. See how our evaluation work runs.
