research · benchmark · jul 18, 2026
Gazelle.
The bespoke model.
General intelligence loses to specific competence. Gazelle does exactly one job — business-rule extraction from legacy code — and does it at a level no general-purpose model has matched on our benchmark.
10 min read · figures from palmbench 2026 · reproducible on your code
business-rule extraction · same production-scale cobol program · same instruction
Why a bespoke model at all?
The frontier labs build for breadth: one model that writes poetry, plans holidays, and reads COBOL — each adequately. But business-rule extraction is not an adequate-is-fine problem. A missed rule is a missed regulation; an invented rule is a defect in your future core. The error budget is close to zero, and general models spend their capacity everywhere except where your estate needs it.
Gazelle inverts the trade. It is trained and tuned for one distribution — legacy enterprise code and the business logic buried in it — and it runs inside Palm Ark's near-deterministic framework, where every answer it gives is grounded in parsed structure and verified against source before it ships. The model is fast and narrow; the framework keeps it honest. That's the design. The framework has its own write-up.
Where the ~65 points come from
The gap between ~30% and ~95% isn't one big miss — it's four compounding ones. On PalmBench we score generic models against the same graded rule set, and their losses cluster in the ways you'd predict:
Gazelle's architecture attacks each directly: grounded extraction sites instead of raw text windows (no context collapse), verification against source (hallucinations don't survive), constrained decoding inside a deterministic pipeline (no drift), and lineage attached to every rule by construction (100% auditable).
The number that matters more than accuracy
Ask Gazelle three times and you get one artifact, one hash. Ask a generic model three times and you get three artifacts — which means your accuracy number is also a random variable. A benchmark score you can't reproduce isn't a score; it's a screenshot. Every figure on this page is reproducible on request, on your code, inside your perimeter, against any model you bring.
What "up to ~95%" honestly means
We publish "up to" because estates differ. Clean structured COBOL with disciplined copybooks scores at the top of the band; estates with decades of dead code and dynamic calls land lower before Palm Key's pre-processing lifts them. We'd rather give you a band you can trust than a point you can't. The ~30% generic-LLM figure is measured the same way, on the same programs, with the same grading.
Bring any model.
We’ll bring the receipts.
PalmBench runs on your code, in your perimeter, against any model you choose.
Run it on your code