Athanar develops two ways to complete a task: model calls and learned procedures. The aim is to turn successful runs into repeatable procedures, with evidence for their quality and cost.
Recent benchmark measurements unavailable.
A routing and validation layer. Recorded benchmarks inform model selection, while structural checks detect malformed output. The figures below describe the recorded benchmark sample; results depend on the task, input and available evidence.
Document extraction through the Athanar protocol. Send text to a supported extraction endpoint and receive structured data. See BookPull for supported tasks and input formats.
bookpull.ai →The method should fit the task. Measure results, compare costs, and reuse a procedure when its evidence supports the job. Athanar is being built around that feedback loop: successful work becomes a candidate for repeatable execution, and new tests check whether it holds up.