Research
Experiments, derivations and cost models behind our recommendations. Every result links to the code that produced it, and every post states its assumptions and limitations.
Experiment
LoRA rank vs full fine-tuning: an experiment on structured extraction
Eight training runs of a 0.5B model, in-distribution and on unseen formats.
Oct 30, 2026SimulationYour LLM judge is shrinking your results
Judge error attenuates real differences; the Rogan–Gladen correction.
Oct 27, 2026Cost modelSelf-hosting vs paying per token: the break-even model
A parametric model with a sensitivity analysis of every input.
Oct 25, 2026Cost modelFine-tuning compute: FLOPs, GPU-hours and cost per run
6N vs 4N, MFU, and where the money really goes.
Oct 23, 2026CalculatorHow much GPU memory does fine-tuning need?
Full, LoRA and QLoRA from eight real configs.
Oct 21, 2026SimulationHow many evaluation examples do you need?
Power analysis, paired vs unpaired designs.
Oct 18, 2026DerivationDPO from first principles, and what β really does
The derivation plus a tabular simulation with a known optimum.
Oct 14, 2026How we publish
- Real runs only. Every number in a research post comes from code in our repository that we actually ran.
- Sources verified. We cite papers and primary documents, and link them.
- Limitations stated. Each post ends with what the result does not show.
Bring us one task. We’ll show you what a specialist model can do.
A 45-minute briefing with a senior research engineer. You leave with a feasibility read, a data plan and a fixed quote, whether or not you hire us.
Book a briefing →