An eval loop withno invoice attached
Evaluation sweeps, fine-tunes and transcription for unpublished work — reproducible because nothing changes under you.
Novel work leaks by inference too. No cloud trail, nothing to subpoena, nothing to train on.
The day, step by step
Each step names the tool and the model doing the work — all from the verified catalog.
- Designllama.cpp · pinned GGUF
Pin everything
Weights, quant and prompts pinned in the harness config — the run that works in March replicates in June, because no provider can deprecate your model.
- NightlyCustom harness · gpt-oss-120b / Qwen3.5
Sweep
The evaluation matrix runs overnight at zero marginal cost — the experiments that cloud pricing quietly discourages become routine.
- WeeklyLocal fine-tune tooling
Fine-tune pass
LoRA adaptation on the lab's material, evaluated against the pinned baseline the same night.
- OngoingWhisper large-v3
Interviews transcribed
Field recordings and lectures transcribed offline — participant consent covers your machine, not a vendor's.
The stack, pre-installed
- gpt-oss-120b
- Qwen3.5 122B
- llama.cpp
- Whisper large-v3
- LoRA tooling
The honest caveat
One box fine-tunes adapter-scale, not foundation-scale — we say so before you buy. For full-parameter training you still want rented big iron; for evaluation and adaptation, this is the economical shape.
This arrives working — models loaded, agents wired, your team walked through it.
Configure yours