finboard · assistant evals

Footprint Ladder

Fix a prompt era and walk models: accuracy against resident memory — the frontier a model choice is made on. A footprint renders validated only when the ledger proves it; the pill is computed, never stored.

Anatomy of a Gate Run — the mechanism Eval Ledger — fix a model, walk time Run Inspector — drill into one run Run Composer — compose the next run

Footprint Ladder

Accuracy vs memory — the frontier

Down the quantization ladder

Context window vs memory — the KV-cache tax

The ladder — one row per model × quantization