Problem
The model name is not a guarantee
Two models with similar benchmark scores can behave very differently when tools, constraints, structured outputs and real operational sequences are involved.
EMPV / AI LabB
Applied R&D / capability evaluation
Local AI
A lab for understanding what a locally executed model can actually do on a specific workload, using reproducible evidence and explicit operating boundaries.
Repository statePublic Core 0.3.1 RC1 · hitch-local-ai==0.3.1 · schema 1.0.1.
Direction
Test the workload before deployment
The path separates discovery, observed candidate, test, preflight, bounded run, evidence and comparison without turning one successful test into operational authorization.
01 / What exists
A versioned and verifiable public core.
The project exposes a Python package, CLI, schemas, job/test packs and output contracts for describing and verifying local-AI work reproducibly.
02 / What we are looking for
A useful decision measure, not a model leaderboard.
The question is not which model is best in general, but which model + runtime + workload combination satisfies a concrete contract with enough evidence.
03 / Potential
A qualification layer for private and on-prem AI.
If the thesis holds across domains, the value could be an independent layer between local models and enterprise workloads: test first, qualify separately, operate only within demonstrated boundaries.
EMPV / NOTE
The public surface is intentionally separated from private R&D state, host inventories and operational recipes.