# Offline LLM-foundations evidence lab

Run `python foundation_cases.py` from the repository root. The lab uses deterministic, authored documents, outputs, model records and task labels for NP-0521–NP-0540. It makes no network request and executes no language model.

The cases test budget arithmetic, schema and semantic validation, evidence/instruction boundaries, fixed evaluation sets, compression invariants, freshness decisions, refusals, abstention, cost/latency/quality selection, prompt-injection boundaries, versioning, adaptation choice and labelled-data review. “Candidate outputs,” quality results and latency/cost values are fixture data for teaching the evaluation method, not model benchmarks.

Tokenizer counts differ across model/tokenizer versions. The context case therefore labels its whitespace units as teaching units and requires production code to use the selected model’s supported tokenizer and documented limits.

The 20-check verifier passed on 20 September 2026 and saved every fixture in `foundation-verification.json`. Human review, provider compatibility and live-model evaluation remain pending.