# Validation scope, 2026-10-01

1. Real fixed-checkpoint CPU generation: first full development execution accidentally omitted --smoke; preserved initial_development/stdout.log and all outputs (exit0), then explicit2-example smoke (exit0), then unchanged64-example formal (exit0). PROTOCOL predates every run. This deviation defeats strict smoke-before-any-full-execution chronology, not the development/confirmation separation; no configuration chosen from scores. All runs stay exploratory development.
2. Formal records contain128 unique sample×contract pairs. Initial/formal output token IDs, decoded text, scoring, prompt hashes and input lengths equal exactly; elapsed times excluded from comparison. The repetitions are not independent seeds.
3. audit.py independently reparses outputs, checks source labels/sentence hashes/prompt hashes/token decoding/denominators/split IDs and8 boundary fixtures; all passed. No human gold or human semantic review performed.
4. Model/tokenizer/data and3 official source files plus license are pinned and hash-locked. Three installed source files byte-identical. Raw data/model not redistributed. run.py records first-token generation=forward argmax and real tensor shapes.
5. CPU timing starts after heavy imports; RSS is process lifetime maximum. Inference phase includes shape probe, scoring and output serialization. Download and process start excluded. No GPU measurements.
6. Code changes after formal inference are additions only: audit.py, docs, dependency locks and validation records. run.py/prepare.py/scoring.py/config/PROTOCOL/resources.lock hashes in results/identity.json still match packaged code. Archive manifest covers every final file separately, outside code/.
7. Missing pip module prevented pip freeze; recovered by enumerating importlib.metadata. No packages added or upgraded. prepare download initially blocked by sandbox proxy; public-download execution succeeded, exit0. Failures are disclosed rather than invented away.
8. Acceptance additionally requires fresh archive extraction syntax/help/audit and a real-model minimal replay, plus public fixed-commit anonymous file/hash checks. These records live beside article.md in package_validation.json and publication_verification.json, avoiding self-referential archive hashes. Completion claims must follow those records.