Quality, correctness and provenance

Four independent questions

Question	Suitable evidence
Did we preserve the chosen weights?	Exact artifact identity, byte hashes and no transformation step
Does this engine compute compatibly?	Numerical comparisons and controlled output/token tests
Is the model useful for this task?	A relevant evaluation set with scoring and uncertainty
Is the system faster?	Repeated controlled latency/throughput/resource measurements

Passing one row does not answer the others. A loader can preserve every byte and still use an incorrect kernel. A correct engine can be slow. A fast short answer can be wrong.

What the smoke test proves

The installed RegesARC 0.0.1 core completed a simple arithmetic prompt, cancellation, recovery, unload and persistence using one exact 35B artifact. That establishes end-to-end functionality on the recorded system. It does not measure broad reasoning quality, coding ability, long-context retrieval or equality to the artifact’s original unquantized source.

Build a qualification set

Choose tasks that represent actual use: short factual answers with ground truth, structured outputs, code tests, long-context retrieval and multilingual cases where relevant. Separate template/tokenizer failures from model limitations. Preserve the test set version and avoid optimizing repeatedly against only one easy prompt.

For numerical work, define acceptable tolerances and compare the same tensor encodings and workload. For answer quality, score outputs consistently and record variance. A temperature of zero is useful for control but does not eliminate every source of nondeterminism across hardware and kernels.

Keep the provenance chain

Record where the artifact came from, its license, immutable revision, complete filenames, hashes and any transformation performed before RegesARC sees it. Name the actual serving artifact. An OpenAI- or Anthropic-shaped protocol does not turn local weights into GPT or Claude and does not establish endorsement by those providers.

The quality policy is deliberately conservative: no silent expert pruning, no automatic lower-quality substitution and no universal “same quality” badge. A future optimization may change placement without changing weights, but it still needs correctness and workload evaluation before its speed is meaningful.