Native provider regression
The shared provider registry preserved all 466 recorded choices and probability vectors across two registered native implementations. Both retained their original closed admission gates.
What changed#
The pack and evaluation harness now resolve a declared provider through one registry. Request and candidate validation, artifact identity and admission use shared contracts. Ontology, playbook and evidence resolution remain in the existing runtime. The exercised path is described in Native language lowering.
| Recorded adapter | Cases | Choices preserved | Maximum probability change | Admitted |
|---|---|---|---|---|
| domain-trained | 233 | 233 | 0.0 | 0 |
| second-native | 233 | 233 | 0.0 | 0 |
Each implementation replayed the same 233 previously exposed synthetic requests through disposable pack copies. “Choices preserved” compares each implementation to its own prior recording, not one model to another. The record shows no live-source calls or business action execution. The canonical pack remained unbound.
Adapter timing#
| Recorded adapter | Pairs | Median before (ms) | Median after (ms) | Median paired change (ms) |
|---|---|---|---|---|
| domain-trained | 12 | 369.99 | 370.70 | +1.17 |
| second-native | 12 | 599.82 | 599.95 | +0.73 |
These are small paired warm samples with alternating execution order on the same workstation. They suggest little adapter overhead in the exercised configuration, but do not establish a statistically reliable speed difference or throughput. The second adapter's new boundary also applies frozen temperature and shared validation, so its before/after work is not identical. A median of paired changes is not the difference between the two medians.
Recorded 2026-10-03 on M5 Max, macOS arm64, Python 3.12.12, 4 CPU threads, batch size 1. Synthetic labels. Provider inference, tokenization and candidate validation; excludes first call, live evidence retrieval, reasoning and business action execution. Evidence and reproduction limits.
The accuracy and admission limits of the domain-trained model remain those in the original frozen evaluation. No new training, calibration or unseen test was part of this refactor.
Evidence scope#
The public regression record contains per-provider replay totals, probability-difference summaries, admission counts and raw paired timings. The manifest pins its digest. Adapter names on this page are descriptive aliases, not package identifiers.
The public export supports arithmetic checks of these recorded summaries and timings. It does not distribute the before/after model resources or complete second-provider prediction trace, and cannot independently rerun either adapter. Both full local replays used generation, decoding, output-parsing and network tripwires. This is interchangeability evidence with explicit limits, not a comparative product ranking.