Native language lowering
Use messy English to select a question your domain already knows how to evaluate. Bound the task with state, a question and allowed answers.
state + English request + question + allowed answers
→ neural encoder vectors
→ finite option scores
→ typed candidate
→ validation and calibrated admission
→ evidence resolution and playbook evaluation
→ application-controlled action
The model produces finite predictions. Ordinary application code serializes the candidate as JSON. The native path does not ask a decoder to write JSON tokens. Recorded inference checks blocked generation, tokenizer decoding, model-output JSON parsing and network access; see the native-path evidence.
What the supply-chain contract supports#
The application selects an order. The language model selects fulfillment to assess whether it can ship from stock, expedite to assess whether expedited delivery is needed, or unsupported. Commands, historical questions, compound requests and unclear or out-of-domain requests are outside this contract.
Selecting fulfillment does not mean “ship.” The playbook still needs ordered quantity, usable stock, quality-hold status and the approved reserve policy. Selecting expedite also needs transit time and the deadline. The current language contract does not extract arbitrary records, quantities or relations from English.
A common provider boundary#
Each registered native provider accepts the same typed request and returns a candidate containing a choice, a score for every allowed answer, model identity and a request fingerprint. Providers retain their own encoder and vector representation. The common boundary is the finite candidate; it does not require different models to share an embedding space.
The pack selects a trusted, registered adapter. A configuration string cannot import arbitrary code. Unknown adapters, missing native scores, incompatible resources and remote or generative implementations fail explicitly at this boundary. Training and export remain separate from inference.
| When you change | What must be checked again |
|---|---|
| Ontology meanings or allowed assessments | The derived language contract and its hash; compatible training and fresh calibration. |
| Model or provider | Model resources, tokenizer/configuration, artifact identity, contract compatibility and that model's calibration. |
| Policy or evidence bindings | The executable policy, required inputs, source permissions and release review. |
The registered-provider replay preserved the recorded predictions. That verifies the refactor's behavior; it does not make the models more accurate.
Confidence is an admission input#
A high score is not evidence that a business fact is true. Temperature fitting and threshold selection use separate calibration partitions. A qualified threshold is bound to the exact model and contract; it is not reusable across models or domains.
The recorded model is not admitted
The supply-chain model achieved 221/233 (94.85%) candidate accuracy on its frozen synthetic test. It did not meet the preset calibration requirement of 98% observed precision with at least 30 admitted calibration cases. Its threshold remains null, so the pack admits no requests. This is an unresolved confidence problem, not a production reliability claim. Read the full result and errors.
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.
Where knowledge and live data enter#
The ontology gives the request its domain vocabulary. Knowledge artifacts supply policy evidence; configured data sources supply observations after interpretation is admitted. Similarity or a nearby vector can suggest a candidate, but cannot establish fact truth or skip validation. The recorded native-model benchmark makes no live source calls and does not measure live-data freshness.
See Prepare a language contract for the assessment boundary and Studio authoring for reviewed rule creation.