AI Engineering Decision Layer for iTools / Codex Workflow

An experimental decision layer for an autonomous AI engineering workflow. I am testing basal as a small System One model for fast classification of problems, failures and findings before harder cases reach larger reasoning models. I am building iTools, a CNC tool-management system, using long autonomous Codex workflows with multiple AI models, tests, reviews and automatic validation.

In such a workflow, large reasoning models often also make simple semantic decisions, such as classifying a problem as a product defect, test-contract defect, environment problem or evidence problem.

I am experimenting with basal as a fast decision layer below the larger models.

The intended flow is roughly:
deterministic software / tooling
→ basal — fast classification and triage
→ a larger model only when the case requires it.

The first use case under test is classification of real failures and findings produced during Codex work. The model should return a category and confidence, leaving low-confidence cases to larger models.

At this stage, it is an experiment and shadow evaluation: basal does not yet control the workflow or make production decisions.

What I measure
- Accuracy on real work cases.
- Confidence, including confidently wrong answers.
- Coverage at thresholds of 0.90 / 0.95 / 0.98 / 0.99.
- Latency.
- Local resource use.
- Above all, how many actual calls to larger models could safely be avoided.