DEIC: Discrete Executive Inference Core

DEIC explores a practical question: when redundant information sources disagree, which state should a system believe, which source is suspect, and what should it check next before committing?

Direct Use: Sensor Trust Diagnostic

The public browser app is a zero-key worksheet for redundant sensor arrays. Enter readings, source-reliability priors, operating thresholds, expected noise, and confirmatory-probe costs. It returns:

• a reliability-weighted robust consensus and uncertainty interval
• COMMIT, ABSTAIN, or ESCALATE with an explicit reason
• the most suspect sensor from deviation and leave-one-out consistency
• the next confirmatory probe ranked by expected diagnostic value per cost
• a downloadable decision trace for review

The implementation is deterministic JavaScript under docs/; all inputs remain in the browser. Run its engine checks with:

node docs/engine.test.mjs

Method

The worksheet combines:

1. reliability-weighted median state estimation
2. median-absolute-deviation robustness
3. leave-one-out consensus shifts
4. threshold-crossing uncertainty checks
5. value-per-cost probe selection

It is a diagnostic planning demonstration, not certified control logic. Real deployments require calibrated noise models, temporal dynamics, authenticated telemetry, domain-specific validation, and human safety ownership.

Research Code

The public repository has multiple historical lines of development. The mature bounded-inference and metacognition benchmark currently lives on the metacog-benchmark-expansion branch, where deic_core/, cross-domain benchmark adapters, regression tests, and result packages are maintained.

This main branch intentionally keeps the immediately usable browser diagnostic small and dependency-free while the research history is consolidated.

License

MIT