CEP-Rank is not a black box of kind

This page exists because the public pitch cannot ask a buyer to trust a nameless ranker. We disclose the kind of model. We do not publish weights, join bonuses, or unpublished thresholds. We do not claim a trained classifier, and we do not put an LLM in the rank path.

What this is not: not laboratory consciousness, not a quantum computer, not AUTO_SAFE live, not a measured precision score from a paid pilot.

Kind

Classical / rules + scored correlation on fused QFC signals. Deterministic enough to journal. Not a neural ranker. Not an LLM vote.

Inputs

Fused plane records: edge, admission PEPs, proxy verdict, IDS, and labeled intel. Same schema, one timestamp. Silos do not rank. This function lives inside Ikaros — it is not a second engine.

Output

An ordered hypothesis list and a proposed act. The AEGIS kernel — not this function — applies HUMAN_ONLY / AUTO_SAFE / DENY and writes the journal. AUTO_SAFE is off.

LLM

Narration only. Translates an already-ranked journal into analyst prose. Explain. Do not rank. Do not gate.

Operational shape (illustrative, not a published coefficient set). score(h) combines weighted correlation of fused signals, a multi-plane join term, and a false-positive-debt penalty. The decision is argmax score subject to the policy gate. That is a rules + scored join story, not “an LLM decided.”

What we will not let this page imply

No pilot MTTD, precision, or hunt-cadence number lives here. Demo scores stay labeled illustrative. Verified numbers replace them only after the first paid Field Mandate pilot, and both labels will stay distinct.

Evidence grounding (CVE / NVD / ATT&CK citations on each ranked option) is the next product increment — retrieval for the HUMAN_ONLY reviewer, not a transfer of authority to a model. A customer-auditable decision journal shape is on journal.html. Graduation numbers are on Govern · AUTO_SAFE.