Commander Saverio

The problem

An AI system can produce a fluent, confident answer that is wrong. Fluency is not evidence. Where people depend on the answer, something has to check it that does not share the model’s weaknesses.

Two layers

Commander Saverio: the floor

Air-gapped and deterministic. It analyzes the output text itself and produces a verifiable integrity verdict. No cloud, no model server, no GPU.

The Witness: the connected layer

A behavioural analysis layer that characterizes eight failure modes in AI output: confabulation, sycophancy, evasion, context drift, source fabrication, temporal contradiction, authority inflation and hedging-without-knowledge.

The architecture is built so that the connected layer can fail or be unavailable, and the deterministic floor still holds.

What it does

• Hallucination detection. Distinguishes fluent, confident wrong answers from correct ones.
• Adversarial robustness. Holds across paraphrase, indirection, sentence reordering and formal-spec rephrasing.
• Deterministic output. Same input, same verdict, byte-identical, every run.
• Fast. Median 9.2 ms on a commodity CPU.
• Signed receipt per evaluation. Timestamp, integrity score, checksum and sensor values, independently re-verifiable.
• Categorical locks. Five hard override categories (safety, coherence, thermodynamics violation, logical paradox, agentic overreach) force a RED verdict regardless of score.
• Model-agnostic. It works on the output text, not the model internals. Replace the generator; the verification layer is unchanged.

Benchmark results

Independent validation

Commander Saverio is open to third-party testing. Because it is deterministic and runs offline, an evaluator can run it on their own tasks, on their own hardware, and get the same verdicts every time.

Patent pending (U.S. non-provisional application; priority June 2025). A capability brief is available without an NDA. The extended technical briefing requires one.