AxoDen Defence
Refusal-capable autonomy
AxoDen Defence
Most defence-AI stacks degrade. AxoDen Defence refuses, then emits the refusal as signed, replayable evidence. The product line covers admissibility, action permission, jam-tolerant flight and mission control, and audit-grade replay.
Pilot-stage runtime. The Defence runtime is currently pilot-stage and intentionally DSP-light. Engagements are scoped as bounded pilots, not production deployments. Procurement or safety-critical use requires source review, integration testing and claim verification.
Four parallel commitments. Every claim, threat model and benchmark on this page reduces to one of them.
Only profile-bound, physically valid, source-attested data reaches the AI.
AI and planner outputs do not execute unless authority, policy and evidence conditions remain valid at the point of execution.
Survival and mission split when RF, GNSS or AI evidence is refused.
Every admission, refusal, degradation and action is recorded once and bit-for-bit replayable from genesis.
Five differentiators for engineering due diligence. Each should be tested against the selected pilot use case and evidence corpus.
Refusals are signed evidence events, recorded once and replayable as first-class outputs.
Survival continuity and mission continuity are handled as distinct control questions when RF, GNSS or AI evidence is refused.
Admissions, refusals, degradations and permitted actions are recorded so the decision path can be replayed from the ledger.
Evidence gates run before downstream reasoning, so inadmissible inputs do not silently enter the graph.
Claims are separated by status, including Established results, Claimed results and proof-obligation-bound work.
The current internal register tracks eight defence-relevant threats. Five are marked closed in the current status record; three remain open with named proof obligations or pilot validation paths. These statuses should be verified against the current engineering record before procurement or safety-critical use.
| # | Threat | Status |
|---|---|---|
| T1 | Hallucinated perception under jamming | Closed |
| T2 | Buffer / overload pressure | Closed |
| T3 | Replay / injection / stale intent | Closed |
| T4 | Single-vendor / monoculture compromise | Closed at L4 |
| T5 | Zombie Intent | Closed; PO2 validated at L1 |
| T6 | Forged negative evidence | Open: PO6, bounded sprint |
| T7 | Predicate-burn race | Open: PO3, bounded sprint |
| T8 | Control-kernel root leak | Open: PO1, pilot programme |
AxoDen Defence is most relevant where AI confidence can be detached from signal reality and where refusal is a safer outcome than action.
A defence pilot is bounded, scoped to one contested-signal use case and one decision point where refusal is acceptable or required.
A defined contested-signal use case, known platform or mission boundary, available RF/GNSS/sensor evidence, and a decision point where refusal is acceptable or required.
Admission predicates, authority-binding model, refusal cascade, replay ledger and failure-mode test harness.
Threat model, evidence boundary design, pilot findings, open proof obligations and a recommendation on whether the use case should proceed, narrow or stop.
