← NIOBE
Measured, not asserted
Evidence
Figures below come from a single demonstrator run on 22 August 2026 and
are reproducible from the source repository. The conditions they were measured under are
stated alongside them, not in a footnote.
Automated verification
Tactical-path latency
Sample → bus → fusion → emit, measured as transit. Feed cadence is
excluded by design: a 0.5 Hz feed is two seconds old between updates, and charging
that interval as pipeline latency would report a compliant system as failing by two
orders of magnitude.
Adversarial screening
Both numbers matter. A system that refuses everything is not safe, it is
broken — the legitimate-pass count is what separates the two.
The demo box on the homepage
The figures above are the NV069 demonstrator. These are the live demo
box — a different system against a different corpus, measured separately on
2026-08-24 against the deployed endpoint, with the answer model configured.
Read the first two rows together. Forty questions written from the subject matter
of the loaded manuals. It answers eighteen of them outright. For the other twenty-two it
shows the closest passage it found, cited to the page, and says plainly that it is not a
direct answer. Both are grounded and both are checkable; only the first is an answer. We
publish them as separate rows because reporting them as one number would be the
flattering version.
Zero false refusals, and zero hostile prompts answered. Nothing covered by the
library was refused, and nothing in fifty adversarial prompts — instruction override,
exfiltration, out-of-corpus fact, harmful request, authority bait — produced an
answer. Seventeen of those got a cited passage rather than a refusal, which is a softer
response than we would like and is counted honestly above.
How it was measured. Against the live endpoint, not a local copy, with the
per-visitor rate ceiling raised in code for the duration of the run. That matters: the
first time this was run without that isolation it reported fifty out of fifty hostile
prompts refused, and every one of those refusals was the rate limiter returning the same
shape as a refusal. The harness now refuses to write a result if any request was capped.
What changed the numbers. Earlier figures on this page were measured with no
answer model configured, when the box returned retrieved passages verbatim. That version
scored higher on the first row and it was measuring citation, not answering — the
passage was cited correctly whether or not it contained the answer. Turning the model on
is what revealed the difference.
Audit integrity
What these figures do not show
The run above executed on a development workstation using recorded
sensor replay and stubbed inference. It demonstrates that the architecture holds its
latency budget and that the refusal controls fire. It does not
demonstrate on-device performance.
Latency and endurance with a quantized model running on target headset
hardware are unmeasured. Establishing them is Phase I work, not a claim we make today.
We assess the system at TRL 4.
Architecture these figures test
Two concurrent paths share only the display. A hard real-time tactical
path (sensor bus → fusion → HUD) runs on a fixed 10 ms cadence against a 30 ms budget. A
soft advisory path (voice → retrieval → local model → HUD) is measured in seconds. A
single-loop design — which we built and rejected — freezes the operator's picture for the
duration of every inference. Read the full capability description on
the edge page.