xenoglyph
Detection tells you something is there. It does not tell you what it means, how sure the machine is, or what the reading rests on. xenoglyph is the layer that answers those questions — and refuses to answer when it cannot.
Three missions.
pyroglyph
Wildfire intelligence for the fleet that fights it. Passive: it reads the aircraft and the public fire record, delivers to the TAK screens crews already run, and says what it cannot know. Its live row flies a real fleet pass again from the aircraft's own broadcast.
Open pyroglyph · Watch the live row
raven
The aerial scout for a ground pursuit. Its one clock is the minutes between a suspect evading and a safe capture. It reads behaviour, never a physical description, and tasks the fewest cameras for the shortest window.
Private demonstration, by request
MEDUSA
Counter-UAS for the drone that carries no signal. A fibre-tethered drone emits no RF: nothing to detect, nothing to jam. MEDUSA does not chase the drone. It finds the pilot. Behead the source to stop the threat.
Private briefing, by requestWhy one company can carry all three.
The three missions share one sky. Fire aircraft, police air units and small drones cross the same airspace. Our own receivers have recorded it every day for 101 days, and our own cameras' clips are joined by time and place to what the aircraft overhead said about itself. The engine reads what that motion means, and says when it cannot tell. Anyone else who wants this record would have to stand where we stand, for as long as we have.
Counted from our own record: receivers 18 July to 27 October 2026, camera clips 3 to 26 October 2026. The pyroglyph page recounts them live.
Legible AI.
Most explainable AI explains after the fact. We invert it. A human authors the dimensions before the reading; the engine returns a profile against them, not a label. The analyst decides.
A bolt-on, not a rip-and-replace.
Not a collection system: an exploitation layer over imagery already in a program of record, adding a per-detection record an analyst — and an accreditor — can defend. Output follows published standards, so a conformant consumer needs no bespoke parser.
Air-gapped and on-premises; data and weights need not leave the customer's boundary. FIPS / NIST 800-53 / GovCloud are a deployment target, not an accreditation we hold. Model upload and validation inside the boundary are not built.
One engine. A family of torches.
Each torch is the engine pointed at a domain.
| xenoglyph | The engine | Shipping · private |
| MANTIS | Visual intelligence for analysts | Shipping · private |
| APTERA | Paperback-sized edge inference, no accelerator (MobileCLIP + ONNX) | Shipping · private |
| Lumen | Manuscript analysis | Preprint of record |
| pyroglyph | Wildfire | Public preview |
| PROGLYPH | Adversarial self-review | In development |
| Nymph | Wearable interface | In development |
| field-sight | Field and naturalist | In development |
| aletheia | Airspace and formation | In development |
Two torches are shipping privately; six are not.
MANTIS and APTERA. The engine is not a torch.
What every reading carries, and what it does not.
Ships with every reading
A confidence score
a ranking score, not a probability. No calibration is claimed, because none has been measured.Model provenance
model id, version and weights hash.Supporting evidence
the observation behind the conclusion, so a reviewer can disagree.Saliency — the “why”
the read behind the detection, or no backing map. The dossier computes an abstention rate; over an empty set it reports null, never a flattering zero. No deployment has produced one yet, so this is a mechanism, not a measurement.Symbology and marking
both self-declare aspirational: a curated symbol subset, and a label binding that detects modification rather than establishing non-repudiation. Under the public default key it is forgeable until the operator supplies one.
Live limits
No customer, no revenue, no third-party evaluation.
None is claimed.No production integration.
Conformance is measured by our own suite. No third party has ingested this output.The encoder reads a fixed window.
77 tokens; axes are encoded as ensembles. 0 of 5,209 encoded texts are over it. On three judgement dimensions, a low score is not evidence of absence.Out-of-sample gating is not evidenced.
The only held-out artifact names a missing file; the claim was removed.
The deployment dossier evidences, per deployment, a fail-closed human gate, honest refusal, label binding, responsible-AI governance and sovereignty.
Checkable without an NDA.
USPTO 64/129,348
Provisional, engine method. Priority March 2026.
USPTO 64/231,134
Provisional, multi-domain profiling. Filed April 2026.
Zenodo preprint of record
Visual semantic analysis of an unread manuscript, with its open dataset.
10.5281/zenodo.19560958
10.5281/zenodo.19560769
The retirement record.
An aerial-imagery capability that conflated modern earthworks with ancient ones was withdrawn. Two research findings were withdrawn before publication after our own review refuted them. Registries are available under NDA.
Our humans.
Jacob Lyons
Co-founder & CEOBuilt the meaning engine. About ten years in data science: data science lead at ALLDATA, wildfire data engineer at PG&E.
Christopher Stephenson
Founding Advisor3× founder · Topos Labs (MIT Top 10 Innovator, acquired by Nuix) · Leads AI strategy and Responsible AI governance at Nuix · US patents, NLP