Sovereign domain superiority

Anomaly detection + track fusion — with an inspectable receipt boundary.

SZL SDA demonstrates a governed Common Operating Picture: it scores anomalies, fuses tracks across sensors, attaches an honest confidence interval and a Λ-gated advisory verdict, and emits a provenance receipt shape. This surface remains unsigned until a real downstream signer supplies evidence.

SZL's clean-room sovereign Domain-Superiority engine — khipu-sda-core — inspired by True Anomaly's Mosaic capability (not affiliated).

Λ = Conjecture 1 (advisory) Receipt schema · UNSIGNED here Sovereign — own-metal only COP · DEMO until live read Orbital SDA = ROADMAP Clean-room · not affiliated

Investor brief · evidence before posture

See the value, then inspect the boundary.

SDA combines anomaly scoring, track fusion, advisory gating, and receipt structure in one reviewable surface. It demonstrates the workflow; it does not establish operational accuracy, signed provenance, orbital capability, or deployment.

Inspect source binding Inspect build identity

LIVE / MEASUREDOnly after a current endpoint read succeeds; the UI records the observation time.

DERIVED / MODELEDScores, confidence intervals, and advisory verdicts are computed outputs, not measured threat truth.

SAMPLE / DEMOIllustrative tracks and synthetic baselines stay explicitly demo data.

UNAVAILABLE / LIMITSMissing feeds remain unavailable. Receipt examples are unsigned and orbital SDA is roadmap.

Developer quickstart python -m http.server 8000 Open /index.html; no build step or credentials are required for the static UI.

The capability — plainly

Spot what doesn't belong, fuse the picture, and prove the call.

SDA's modeled pipeline accepts sensor streams, flags abnormal tracks, merges duplicate sightings, and attaches a confidence band plus receipt structure. This Space demonstrates the air / maritime / counter-UAS workflow; the orbital extension is roadmap, not a live claim.

Detect (anomaly score) Fuse (tracks → COP) Λ verdict (advisory) Receipt + signing state

01 · Detect

Multivariate + graph anomaly

An ensemble (isolation forest + autoencoder + robust z-score) scores each track on stationary behavioural features — speed, RCS, acceleration, heading-rate. A separate graph-relational channel (GDN-style) flags tracks that deviate from their velocity-space neighbours — explainable, not a black box.

MODELED · FEED-DEPENDENT

02 · Fuse

Multi-sensor track fusion → COP

Nearest-neighbour association plus a compact Kalman filter merge noisy, duplicate sensor reports into one Common Operating Picture: each real object becomes a single fused track with a clean state estimate, even when two sensors disagree.

MODELED · FEED-DEPENDENT

03 · Orbital seed

SGP4 conjunction (roadmap)

A python-sgp4 TLE-propagation + conjunction-flag stub seeds future orbital SDA. It is shipped clearly labelled as a space roadmap; SZL does not claim an operational orbital feed.

SPACE ROADMAP

04 · Gate

Λ-gated advisory confidence

Every verdict carries a split-conformal interval plus a PAC-Bayes bound — both labelled ESTIMATE. The Λ axis (Conjecture 1) returns allow / advisory / deny as an advisory only — never "proven trust", never a false green.

MODELED · ADVISORY

05 · Prove

Provenance receipt schema

Each detection emits an in-toto Statement v1 receipt — method, input hash, anomaly score, fused track id, Λ verdict, confidence band. It is honestly UNSIGNED until a real DSSE signs it downstream; no signature is ever fabricated.

SCHEMA · UNSIGNED HERE

06 · Deploy

Sovereign + air-gap

The static UI uses 0 runtime CDN. Separate repository artifacts describe an szl-sda UDS / Zarf bundle; this Space does not verify or claim that external deployment.

REPORTED · NOT VERIFIED HERE

Attribution (clean-room — cite, never plagiarize): SZL SDA is inspired by True Anomaly's Mosaic capability (sensor-fusion Common Operating Picture, OODA acceleration, ML threat-warning) — described publicly by Eric Hilmer / True Anomaly. SZL is not affiliated with True Anomaly and no proprietary code was seen or copied; the capability was re-built from public descriptions only. Methods were independently re-implemented from a verified-permissive lineage: PyOD (BSD-2), PyGOD (BSD-2), Merlion (BSD-3), TODS (Apache-2.0), tsod (MIT), GDN (MIT), GraGOD (MIT), python-sgp4 (MIT). alibi-detect is excluded (Seldon BSL-1.1) and CI-enforced absent.

The picture, drawn honestly

A Common Operating Picture with explicit source state.

Fused tracks as blips, an anomaly overlay coloured by score, a per-track score panel, and how multiple noisy sensor reports converge into one confirmed track. This is a labelled demonstration driven by SDA's real scoring logic on illustrative tracks — it reads the live killinchu /mosaic endpoints when reachable, and says so when it is not.

Common Operating Picture — anomaly overlay

DEMO DATA

benign (low score) elevated anomaly (Λ advisory) fused COP track

Anomaly score · Λ verdict

DEMO

Score ∈ [0,1] from the multivariate + graph ensemble. The Λ verdict (allow / advisory / deny) is advisory only — Λ is Conjecture 1, never a proof of trust.

Demo / illustrative data. The blips, scores, and fusion above are an honest demonstration of the SDA pipeline on illustrative tracks — they are not a live operational feed and contain no fabricated measured numbers. When the live killinchu /api/killinchu/v1/mosaic/* endpoints are reachable, this panel reads them and re-labels itself LIVE; until then it stays honestly labelled demo.

Benign tracks (cyan) drift normally; an anomaly (violet, ringed) trips the ensemble and graph channels; two noisy sensor reports for one object are associated and Kalman-fused into a single COP track (teal).

The moat

Every detection can carry a checkable receipt.

A threat call you cannot inspect is an opinion with extra steps. SDA defines a provenance receipt shape for method, input hash, anomaly score, fused track id, Λ verdict, conformal interval, walltime, and verification state. Below is an unsigned example shape, not a live verdict or signature claim.

Verified verdict receipt — fields

EXAMPLE SHAPE — NOT A LIVE VERDICT

methodsda.anomaly.ensemble+graph.v1

enginekhipu-sda-core

inputs_hashsha256:<track digest>

anomaly_score<value ∈ [0,1]>

fused_track_idcop:<fused track id>

lambda_verdictallow | advisory | deny ▸ advisory

confidence{lo, hi} conformal · ESTIMATE

sgp4null (air/maritime) · roadmap for orbital

walltime_ms<measured>

attestationin-toto Statement v1 / DSSE

_signing.statusUNSIGNED ▸ until real DSSE downstream

verifiedfalse ▸ structural-only, advisory

anomaly_score and the confidence band are ESTIMATES, never a proof of threat. walltime_ms is measured. The Λ verdict is advisory (Λ = Conjecture 1) — never "proven trust", never a false green. The receipt ships honestly UNSIGNED / verified:false until a real DSSE signature is applied downstream by the szl_lake / khipu path (Khipu BFT settlement is Conjecture 2); no signature is ever fabricated. You can check a real receipt against the live fabric below.

Honest baseline metrics — synthetic data

SYNTHETIC-DATA BASELINE

Real numbers from `szl_mosaic_validate.py` on synthetic tracks with injected anomalies (maneuver, RCS spike, weave). Thresholds set from the calibration distribution, not tuned to the test anomalies.
Channel	Precision	Recall	F1
Point ensemble iForest + AE + robust-z	0.44	0.74	0.55
Graph-relational GDN-style velocity dev.	1.00	0.35	0.52
Fused consensus 0.5·point + 0.5·graph	0.40	0.70	0.51

These are honest synthetic-data baselines, not operational accuracy. The point ensemble carries recall; the graph channel is high-precision / low-recall and correctly misses the RCS spike (no kinematic signature) — the right behaviour, not a bug. Numbers improve with real training data; they will be re-measured on the live GPU fabric and re-issued with a real signed DSSE receipt — never inflated.

Track fusion: 6 fused COP tracks recovered from 6 true objects across 2 noisy sensors. Source: khipu-sda-core validation harness (synthetic).

Where it fits

A governed domain-awareness layer for the sovereign estate.

SDA is a capability, not a silo. It is already wired into the estate.

Powers killinchu

Feeds the killinchu Track Board, Threat-DB and the new Mosaic / Domain-Superiority COP view via /api/killinchu/v1/mosaic/* (PR #118) — anomaly overlay on every fused track. killinchu effectors remain simulated.

Governed by a11oy

Every verdict is policy-evaluated in a11oy's Governed Anomalies tab via /api/a11oy/v1/mosaic/governed (PR #356): 13-axis Λ advisory, DSSE-shaped receipts, Khipu BFT 3-of-4 quorum (Conjecture 2), and a human-approval gate. receipts.in ≡ receipts.out.

Air-gap deployable

Ships as the szl-sda UDS / Zarf bundle (cosign-signed, OSCAL, Section-889 denylist, λ-gate policy) for classified or air-gapped install. Image digest stays blank until founder-signed — never fabricated. Canonical version uds-v0.4.0.

Sovereign GPU fabric this runs on

SNAPSHOT

Live read of https://a-11-oy.com/api/a11oy/v1/compute-pool. sovereign GPU = own-metal only: nodes you actually own are flagged sovereign; hosted APIs are labelled hosted fallback — not compute you own and never counted as sovereign. If the live read fails, this panel falls back to a captured SNAPSHOT and says so — it never invents a live status. The MEASURED SDA validation will be re-run here on the real GPU fabric and re-issued with a signed DSSE receipt.