NGAD sensor-fusion briefing scaffold

One frozen geometric operator, applied vehicle-wide, catches co-signals per-subsystem models cannot.

VW-CSIM is a deterministic, non-actuating monitor that scores every telemetry substrate on the same platform through a single measurement kernel. The invariant it computes cannot be inherited from any one substrate, so a coincident excursion across physically unrelated channels is exactly what the operator was built to name.

Meeting
Lockheed Martin · 22 Sept 2026
Framing
NGAD-forward, Starship as proof point
Kernel status
v1.0.0 · Frozen · Vendor-neutral
Vocabulary
Deterministic geometric measurement

The invariant does not live inside any one sensor.

Every substrate on a complex vehicle (a Raptor cluster, a wing-box strain field, an infrasound array, an AESA/EO-IR/DAS/RWR/MADL sensor-fusion stack) has its own manufacturer, tuning, and normal-behaviour model. Faults inside any one of those views are typically caught by that view's own supervisor. Coincident excursions across views, with the right physical lags, are not.

VW-CSIM computes an affine-invariant SPD-cone distance over a rolling representation of each substrate, then combines them through a Cross-Substrate Coincidence Statistic evaluated at pre-declared physical lag lists. The measurement operator is fixed, sealed and version-pinned. It is used as a floor over per-substrate supervisors, never as a controller and never as a predictor.

Per-subsystem tuned models cannot supply a vehicle-wide invariant.

Sensor-fusion, engine health-management and structural monitoring today are strong within their own boundaries and increasingly rely on statistical, non-deterministic components. NASA's own complex-autonomous-systems assurance work is explicit that such components cannot be certified at algorithm level. A supervisor sitting above them must be simple, sealed and universally scored.

Sensor fusion · AESA / EO-IR / DAS / RWR / MADL

Tracker-level correlation, per-sensor confidence, contested-EM discipline. Powerful within the fusion boundary.

Does not score whether the platform is behaving consistently with structure, propulsion or acoustic signature.

Engine health monitoring

Per-engine redlines, chamber-pressure envelopes, per-cluster statistical models tuned to that fleet.

No shared coordinate with structural, acoustic or mission-sensor state. Precursors that live across those channels are invisible.

Structural health monitoring

Strain-gauge envelopes, mode-shape trending, damage-index scoring.

Not built to co-witness a propulsion ensemble or fuse with a mission-sensor bus in the same measurement.

Watch four substrates through one operator.

Four substrates stream in parallel through the same measurement kernel. The single bold trace is the Cross-Substrate Coincidence Statistic, the invariant score. Switch scenarios to see what the same operator does in nominal, precursor, degraded-link and spoofed-track conditions. Toggle the pre-declared floor on and off to compare "detector-armed" versus "comparison-only" behaviour.

Floor engaged Synthetic scaffold
Rolling geometric residual + CSCS overlay T+0.00 s / 24 s window
CSCS (invariant score) Pre-declared floor Substrate residuals ~ dimensionless RMS
CSCS · current 0.00 Dimensionless, single-detector
Floor state ARMED Pre-declared, sealed with kernel
Lead vs per-engine redline Illustrative reference sample only
Verdict Attribution held to floor plane

The four scenarios are canned. They exist to make the operator's behaviour legible in a briefing, not to substitute for real telemetry. Under a supported programme, VW-CSIM would ingest whatever native cadence each substrate produces, run the same kernel, and hold the same floor.

Per-substrate residuals, one kernel.

The same operator is applied to each substrate independently. What differs is the raw signal: engine chamber-pressure ensembles, strain-gauge fields, ground acoustic arrays, sensor-fusion track states. The residual family is directly comparable across substrates because it is computed identically.

Raptor cluster 33-engine
Residual RMS 0.000
Structural strain Field
Residual RMS 0.000
Infrasound array Ground/air
Residual RMS 0.000
NGAD sensor fusion Fusion
Residual RMS 0.000

Faint background traces show the raw substrate signal. Coloured traces show the rolling residual, the quantity the invariant operator consumes. Residuals are pre-attribution: the CSCS asks whether they climb together at the correct physical lag, not whether any single one is above its local envelope.

Three claims, each with a pre-declared null.

VW-CSIM is a research instrument. Its value depends on being wrong in specific, measurable ways. Every claim below is stated so it can be preserved as a negative result if the test fails, and each protocol is fixed before data is touched.

Claim I

Substrate invariance

A single frozen operator applied to physically unrelated substrates produces residual families whose baseline geometry is statistically indistinguishable. Distinguishability under a fixed KS-test threshold falsifies the invariance argument.

Null preserved
If distinguishable, invariance argument is retracted; per-substrate operators remain admissible in their own domains. Result reported as measured, not tuned away.
Claim II

Coincidence at pre-declared lag

The CSCS crosses a fixed floor before any per-subsystem redline in a pre-registered event set, and does not cross the floor in a matched null set. Lag list L is declared with the kernel and not adjusted after the fact.

Null preserved
A single false alarm in a size-matched null set forfeits the claim for that lag pair; the pair is retired to the failure register with its recorded margin.
Claim III

z-Stack triage separation

The synchronous CSCS acts as a detector; asynchronous z-Stack subspaces (pair, counterfactual, Janusian) run as triage and never re-enter the detector vote. Their scores are attribution, not evidence.

Null preserved
Any measured leakage of a z-Stack score into a CSCS decision voids the run; the leakage path is documented before the register is closed.

Transfer is the argument.

Because the kernel is frozen and vendor-neutral, the mere fact that the same operator produces meaningful residuals on a Raptor cluster and on an NGAD-generic sensor-fusion bus is the invariance claim's operational witness. Below are the two proof surfaces we run in parallel.

NGAD-forward proof surface

Sensor-fusion supervisor above the fusion boundary

The NGAD substrate is treated as a heterogeneous ensemble: AESA range-Doppler cells, EO-IR track states, DAS aperture confidence, RWR emitter classification, MADL link geometry. VW-CSIM does not replace the fusion tracker; it sits above it and scores whether the fusion output is consistent with structural, propulsion and environmental substrates through the same measurement.

  • Contested-EM degradation Operator isolates rank collapse to NGAD substrate. CSCS does not fire because no cross-substrate co-signal exists.
  • Spoofed-track injection Operator rejects coherent NGAD anomaly for lack of matched structural or acoustic response at declared lag.
  • Genuine vehicle-side excursion Cross-substrate lift confirms. NGAD-attributed floor cross would only ever be a warning against the operator vote.
Starship as concrete proof point

Same operator on an unrelated vehicle

Starship is the public-record witness. The operator applied to the Raptor 33-engine ensemble, vehicle-wide structural strain, and IMS-class infrasound produces a CSCS trace whose coincidence structure, at pre-declared propellant-to-chamber-to-structure lags, climbs above the pre-declared floor 13 to 49 seconds before per-engine redlines would fire on landing-burn no-relight scenarios.

  • Public flight witnesses IFT-1, IFT-2, IFT-3, Flight 12 (5-of-33 no-relight), Flight 13 (hot-stage, hard ship splashdown).
  • Independent infrasound record IMS network detected IFT-2 at ~4000 km with resolvable stage separation and booster explosion as distinct signal groups.
  • Documented seismic proxy Texas state seismological network recorded IFT-5 launch at Hebbronville ~192 km; Linares ~257 km.