Nervous Machine encodes the operational physics — NRLMSISE-00 for LEO drag, SWPC REFM for GEO electrons, NAIRAS for Mars dose, naive ballistic for cislunar fields — as the starting prior for a per-edge causal substrate, then calibrates certainty against real archive telemetry from GRACE-FO, GOES-19, MSL/RAD, and ARTEMIS-P1/P2. The substrate runs byte-identical across all four regimes, surfaces regime-dependent physics that the operational priors didn't encode (magnetotail decoupling at the Moon; the voxel-dependent Martian Forbush decrease), and produces a per-regime measurement-gap diagnostic showing where the current science can't yet support autonomous decisions.
Each regime's benchmark runs the substrate as a single prequential pass: the operational model is the starting prior, predictions are made before the observation they target, and edge certainty evolves as actuals arrive. There is no train/test split. The "external comparator" column names the operational prior NM rides on; the precision and lift columns measure how much information the substrate extracts beyond what the prior already provides. Where no time-aligned operational comparator exists publicly (Mars surface dose), the substrate reports lift over its own prior-W internal baseline and the missing-comparator status is documented explicitly.
| Regime | Ground truth | Window | Observations | Edges | Operational prior | Substrate evaluation |
|---|---|---|---|---|---|---|
| LEO | GRACE-FO density | 7.5 yr | ~510k | 356 | NRLMSISE-00 | 92.4% anomaly-flag precision against MSIS errors; 7.9× over base rate on the strongest storms |
| GEO | GOES-19 SEISS + MAG | 7 d | 17,809 | 360 | SWPC REFM Day-1 | Substrate prediction tracks GOES-19 at log-MAE 0.043 on the REFM-overlap window (REFM Day-1: 0.169) |
| Mars | MSL/RAD dose-rate | 1 yr | 30,067 | 24 | (no time-aligned public model) | Best-available prior from sparse data; substrate autonomously learned Forbush coupling W = −0.071 at Z = 1.0 |
| Cislunar | ARTEMIS-P1/P2 FGM | 1 mo (May 2024 G5) | 2,976 | 72 | Naive ballistic L1 → lunar (voxel-dependent) | Voxel-dependent W flip: outer +0.55 at Z = 1.0, magnetotail −0.08 at Z = 0.84 |
LEO and GEO have operational baselines the substrate can be compared against directly. Mars and cislunar do not have time-aligned public operational models for the full window — the substrate's role in those regimes is to build the best-available prior from sparse archive data and provide the learning structure for future in-situ telemetry. Detailed per-edge tables and reproduction instructions are in the per-regime Zenodo bundles linked in the footer.
The four regimes split into two structurally different cases for the substrate. In LEO and GEO, decades of operational modeling have produced trusted baselines. In Mars and cislunar, there is no legacy operational model the substrate could possibly beat — the problem in those regimes is data sparsity, and the work in front of the community is building the structure that lets future missions learn the local environment as their telemetry comes online.
NRLMSISE-00 and SWPC REFM are the priors. The substrate evolves a per-edge certainty around them and surfaces anomaly flags where the operational model will drift. Operators continue to use their existing forecasts — the substrate adds a per-forecast trust signal and a voxel-aware structure that the point-prediction outputs of the operational models don't carry.
Where the comparison is meaningful (operational baseline present): LEO 92.4% anomaly-flag precision against MSIS errors with 7.9× lift; GEO log-MAE 0.043 on the REFM-overlap window vs REFM Day-1 log-MAE 0.169.
There is no operational model with a time-aligned public archive at Gale Crater or at lunar distance for the full window. The substrate's role here is different: build the best-available prior from sparse archive data, and provide the per-edge learning structure that a future mission can pick up and immediately learn from when new telemetry arrives.
The per-edge primitives are small enough to run on a radiation-hardened flight CPU. About 50 kB of state including the learning loops, microseconds per update at typical telemetry cadence, no GPU, no constant cloud connection, no massive data movement — the spacecraft can learn the local environment from its own telemetry, against the priors the substrate already converged on the ground.
Across both modes, the substrate produces the same four outputs: per-edge calibrated certainty (W and Z), voxel-aware structure, surfacing of unmodeled physics that the priors didn't encode, and a measurement-gap diagnostic showing where current data is too thin to support autonomous decisions. The two modes differ in what they compare against, not in what the substrate does.
The defining metric for an autonomy-grade physics substrate is not curve-fitting against known physics, but discovery of physical interactions the priors did not encode. Across the four regimes, the substrate independently surfaced two such phenomena. Both are consistent with the published physics literature; both were arrived at from data alone, with priors identical across voxels.
Operational physics for cislunar IMF assumes the L1 measurement convects unchanged to lunar distance. The substrate confirmed this with high certainty outside the magnetotail. Inside the magnetotail, the L1-to-lunar coupling flipped sign autonomously as ballistic propagation broke down — arriving at the published tail-lobe physics result without prior coding.
With neutral priors (W = 0) for the Ap → dose relationship across all seasonal voxels, the substrate registered persistent over-predictions during active heliospheric storms in the ls_90–180 voxel. It settled on a negative coupling at Z = 1.000 — independently surfacing the Forbush decrease, where CMEs sweep away background galactic cosmic rays and lower surface dose against operator intuition.
A single-coupling forecaster forced to average across voxels would learn W ≈ 0 in both cases — neither signal — and would be confidently wrong in both regimes. The substrate's voxel structure makes the regime-dependence structurally unavoidable.
The four benchmarks share a structural feature that turns the substrate from a forecaster into a measurement-gap diagnostic. Edges in voxels with sparse coverage stay at the prior; the framework's W = 0 in such a voxel tells you not that the coupling is null but that the data does not yet say. The pattern of where Z-convergence stalls per regime maps directly to a per-regime instrumentation prescription, prioritized by mission-autonomy criticality.
| Regime | Top Z-convergence gap | Recommended instrument or data path | Autonomy criticality |
|---|---|---|---|
| LEO | Storm-time density resolution at Dst ≤ −200 nT; GRACE-FO aging | Secondary accel-density platforms (CHAMP-successor, MAGIC, CIRCE) producing GRACE-FO-equivalent error signals | Medium |
| GEO | 1–50 keV MPS-LO band; multi-year SEP-event coverage | Operational JSON publication of GOES-R MPS-LO L1b (data-distribution gap, not a new instrument) | Medium-high |
| Mars | Driver-side continuity post-MAVEN | ESCAPADE partnership access (arrives 2027); longer-term continuous Mars-distance SW/SEP spectrometers | High |
| Cislunar | Surface dose; composition resolution; post-2012 CRaTER pathway | CLPS dosimetry; Artemis crew dose archive release; ESCAPADE loiter-phase access | High |
Three of the four recommended actions are data-distribution gaps rather than new-instrument gaps. GOES-R MPS-LO L1b exists; it is not yet published in a time-aligned operational feed. Post-2012 CRaTER data exists; it requires team outreach to access. ESCAPADE will measure Mars-distance solar wind; the partnership engagement is the gating step. Only the LEO redundancy recommendation requires actual new instrument flight. Publication-side investment may yield more mission-autonomy lift per dollar than new instrument programs for the next 3–5 years.
LEO covers 7.5 years of GRACE-FO; Mars 1 year of MSL/RAD; GEO 7 days of the current SWPC rolling window; cislunar 1 month over the May 2024 G5 superstorm. Multi-year backfill is the obvious next step for the three shorter regimes — bounded by disk space and ingest engineering, not data availability.
Across the four regimes, converged-and-signed edges show 67% correct sign at GEO (4 of 6 voxels on dst → B-field magnitude) and 67% signed correct plus 83% null-held on cislunar no-coupling-expected edges. The substrate has not yet been challenged with adversarial test data designed to flip a sign at high Z. That is a worthwhile follow-on falsifiability study.
What would falsify the architecture: a regime where edges with strong a-priori physics expectations consistently converge to wrong-sign W at high Z. None of the four regimes showed this pattern. Had the substrate not surfaced the magnetotail-decoupling or the Mars Forbush coupling from data alone, the autonomous-discovery claim would not hold. Both did, on real archive data, in single prequential passes.
Every operational space-weather forecast received today — Kp predicted, REFM fluence forecast, NAIRAS dose estimate — is a point prediction with no per-forecast trust signal. The substrate's per-edge certainty is exactly that, evolved per (voxel, time) and conditioned on the operational model. Operators see different trust levels for different parts of the same forecast, and where the substrate's certainty stays low, the gap diagnostic tells them why.
Substrate riding on NRLMSISE-00 as prior surfaces storm-window anomaly flags with 92% precision against MSIS errors and +4.6 day median lead time on the strongest storms. Direct use case: conjunction-analysis confidence intervals, lifetime-budget refresh, maneuver scheduling under storm risk — with a per-forecast trust gate the operator can act on.
Substrate prediction tracks GOES-19 daily fluence at log-MAE 0.043 on the REFM-overlap window, with REFM as prior; per-LT-voxel certainty surfaces pre-midnight growth-phase substorm risk separately from afternoon-side. Direct use case: thruster-firing scheduling during high-charging windows, deep-dielectric precursor awareness for sensitive payload modes.
No operational dose model has a public time-aligned archive at Gale Crater. The substrate provides the best-available prior from MSL/RAD plus the per-edge structure that future Mars missions can deploy on-board to learn the local SSA as their own telemetry comes online — against the priors the substrate already converged. Direct use case: rover-instrument scheduling during SEP windows, future crew shelter timing, dose-budget tracking that improves as the mission collects data.
The cislunar regime is where Artemis and CLPS are heading and where the operational community has no time-aligned model. The substrate provides the voxel-aware prior (magnetotail-vs-outer regime distinction surfaced from ARTEMIS data alone) and the on-spacecraft learning structure that lets Artemis crew transit and CLPS surface missions improve their local SSA as they collect new telemetry. Direct use case: voxel-aware transit dose forecasting, comm-degradation prediction during magnetotail passage, autonomous local calibration without ground-loop latency.
The per-edge primitive is small and deterministic. The cross-regime substrate as benchmarked here carries roughly 300–400 edges total across all four regimes, with a full flight-memory footprint of about 50 kB — the serialized edge state plus the learning loops that update W and Z on each observation. Update budget per observation is microseconds on a RAD750-class radiation-hardened flight CPU. The architecture is split into two loops that share byte-identical primitives.
Ingests public archives plus spacecraft telemetry where available. Maintains the per-edge environmental graph: driver → environmental observable, voxelized per regime. This is what the four-regime benchmark exercises. Converges priors that are then deployed to the inner loop on-board.
Same primitives, mechanical driver-observable graph: battery degradation, attitude-controller bias, propulsion drift per burn, thermal cycles, gimbal residuals. Starts from the outer loop's converged priors and learns the vehicle-specific structure from on-board telemetry the spacecraft already produces. No retraining cycle, no model rebuild — outer-loop updates arrive over the comm link as compact per-edge (W, Z) messages.
The structural property that makes this matter for Mars and cislunar: missions in those regimes cannot wait for ground-loop physics-modeling cycles to catch up. The substrate's per-edge structure means a spacecraft can learn the local SSA from its own telemetry, against the best-available priors converged on the ground — without depending on continuous comm, without GPU compute, and without exporting raw telemetry off the vehicle. Outer-loop pilots are available today; the inner-loop deployment is the architectural follow-on this work points at.
The walk-through runs the substrate against an operator's regime — constellation drag forecasting, charging-anomaly precursors, surface dose, cislunar transit — and shows where it flags, where it stays silent, and how its calibrated certainty compares against the operator's current forecast process.