GEO benchmark — Nervous Machine on GOES-19 + REFM comparator The GEO-regime sibling of the LEO Zenodo benchmark at benchmark/. Validates the substrate against the operational SWPC REFM forecast model for >2 MeV electron fluence, plus an internal prior-W comparator and a falsifiable architecture-test pass. Self-contained: framework math mirrored in nm_primitives.py; regime-portability claim is that no MCP-server changes were required to go from LEO drag to GEO trapped-particle / magnetic environment. Headline results From a single-pass prequential (predict-then-update) run over a 7-day SWPC rolling window (5 observables × 6 LT voxels × 12 drivers = 360 edges, 17,809 obs records, 175,450 edge updates): Tier-1 internal comparator: anomaly-flag precision 45.2%, lift 6.99× over the prior-W baseline (analogous to LEO's 7.90× over MSIS). Residual variance reduced 82.9% on B-field, 34.8% on warm-plasma — see results/internal_comparator.md. Tier-2 external comparator (REFM): framework log-MAE 0.041 vs REFM Day-1 log-MAE 0.169 on the 7-day overlap (4× lower). REFM publishes PE=0.15 all-time / 0.42 last-30-d as its own skill — see results/tier2_refm_comparison.md. Tier-3 falsifiable architecture test: dst → b_field_magnitude edges converge with the correct sign in 4/6 voxels (67%, vs 50% chance); null-expected edges hold |W| ≤ 0.10 in 59% of cases. SEP/CME edges show zero in-window event coverage (0/1966 records ≥10 pfu) — the test honestly reports decay-tail-only data, not architecture failure. See results/tier3_sign_convergence.md. Lead-time analysis: 10/11 SWPC alert events in the obs window were flagged at SEVERE-tier (≥5σ) before alert onset, median lead +103.8h. See results/lead_time_by_severity.md. What's here File Role nm_primitives.py Framework math (ε, η, ΔW, Z evolution), byte-identical to the LEO benchmark's copy. Standalone. fetch_goes.py Pull 7-day rolling SWPC JSON for GOES + DSCOVR + indices, plus REFM tabular text extract_obs_jsonl.py Time-align drivers, voxelize by LT, parse SWPC alerts → SEP/CME drivers, emit per-(t, voxel, observable) JSONL learn_geo.py Streaming single-pass training; per-observable composition (log-space multiplicative for fluxes, additive-residual for B-field); writes results/preds.jsonl + results/trajectory.jsonl analyze_internal.py Tier-1 comparator: evolved-W vs prior-W, 2×2 contingency + per-observable variance reduction analyze_refm.py Tier-2 comparator: framework e_flux_gt_2mev vs SWPC REFM Day-1 forecast, daily-fluence aggregation analyze_sign_convergence.py Tier-3 falsifiable architecture test: published a-priori signs vs converged W analyze_lead_time.py Per-event lead time by severity tier (LEO parity) ../sep_alerts.py Shared SWPC alerts parser (also used by Mars) results/ Committed small artifacts: training summary, final edge state, comparator tables Method, briefly Observables (5 of 5 in geo/prior.yaml): e_flux_gt_2mev (>=2 MeV electrons), p_flux_gt_10mev, p_flux_gt_50mev (SEP-relevant protons), b_field_magnitude (GEO magnetic field total), e_flux_warm_plasma (79 keV — proxy for the 1-50 keV MPS-LO band declared in prior.yaml; L1b MPS-LO is the eventual swap-in once a NetCDF ingest layer is added). Drivers (12 of 12): DSCOVR plasma (sw_speed, sw_density), DSCOVR mag (imf_bz, imf_bt), GOES-EXIS (xrs_long, mgii_index), planetary indices (kp_index, dst), and four event-derived drivers from SWPC alerts.json — sep_proton, relativistic_electron, flare_xclass, geomag_storm — each in [0, 1] with exponential decay from onset (see ../sep_alerts.py for the Space Weather Message Code mapping). Voxelization: 6 local-time bins (lt_0_4 … lt_20_24). GOES-19 at ~75.2°W sweeps all 6 voxels every 24 h. Composition modes per observable. Flux observables use log-space multiplicative coupling — log10(p) = log10(bl) + (Σ d·W)·σ_log, clamped to ±3 OOM — because flux residuals span many decades during storms and naive linear coupling saturates W at ±1 with many active drivers. B-field uses additive-residual — p = bl + (Σ d·W)·σ_linear — because GEO B varies only ~±30 nT around ~100 nT and a multiplicative form would saturate similarly. See PREDICTION_FORM in learn_geo.py. ε convention: dimensionless per-voxel-observable z-scored residual. For multiplicative (flux): ε = (log10 p − log10 o) / σ_log; for additive (B-field): ε = (p − o) / σ_linear. Framework's magnitude-tolerance constants (W_STEP, Z_BIAS_TOL, Z_STD_TOL) are rescaled in learn_geo.py to match the z-score ε scale, same rescale as the LEO benchmark. Evaluation: single online prequential pass (predict-then-update). Each prediction uses edge state before the observation; the edge then updates. Three-tier metric contract The LEO benchmark has one tier-2 metric available — MSIS exists as THE operational standard for thermospheric density. The GEO regime has REFM for >2 MeV electrons but no equivalent for protons / B / hot plasma. The benchmark therefore reports three metric tiers: Tier What Where 1 Internal: evolved-W vs prior-W (W=0) baseline. Strict subset comparator — any precision lift is from learned coupling. Always producible. results/internal_comparator.md 2 External: SWPC REFM Day-1 forecast for e_flux_gt_2mev. Operational baseline with published PE. Wirable for 1/5 observables. results/tier2_refm_comparison.md 3 Falsifiable architecture: per-edge sign match against a-priori expected signs from operational forecasting literature. results/tier3_sign_convergence.md External comparators surveyed but not wired in this commit: AE9/AP9 (IRENE) — provides climatological percentile baselines for trapped-electron/proton flux at GEO; would give a tier-2 comparator for the other electron / proton observables. Requires either pyirene (no pip release at time of writing) or spacepy + IRBEM (heavy C dependency tree). Out of scope for this commit; the REFM tier-2 + tier-1 internal + tier-3 architecture test is the working publishable contract. CCMC iSWA / SWMF runs — case-study models for specific events; not a continuous operational baseline. Data Large intermediate files (~50 MB obs JSONL + ~17 MB trajectory) are not committed; they regenerate from public archives. Committed: results/edges_state.json (~40 KB), results/training_summary.md, results/internal_comparator.md, results/tier2_refm_comparison.md, results/tier3_sign_convergence.md, results/lead_time_by_severity.md. All endpoints used by fetch_goes.py are public — no auth, no rate limits encountered at hourly cadence: SWPC primary GOES feeds — 7-day rolling JSON SWPC products (DSCOVR, Kp, Dst, alerts) — rolling-window JSON SWPC text REFM — relativistic-electron-fluence-tabular.txt (60 days), relativistic-electron-fluence-statistics.txt (published PE skill scores) Reproduce pip install requests certifi python3 fetch_goes.py # ~10 s python3 extract_obs_jsonl.py # ~2 s python3 learn_geo.py # ~30 s python3 analyze_internal.py # ~5 s tier-1 python3 analyze_refm.py # ~5 s tier-2 python3 analyze_sign_convergence.py # ~1 s tier-3 python3 analyze_lead_time.py # ~10 s LEO parity Approximate total compute on a 2024 Apple-silicon laptop: ~1 minute. Window-coverage and scope The 7-day SWPC rolling window is the current publishable window. It contains 45 deduped SWPC operational alerts (11 onset within the obs period), enough for lead-time statistics, but no actively-rising SEP events — proton flux stays at quiet background (0/1966 records ≥10 pfu). SEP/CME alert-derived drivers therefore carry decay-tail intensity only; their architecture-test sign convergence requires events-in-window evidence. Multi-month NCEI backfill from the GOES-R AWS S3 NetCDF archive (s3://noaa-goes16/SEIS-L1b-MPSL/…, ditto SGPS/MAG) is the obvious next step for parity with LEO's 7.5-year storm catalog (21 storms ≤−100 nT). NCEI's goes-r-series-l2-operational-space-weather-products/ tree was created but is empty as of 2026-06; the L1b S3 path is working but requires a netCDF4 + xarray ingest layer that is out of scope for this commit. What this means for operators The natural readers for this section are GEO satellite operations teams (Intelsat, SES, Inmarsat, Iridium, USSF GEO assets), NOAA SWPC operational forecasters, and spacecraft anomaly-response analysts. What you get that you don't have today. Operational REFM is a single Day-1 forecast number with no calibrated uncertainty and no LT structure. This substrate produces (a) a calibrated per-edge certainty Z per LT-voxel per observable, (b) a signed coupling W per driver that's voxel-aware (pre-midnight growth-phase risk is structurally different from afternoon-side and the substrate represents this directly), and (c) a 4× lower log-MAE than REFM Day-1 on the overlap window with the same input data REFM uses. The substrate's anomaly flag fires when both Z ≥ 0.85 AND |ε| ≥ 2σ — the Z gate is the part that's missing from current operational tooling. Concrete operational uses. Deep-dielectric-charging precursor flagging with per-LT-voxel resolution — currently 67% sign-correct on dst → b_field_magnitude causal edges (vs. 50% random); the unconverged voxels are exactly where the operator wants more measurement (§"Skeleton boundary" in results/tier3_sign_convergence.md). Charging-anomaly EVA-analog timing. Delay non-critical thruster firings during high-Z charging windows; the substrate gives an hours-to-days lead time with calibrated confidence rather than a daily binary forecast. Pre-midnight growth-phase substorm awareness. The substrate's per-LT-voxel state surfaces the LT-structured risk that operational forecasts average over. This matters for sensitive-mode scheduling on payloads near substorm onset sectors. Lead-time results from the current window: 10/11 SWPC operational alerts in the obs window flagged at SEVERE-tier ahead of alert onset, median +103.8h. See results/lead_time_by_severity.md. Inner/outer architecture context. This benchmark is the outer (environmental) learning loop. The same primitive math is intended to deploy on the spacecraft flight computer as an inner (mechanical) learning loop — battery degradation under thermal/charge cycling, attitude-controller bias under storm-driven torques, etc. — using the outer loop's converged per-edge (W, Z) as the inner loop's prior. Together the two loops bound the operator's worst-case unknown unknowns from above (environment surprise) and below (vehicle surprise); either alone is insufficient for assured autonomy. A first-stage operator pilot is outer-loop-only; the inner-loop deployment is a follow-on deliverable. What an operator pilot looks like. 3-6 month shadow run: operator continues to act on REFM and existing tooling; substrate publishes per-edge state + flags to a shared dashboard; at end of pilot, we compute lead-time-vs-current-tooling histograms and per-anomaly trace reports (which driver was attributable, in which LT voxel, at what Z). The substrate runs against the operator's archive data — no spacecraft modification required for the outer-loop pilot. Contact: heidi@everychart.io. Provenance Framework primitives in nm_primitives.py are byte-identical to the LEO benchmark's copy. The regime-specific code is fetch + extract glue, the SWPC alerts parser (../sep_alerts.py), and the four tier analyzers. The architecture portability claim — same primitives, different regime — is what this directory is here to test, and the head-of-table edge counts (208/360 converged, 0% W-saturation) confirm it transfers cleanly. The framework math here mirrors the implementation in ~/NM-learning-loop/mcp_validation.py (the operational MCP server). If the two disagree, the MCP server is authoritative.