# Validation status
> **On tolerances.** "Validated" below means a calibration test asserts the simulated
> statistic matches the reference relation within an **enforced gate**, currently
> **20–25% relative error** (seed-averaged) for the Allan/noise-budget relations. Median
> observed agreement is often much tighter (a few percent), but the *guaranteed* bound is
> the gate, not the median. Where a number like "~2%" appears it is a typical observation,
> not the enforced tolerance. See [Claims vs. reality](#claims-vs-reality--quick-reference).
| Noise term | Status | Evidence |
|------------|--------|----------|
| Allan estimator parity (ADEV/MDEV/TDEV/OHDEV) | `validated` | Two primary-source checks against **NIST SP 1065** (Riley, *Handbook of Frequency Stability Analysis*, 2008), both to a **1e-4** relative tolerance. (1) `tests/allan_reference.rs`: the overlapping ADEV, modified ADEV, time deviation, and overlapping Hadamard estimators reproduce the deviations for the canonical 10-point **NBS14** data set (**SP 1065 Table 29/30, p. 107**) at tau = 1, 2 — agreement is actually ~1e-6 (e.g. OADEV(2) 85.952868 vs 85.95287). (2) `tests/allan_nist_sp1065_1000point.rs`: the same four estimators reproduce the **SP 1065 §12.4 1000-point data set** (**Table 31, p. 108**) at averaging factors 1 / 10 / 100, where the data set is regenerated in code from the SP 1065 LCG (Eq. 73) — hermetic, no fixture. The same data and Table-31 numbers are the regression target in aewallin/allantools (`tests/nbs14`), reproduced here with no third-party code. This pins the *estimator maths* against the reference, distinct from the noise-*calibration* rows below. |
| Allan estimator parity on **real measured hardware** | `validated` | `tests/cs5071a_reference.rs`: the overlapping Allan and overlapping Hadamard estimators reproduce the **Stable32** reference deviations for a **real 5071A caesium primary standard measured against a hydrogen maser** (556 990 phase samples at τ₀ = 1 s; 1 PPS into a 53230A counter, Feb 2014; collected by A. Wallin, distributed with `allantools`). Checked across the **whole decade τ ladder (16 averaging factors, τ = 1 s … 4×10⁴ s)**: well-sampled factors to a **1e-3** relative gate (observed agreement is actually ≤ 3e-5, e.g. OADEV(1 s) 3.317111e-10 vs Stable32 3.3171e-10), long-τ sparse factors required to land inside Stable32's printed 1‑σ confidence band. This complements the synthetic NBS14 parity above with a measured-clock check. The raw phase file is third-party data with no redistribution licence, so it is git-ignored; `scripts/fetch_cs5071a.sh` reproduces it and the test prints a skip notice (CI stays green) when it is absent. Only the public Stable32 reference numbers are committed (`tests/fixtures/cs5071a/`); no third-party code is used. |
| Allan estimator parity on the **canonical Stable32 PHASE.DAT** | `validated` | `tests/phasedat_reference.rs`: the overlapping Allan, modified Allan, and time-deviation estimators reproduce the **Stable32** reference deviations for **PHASE.DAT** — the 1001-point series distributed with Stable32 that independent tools (e.g. `allantools`) use as their standard regression target — across the **full averaging-factor ladder (139 factors)** to a **1e-3** relative gate (observed ≤ 5e-5). A second, independent real reference series alongside the measured-caesium check above. PHASE.DAT ships with the commercial Stable32 tool, so the raw series is git-ignored; `scripts/fetch_phasedat.sh` reproduces it and the test skips green (CI stays green) when it is absent. Only the public Stable32 reference numbers are committed (`tests/fixtures/phasedat/`); no third-party code is used. |
| Optical-clock stability on a **real measured curve** | `validated` | `tests/optical_clock_adev_reference.rs`: the clock-noise parameter inversion (`quantum_trade::qparams_from_adev_curve`) reproduces the **published ⁸⁸Sr optical-lattice-clock σ_y(τ) of Norcia et al., _Science_ 366:93 (2019)** (Zenodo 10.5281/zenodo.3382347, **CC-BY-4.0**, curve vendored at `tests/fixtures/optical_clock_adev/`): the recovered white-frequency coefficient √q_wf ≈ 4.33e-16 matches the paper's ≈4.7e-16/√τ short-term stability (~10% RMS over the curve) and the fit detects a non-zero measured floor (q_rw > 0). This pins the **measured-ADEV → noise-parameter inversion against a real optical clock**; the holdover-to-threshold benefit built on it stays MODELLED. |
| Anomaly-detection ROC AUC on **real ESA OPS-SAT telemetry** | `validated` (reproduces-labels) | `tests/opssat_ad_reference.rs`: on the **OPSSAT-AD** dataset (Ruszczak et al. 2025, Zenodo 10.5281/zenodo.12588359, **CC BY 4.0**) — real OPS-SAT housekeeping segments with ground-truth anomaly labels — Kshana's Mann–Whitney ROC AUC (`impairment_eval::auc`) reproduces **scikit-learn `roc_auc_score`** to **< 1e-9** on the held-out test split (529 segments, 113 anomalies), for two deterministic scores: a transparent single-feature detector (segment peak count, **AUC ≈ 0.851**) and an 8-feature diagonal-Mahalanobis score fitted on the normal training segments (AUC ≈ 0.556); `eval_stats::bootstrap_auc_ci` brackets the point estimate with its lower bound clearing chance. This validates the **AUC computation on real ESA data** and a transparent detector's **labelled separation** — it does **not** reproduce the OPSSAT-AD paper's best published metric (a supervised FCNN, F1 ≈ 0.95), which needs their trained model. CC-BY data vendored (`tests/fixtures/opssat/`, hermetic); oracle regenerated by `generate_opssat_oracle.py`. |
| Gravity-field functional synthesis (gravity-aided / GNSS-free nav map) | `validated` | `tests/icgem_gravity_reference.rs`: the spherical-harmonic gravity-functional kernel (`gravity_sh::gravity_magnitude` / `gravity_disturbance_mgal`) — the "map reader" a gravity-aided navigator matches against — is validated against the **GRS80 normal-gravity standard** (Moritz 1980, IAG-adopted). Building the GRS80 normal field from its published even zonal harmonics J₂…J₈ and synthesising gravity (gravitation + centrifugal) on the ellipsoid reproduces the closed-form **Somigliana** normal gravity, and the published equatorial/polar values γ_e = 9.7803267715 / γ_p = 9.8321863685 m/s², to **3.5e-12** relative across all latitudes (gate 1e-9). The same kernel loads the flagship **real ICGEM Earth model EGM2008** (NGA public-domain via ICGEM; the `from_gfc` reader is also exercised on the GRAIL lunar field in `tests/agency_lro.rs`) and produces a physically-bounded gravity-disturbance map (RMS ≈ 26 mGal, max ≈ 89 mGal at d/o 70). Validates the **synthesis code correctness**; gravity-aided *navigation fix accuracy* stays MODELLED (no public gravimeter-on-a-platform trajectory with ground truth). |
| Noise-type-specific edf (confidence intervals) | `validated` | `src/allan.rs`: the χ² confidence band on each overlapping-ADEV point uses the noise-type-specific effective degrees of freedom (NIST SP 1065 Table 5 closed forms for WPM/FPM/WFM/FFM/RWFM — the Stable32 simple set), with the record's power-law type identified from its MDEV log-log slope. Three independent checks: (a) the five formulas reproduce hand-evaluated values at N=64,m=4 to 1e-12; (b) a **Monte-Carlo white-FM ensemble** (4 000 records) measures the estimator's actual chi-squared edf = 2·mean²/Var(σ²) and matches the formula within 20% (and materially beats the conservative non-overlapping count); (c) `tests/allan_nist_sp1065_1000point.rs` reproduces the printed **SP 1065 Table 32 (p. 109)** edf = 146.177 for the 1000-point data set at m=10 (white FM) to 5e-3, and the table's 95% confidence bounds (8.223942e-2 / 1.035201e-1) to <0.2% (the small residual is the Wilson-Hilferty χ² approximation vs NIST's exact χ²). The exported ADEV curve carries the identified noise type, edf, and 95% band per τ. |
| White FM (short-term) | `validated` | `tests/calibration.rs`: simulated overlapping ADEV reproduces published sigma_y(1 s) — typically within a few percent, **enforced gate 25%** — and the white-FM curve sigma_y(tau)=sigma_y(1s)/sqrt(tau) across tau = 1, 10, 100 s within the same 25% gate (matches CSAC datasheet 3e-10 / 1e-10 / 3e-11). |
| Random-walk FM (long-term) | `validated` | `tests/calibration.rs`: simulated ADEV matches sigma_y^2(tau)=q_rw*tau/3 (Riley NIST SP 1065) to ~20% (seed-averaged). |
| Aging / linear drift | `modeled` + calibrated-out | Deterministic; the holdover estimator removes offset and aging via a quadratic predictor, so the residual is the stochastic limit. Tested in `src/estimator.rs` / `src/models.rs`. |
| Kalman estimator + integrity bound | `validated` | `src/kalman.rs`: a two-state (phase, frequency) filter whose exact van Loan process noise matches the truth model; coasting reproduces the analytic holdover variance `q_wf*T + q_rw*T^3/3` to 1e-9. The covariance update is in **Joseph stabilised form** `P⁺ = (I−KH)P(I−KH)ᵀ + KRKᵀ`, which stays positive-semidefinite (Cholesky-checked) through 500 predict/update steps at an extreme `R=1e-26 / Q≈1e-30` ratio, and agrees with the naive form to 1e-9 where the latter is well-conditioned. The run layer reports Integrity as the fraction of outage samples inside the 3-sigma protection bound (`src/run.rs`). |
| Filter self-consistency (NIS / NEES) | `validated` (model) | `src/filter_health.rs`: a Monte-Carlo consistency assessment (Bar-Shalom §5.4). The matched filter's pooled **NIS → 1** and **NEES → 2** land inside their 95% χ² bands (the NIS band uses the pooled white-innovation count; the NEES band uses the independent-run count, since estimation errors are temporally correlated). A **Q/R-mismatch sweep** test verifies the monitor reports `consistent=true` only at unit tuning and `consistent=false` at ×0.1/×0.5/×2/×10 mistuning, with NIS scaling as ≈1/q_factor as derived. Surfaced as `filter_health` in the clock result JSON and a playground card. The χ² bands use the Wilson–Hilferty quantile (`detection::chi2_inv_cdf`, table-checked). |
| Two-way time-transfer model | `validated` (model) | `src/timetransfer.rs`: the reciprocal common-mode delay cancels exactly in the `(m_AB - m_BA)/2` estimate (hand-derived, exact) and two one-way measurements average to `1/sqrt(2)` (seed-averaged); the non-reciprocal differential delay is a colored white-FM + random-walk-FM process whose Allan deviation follows `sigma_y^2(tau) = q_rw*tau/3` (seed-averaged, ~20%, via the link's own `step()`). The *parameters* (per-link `sigma_j`, `q_rw`) are representative TWSTFT/optical figures, not fitted to a specific terminal. |
| Security (spoof-detection score) | `validated` (model) | `src/security.rs`: the monitor 1-sigma floor `sqrt(r/m + q_wf*tau + q_rw*tau^3/3)` and the resulting `k`-sigma minimum-detectable offset and `[0,1]` score are hand-derived and unit-tested. The model (innovation / RAIM-style clock-aided detection) is sound; the *parameters* (monitoring window `tau`, detection multiplier `k`, measurement noise `r`) are representative, not fitted to a specific receiver. |
| Spoofing-attack detection | `validated` (model) | `src/spoof.rs`: a ramping false-time spoof is flagged when its offset exceeds the clock's detection bound; the detection time matches the hand-derived `start + bound/rate` to one grid step, and a clock whose bound exceeds the spec lets the spoof reach the threshold undetected. Same `representative-parameters` caveat as the Security score. |
| Flicker FM (floor) | `modeled` (off by default) | Synthesised as a sum of log-spaced Ornstein-Uhlenbeck processes calibrated to a configurable flat ADEV floor; `src/models.rs` validates the floor is flat across averaging time and sits at the configured level (seed-averaged). Enabled per clock via `flicker_floor`. The cited reference scenarios leave it off: CSAC is white-FM-dominated across its datasheet range (1-1000 s) and the optical-clock systematic floor (~5e-17) is represented by its accuracy figure. |
| Monte Carlo ensemble statistics | `validated` (aggregation) | `src/ensemble.rs`: nearest-rank percentiles and mean over `runs` reproducible realizations; a single run collapses the spread, the per-timestep band is ordered `p05 <= p50 <= p95`, and reruns are bit-identical. This is statistical aggregation over the already-validated single run, not a new physical model. |
| Clock | sigma_y(1 s) | Source |
|-------|--------------|--------|
| csac-sa45s (CSAC) | 3.0e-10 | Microchip SA65 / SA.45s datasheet |
| optical-sr-lattice (Sr lattice) | 1.0e-15 | strontium optical lattice clock, space-oriented goal, arXiv:1503.08457 |
**Status: white FM and random-walk FM validated; aging modeled and calibrated-out; flicker FM modeled and validated (off by default in the cited scenarios).**
Maturity: the optical-clock figures are the *space-oriented goal* on ground hardware --
no strontium optical clock has flown. Laboratory Sr clocks reach 4.8e-17 (Oelker et al.
2019, Nature Photonics). CSAC figures are from a deployed commercial part.
Relations: Riley, NIST SP 1065, Eq. 67 -- white FM sigma_y^2(tau)=h0/(2 tau);
random-walk FM sigma_y^2(tau)=(2 pi^2/3) h_-2 tau, equivalently q_rw*tau/3 for a
frequency Wiener process of diffusion q_rw.
## Pack 2 — inertial dead-reckoning (quantum-IMU)
| Term | Status | Evidence |
|------|--------|----------|
| Constant/residual bias -> position | `validated` | pos error = 0.5*b*T^2 family (Groves AESS Tutorial); hand-derived discrete test in `src/inertial.rs`. |
| Velocity random walk (white accel) | `validated` | `src/inertial.rs`: simulated position-error SD matches sigma_x(T)=sqrt(S_a*T^3/3) (Groves eq.54) to ~12% (seed-averaged). |
| CAI accelerometer physics (first-principles) | `validated` (model) | `src/inertial/quantum_imu.rs`: a three-pulse Mach-Zehnder cold-atom interferometer. Hand-verified against textbook geometry — Rb-87 `k_eff = 4pi/lambda ~ 1.611e7 rad/m`; `Phi = k_eff*a*T^2 ~ 1.580e4 rad` at 1 g, `T = 10 ms`; projection noise `sigma_Phi = 1/(C*sqrt(N)) = 2e-3 rad` at `C = 0.5`, `N = 1e6`; per-shot `sigma_a ~ 1.24e-6 m/s^2`; the `1/T^2` and `1/sqrt(N)` scaling laws; exponential contrast decay. `q_va()` **derives** the white-acceleration PSD the classical `AccelModel` consumes. Honest scope: the **quantum-projection-noise floor** only (~0.09 ug/sqrtHz here) — far below real, vibration-limited devices (1-50 ug/sqrtHz); the vibration tensor, Coriolis and light-shift systematics are not yet modelled (see [`QUANTUM.md`](QUANTUM.md)). |
| Gyro bias + angular random walk -> tilt -> gravity coupling | `modeled` (off by default) | A residual gyro bias and ARW drive an attitude error; the tilt couples gravity (`g*theta`) into horizontal specific-force error. `src/inertial.rs` validates the pure-bias cubic position growth exactly and the ARW attitude growth as a Wiener process (seed-averaged). Enabled per sensor via `gyro_bias` and `q_arw`; off in the cited accelerometer-only scenarios. |
| Bias instability (1/f flicker floor) | `modeled` (off by default) | `src/inertial.rs`: a 1/f flicker process (the same OU synthesis validated for the clock) whose flat Allan-deviation floor sits at the bias-instability coefficient; zero is a no-op, a non-zero floor changes the trajectory, both reproducible. Enabled per sensor via `bias_instability`. |
| Acceleration random walk (rate random walk) | `validated` | `src/inertial.rs`: the bias is a Wiener process with `Var(bias_rw(T)) = q_aa*T`, checked seed-averaged to ~20%. Enabled per sensor via `q_aa`. |
| NaveGo cross-validation (IMU noise coefficients) | `validated` | `tests/navego_imu_crossval.rs`: an external cross-check against **NaveGo** (R. Gonzalez, open-source INS/GNSS toolbox, `github.com/rodralez/NaveGo`). Reproduces the synthetic round-trip of NaveGo's `navego_example_allan.m` on its **published Microstrain 3DM-GX3-35 reference profile**: driving our (NBS14/Stable32-validated) overlapping-ADEV estimator with white sensor noise at NaveGo's 1-σ levels recovers the velocity- and angle-random-walk coefficients (`ADEV(1 s) = σ·√dt`) to **< 5%** (actual ~0.4%), with the white-noise branch at the −1/2 slope. Confirms our Allan pipeline and NaveGo's VRW/ARW definitions agree. The 40 MB recorded STIM300 `.mat` log is not ingested (binary-format-gated). |
| Scale factor, finer cross-axis terms | `not modeled` | Scale-factor and cross-coupling errors are future work. |
| Sensor | bias stability | noise root-PSD | Source |
|--------|----------------|----------------|--------|
| cold-atom-quat (quantum) | 5.88e-7 m/s^2 (60 ng, 24 h) | 22 ug/sqrtHz = 2.16e-4 (m/s^2)/sqrtHz | Templier et al. 2022, Science Advances (arXiv:2209.13209) |
| nav-grade-quartz (classical) | 1.57e-3 m/s^2 (~160 ug) | ~20 ug/sqrtHz = 1.96e-4 (m/s^2)/sqrtHz | Honeywell QA-2000 / Groves AESS Tutorial |
Honest framing: the cold-atom advantage is **long-term bias stability** (~2600x lower), which dominates a long GNSS outage via 0.5*b*T^2. Short-term noise is comparable (quantum ~22 vs classical ~20 ug/sqrtHz) — the quantum sensor wins the marathon, not the sprint. Maturity: cold-atom accelerometers are laboratory/early (JRC122785), navigation-grade quartz is deployed.
## Pack 3 — time transfer (optical vs RF, optical inter-satellite link)
| Term | Status | Evidence |
|------|--------|----------|
| White timing jitter -> sync precision | `validated` | `src/timetransfer.rs`: simulated sync RMS reproduces the link jitter sigma_j; sample-mean averages as sigma/sqrt(N) to <20% (seed-averaged). |
| Timing -> one-way ranging | `validated` | range = c * dt, c=299792458 m/s; 1 ps = 0.299792458 mm (exact, hand-derived test). |
| Flicker/TDEV floor, two-way reciprocity residual | `not modeled` | Jitter is modeled as white; long-averaging floors and reciprocity residuals are future work. |
| Link | single-sample jitter | type | source |
|------|---------------------|------|--------|
| optical-isl | 1 ps (1e-12 s) | on-orbit-credible target | optical inter-satellite link, picosecond sync target; lab O-TWTFT ~1 fs (Giorgetta 2013 / Deschenes 2016) |
| twstft-rf | 0.5 ns (5e-10 s) | measured single-session | BIPM/PTB/NIST TWSTFT |
Honest framing: the optical figure is a picosecond-level on-orbit synchronization target (not flown). The terrestrial optical lab floor is ~1 fs (far better); a well-engineered microwave link (ACES MWL, ~0.3 ps) can rival optical, so the "RF = 0.5 ns" baseline is specifically ordinary TWSTFT. Ranging conversion is one-way (range = c*dt); two-way/round-trip halves it (range = c*dt/2).
## Pack 4 — hybrid fusion (capstone)
Composes Pack 1 (clock), Pack 2 (inertial), and Pack 3 (time-transfer) into one PNT
suite. The suite must keep BOTH timing (< timing spec) and position (< position spec)
within bounds; `pnt_holdover_s` is the time until either breaches. Optional optical
inter-satellite time-transfer re-syncs the clock during the outage (time aiding only —
position is not re-synced, since time transfer gives time, not position).
| Aspect | Status | Evidence |
|--------|--------|----------|
| Combined PNT scoring (timing AND position) | `validated` | `src/hybrid.rs` hand-derived `score_hybrid` test (pnt_holdover = first of timing/position to breach). |
| Integrity + Security for the hybrid pack | `validated` | `src/hybrid.rs`: a Kalman timing estimator disciplined to truth (nominal) and re-anchored at each optical re-sync; Integrity is the protection-bound containment (bound includes the link-jitter floor), Security the spoof-detection score. Tested for both suites including the link-aided case. |
| Joint fusion estimator | `validated` (model) | `src/fusion/mod.rs`: a single joint Kalman filter ([phase, freq] ⊕ [pos, vel]) is the navigator, disciplined by GNSS (learning the offsets from non-zero initial covariance) and aided by time transfer; the joint covariance gives a joint integrity that is tested reliable (≥0.9) for both suites with noise-consistent sensors. The pack observes position and time **separately**, for which the optimal filter is block-diagonal; augmented-state constant-bias estimation and dynamic cross-aiding are future work. |
| Coupled clock+position filter (cross-block covariance) | `validated` (model) | `src/fusion/coupled.rs` `CoupledPntFilter`: a single stacked `[pos, vel, phase, freq]` Kalman filter (Joseph-form updates) whose **pseudorange** measurement `ρ = g·pos + c·phase + noise` couples the position and clock blocks. Tests: a shared pseudorange drives `P[pos,phase]` non-zero (decoupled filters keep it exactly 0); two distinct geometries jointly resolve injected position and clock offsets a single range cannot separate; a **clock-only** fix sharpens the **position** through the cross-covariance (the payoff coupling provides and decoupled filters cannot); and a Monte-Carlo **NEES is χ²(4)-consistent** (Bar-Shalom §5.4, run-based band). 1-DOF (the pack's dimensionality); not yet wired into the runnable pack, and the 3-D 8-state extension is future work. |
| Composition of validated sub-models | inherits | clock/inertial/time-transfer terms are validated in their own packs. |
| Loosely-coupled GNSS/INS (`gnss-ins` pack) | `validated` (model) | `src/fusion/{gnss_ins_ekf,closed_loop,pack}.rs`: a 15-state error-state EKF (δp, δv, ψ, accel/gyro bias; Groves §14) drives the three-axis strapdown navigator with feedback. Tests: an injected position error nulls under perfect-truth aiding; the aided solution stays metre-bounded while a free-running INS diverges past 100 m; and the fused position RMS over a 60 s outage beats unaided dead-reckoning by ~4× (ensemble). Honest limit: loosely-coupled accel-bias/tilt are weakly separable, so per-bias calibration is not claimed. |
| Tightly-coupled GNSS/INS (pseudorange) | `validated` (model) | `src/fusion/gnss_ins_ekf.rs` `update_tightly_coupled` + `closed_loop.rs` `fuse_tightly_coupled`: the innovation is formed in the range domain (line-of-sight Jacobian on δp). Tests: four satellites null an injected 8 m/−5 m horizontal error to < 0.1 m; **two satellites** (no PVT fix possible) still cut the horizontal error by > 5×; a single overhead satellite observes only the along-line-of-sight (vertical) component and leaves the horizontal untouched; malformed inputs are rejected. Pseudorange-only — carrier phase and an explicit receiver-clock state are roadmap. |
| Sensor cross-aiding fidelity (full Kalman/factor-graph fusion) | `not modeled` | This is a system-level composition + time-aiding, not yet a full optimal estimator. |
Result: the all-quantum suite holds full PNT through a 1.8 h outage; the all-classical
suite is **position-limited** (nav-grade IMU breaches first). Optical ISL time-transfer
keeps even the classical CLOCK locked, isolating the inertial sensor as the classical
suite's weak link — the core argument for quantum inertial + optical timing together.
## Geometry — GNSS availability from orbits
| Aspect | Status | Evidence |
|--------|--------|----------|
| Circular two-body propagation | `validated` | `src/orbit.rs`: period `T=2 pi sqrt(r^3/mu)` (mu = 3.986004418e14), position returns after one period, equatorial/polar planarity — hand-derived tests. |
| Eccentric (Keplerian) propagation | `validated` | `src/orbit.rs`: Kepler's equation `M = E - e sin E` solved by Newton (residual < 1e-12); perigee/apogee radii `a(1∓e)`; circular case matches the closed-form path to 1e-9. |
| J2 secular nodal/apsidal drift | `validated` | `src/orbit.rs`: `Omega_dot = -1.5 n J2 (Re/p)^2 cos i`, `argp_dot = 0.75 n J2 (Re/p)^2 (5cos^2 i - 1)` (Vallado); node regresses (prograde) / advances (retrograde) / is stationary (polar), and apsides freeze at the critical inclination 63.4 deg — hand-derived sign/zero tests. Two-body + secular only, not osculating. |
| Line-of-sight visibility (Earth occultation + elevation mask) | `validated` | Antipodal sat occulted, radially-outward sat at 90 deg elevation, tangential sat on the horizon — exact hand-derived tests. |
| Visibility -> GNSS state -> timeline | `validated` | `>=4` visible = nominal, 1-3 degraded, 0 denied; Walker-delta generator; integration test drives a clock-holdover run from the derived timeline. |
| Constellation-design optimiser + streets-of-coverage geometry | `validated` (model) | `src/walker.rs`: `optimize_walker_design` searches the `{planes × sats × inclination}` grid and returns the cell that best meets a `DesignObjective` (min satellites for a coverage target / max coverage / min worst-PDOP) — a test asserts the pick equals an independent brute-force scan of the same sweep, and a worked example confirms a GPS Walker 24/6/1 design covers at least as well as a thinned 18-satellite one. The analytical **streets-of-coverage** closed forms (`coverage_half_angle_rad` `λ = arccos(Re/r·cos ε) − ε`; `street_half_width_rad` `cos c = cos λ / cos(π/s)`, Rider/Beste) are hand-verified against textbook geometry (GPS at a 5° mask: λ ≈ 71.16°, 4-sat street half-width ≈ 62.83°) and detect the under-population gap (`λ < π/s ⇒ None`). The full Rider minimum-satellite *global*-coverage solver (the seam-sensitive plane count) is not yet implemented. |
| Dilution of precision (GDOP/PDOP/HDOP/VDOP/TDOP) -> position accuracy | `validated` | `src/orbit.rs`: `Q=(HᵀH)⁻¹` from the line-of-sight design matrix. **External oracle** (`tests/dop_reference.rs`): the five DOP factors reproduce those computed by **gnss_lib_py 1.0.4** (Stanford NAV Lab — an independent, peer-reviewed GNSS library, used here as ERFA is for the frames) across **8 geometries** spanning well-conditioned to near-singular (DOP in the hundreds), to **1e-6** relative; provenance + regeneration in `tests/fixtures/dop/`. As an additional closed-form check, a regular-tetrahedron geometry reproduces the analytic DOPs (PDOP 1.5, TDOP 0.5, GDOP √2.5, HDOP √1.5, VDOP √0.75) to 1e-9. Position sigma = PDOP × user-equivalent range error. |
| Multi-constellation availability | `validated` | `src/orbit.rs`: additional `[[constellations]]` are merged into one satellite set for visibility/DOP; a combined-count test confirms the union. |
| Broadcast ionosphere (Klobuchar, IS-GPS-200) → L1 slant delay | `validated` | `src/gnss_sim.rs`: `klobuchar_delay_m`, the eight-coefficient broadcast single-frequency model. **External oracle** (`tests/klobuchar_reference.rs`): the slant L1 delay reproduces **RTKLIB's `ionmodel`** (tomojitakasu/RTKLIB, `src/rtkcmn.c` — the de-facto open GNSS reference, compiled from source and run independently) across **10 cases** spanning elevation, azimuth, local time and two coefficient sets (kshana's default + RTKLIB's), to **< 1e-4 m**. The IONEX TEC-map *reader* for real 80-column GIM files (and a real-CODE-GIM cross-check of the bilinear/time interpolation) is a tracked follow-up; the broadcast model itself is externally validated. |
| RAIM/ARAIM detection kernel (χ² / non-central χ² / normal) → thresholds & multipliers | `validated` | `src/raim.rs`: the snapshot / solution-separation / ARAIM detection threshold (`chi2_quantile`), the missed-detection non-centrality `pbias=√λ` (`noncentral_chi2_cdf`), and the K_fa/K_md/K_V multipliers (`normal_quantile`) are all evaluated from a dependency-free regularized incomplete-gamma function. **External oracle** (`tests/raim_reference.rs`): **171 cases** reproduce **SciPy 1.17.0** (`scipy.stats.chi2`/`.ncx2`/`.norm` + `optimize.brentq` — an independent Cephes/Boost implementation, a different algorithm) across the P_fa / P_md / redundancy operating ranges, to **≤ 1e-6 relative**. The geometry that wraps the kernel (the slope/covariance from `(GᵀG)⁻¹`) is the gnss_lib_py-validated DOP kernel. **Scope:** this pins the *statistical kernel*; the protection-level *value* and the ARAIM MHSS budget allocation are not pinned to a published reference value (see the standards note below). |
| SBAS protection level (DO-229E Appendix J) → HPL / VPL | `validated` | `src/sbas.rs`: `sbas_protection_level`, the weighted-LS `D=(GᵀWG)⁻¹` → horizontal error-ellipse major axis + vertical σ, scaled by the DO-229E K-factors. **External oracle** (`tests/sbas_reference.rs`): kshana's HPL reproduces the **RTKLIB SBAS-PL fork** (`zsiki/rtklib_ws` `waasprotlevels()`, the Siki & Takács 2017 "DO-229D Appendix J" implementation), run by `rnx2rtkp -ws` on **real EGNOS data** (GEO PRN120 broadcast messages + real BUTE/Budapest RINEX, 2017-02-19), across **6 epochs** to **< 2e-3 m**. The σ per satellite is the oracle's own full SBAS budget (UDRE + GIVE iono + tropo + airborne). ESA gLAB v6.0.0 (`core/filter.c`) was compiled and confirmed to use the identical convention. **K_V note:** both oracles round the MOPS K_V → 5.33; kshana uses the exact `Φ⁻¹(1−5e-8)=5.3267` (~0.06 % smaller VPL), so the vertical is validated as the K-factor-free `d_U` (= oracle VPL / 5.33). HPL is matched directly (K_H = 6.0 on both sides). |
| Real constellation geometry from TLEs (line-2 / Keplerian) | `validated` (parsing) | `src/tle.rs`: parses the line-2 mean Keplerian elements (semi-major axis from the mean motion); a known ISS element set round-trips to the correct elements and period. For a line-2-only block, propagation is the engine's two-body (+ optional J2) of the *mean* elements — sound for a snapshot from a common epoch, drifting from SGP4 over time. |
| SGP4 / SDP4 propagation (full TLEs) | `validated` | `src/sgp4.rs`: full near-Earth SGP4 + deep-space SDP4 (lunar-solar secular/periodic + 12 h / 24 h geopotential resonance). Validated in `tests/sgp4_verification.rs` against the official AIAA 2006-6753 ("Revisiting Spacetrack Report #3") vectors — all 666 reference states (near-Earth, deep-space, resonant, error-code cases) match to a worst-case position error of ≈ 4.1 mm. A full two-line set (line 1 + line 2) is propagated with SGP4; line-2-only stays Keplerian; the two can be mixed. Frame is TEME, used consistently for the user and satellites (no TEME→ECEF reduction — adequate for availability/DOP geometry). |
| IAU 2000A / 2000B nutation + TEME→GCRS reduction | `validated` | `src/nutation.rs`: the **full IAU 2000A** series (678 luni-solar + 687 planetary terms, `nutation_iau2000a`) and the 77-term 2000B truncation (`nutation_iau2000b`) are each validated **bit-for-bit** against the published SOFA/ERFA reference vectors at JD_TT 2453736.5 — `eraNut00a` gives Δψ = −0.9630909107115518e-5, Δε = 0.4063239174001679e-4; `eraNut00b` gives Δψ = −0.9632552291148363e-5, Δε = 0.4063197106621160e-4 — both to **1e-13 rad**. The 2000A table is machine-generated from the ERFA `nut00a` source by `tools/gen_nut00a.py` (reproduces the committed `nutation_iau2000a_data.rs` bit-for-bit). The nutation matrix (`iauNumat`), equation of the equinoxes, and the full TEME→TOD→MOD→GCRS chain (`teme_to_gcrs`, Vallado AIAA-2006-6980) are tested for proper-rotation/round-trip/precession-plus-nutation properties; 2000A vs 2000B agree to < 1 mas as required. The chain is equinox/GMST-based; the equinox-free CIO (X,Y,s) reduction and its independent ANISE/SPICE numerical cross-check (≤ 3.6 m at GNSS orbit, well inside the < 10 m target) are in the CIO row below. |
| IGRF-14 geomagnetic main-field model | `validated` | `src/igrf.rs`: the IAGA Schmidt-normalised spherical-harmonic field (degree/order 13, 2025.0 + secular variation; coefficients machine-generated from the official `igrf14coeffs.txt` by `tools/gen_igrf.py`, reproducing `igrf_data.rs` bit-for-bit). Validated on five independent fronts: (1) exact coefficient spot-checks vs the IAGA file (g₁⁰=−29350.0, g₁¹=−1410.3, h₁¹=4545.5 nT); (2) the degree-1 synthesis reproduces the **exact closed-form tilted dipole** field `(B_r,B_θ,B_φ)` to 1e-6; (3) the **full degree-13 analytic field equals −∇V** of the scalar potential (finite-difference, to 1e-4) — exercising the Legendre derivatives and the 1/sinθ term end-to-end; (4) the dipole axis reproduces the known **geomagnetic north pole** (~80.7°N, −72.7°E) and **dipole strength** (~29.7 µT), and the global field lies in the physical 22–67 µT band with the correct hemisphere dip sign; and (5) **external oracle** — the full synthesis reproduces the official **British Geological Survey IGRF-14 web-service** values at 2025.0 (geodetic, WGS-84 altitude) to **< 2 nT** on X/Y/Z/F and **< 0.02°** on D/I at four points including 400 km altitude (`synthesis_matches_the_official_bgs_igrf14_values_at_2025`). |
| CIO-based IAU 2006/2000A GCRS↔ITRS reduction | `validated` | `src/cio.rs`: the equinox-free celestial-to-terrestrial chain. The CIP coordinates `(X, Y)` and the 66-term CIO-locator `s` series are validated **bit-for-bit** against the published SOFA/ERFA `eraXys06a` vector at JD_TT 2453736.5 — X=0.5791308482835292617e-3 (1e-14), Y=0.4020580099454020310e-4 (1e-15), s=-0.1220032294164579896e-7 (1e-18); the GCRS→CIRS matrix against `eraC2ixys` (all nine elements to 1e-12); the Earth rotation angle against `eraEra00(2400000.5, 54388.0)` = 0.4022837240028158102 (1e-12). The `s` table is machine-generated from the ERFA `s06.c` reference by `tools/gen_s06.py` (reproduces `cio_s06_data.rs` bit-for-bit). The full `gcrs_to_itrs_matrix` (CIO `eraC2tcio` = polar-motion · ERA · GCRS→CIRS) is tested for proper-rotation + round-trip, and shown consistent with the legacy equinox/GMST-1982 TEME reduction up to the documented ≈2·(equation of equinoxes) sidereal-convention difference. **Independent third-party cross-check (`xval/anise-frames/`):** the same `gcrs_to_itrs_matrix` is compared against **ANISE** (the pure-Rust NAIF/SPICE reimplementation) rotating GCRF→ITRF93 from JPL's `earth_latest_high_prec.bpc`, with **identical IERS `finals2000A` Earth-orientation parameters fed to both sides**, over eight quarterly epochs 2020–2023. The two independent frame realizations agree to a **maximum relative rotation of 0.028″ — ≤ 0.86 m on the ground, ≤ 0.93 m at LEO, ≤ 3.6 m at GNSS orbit** (mean angle 0.023″). This is the ROADMAP "< 10 m" frame cross-check, **delivered**; the residual is the expected ITRF93-vs-IERS-2010-CIO model/datum difference, not a defect (the bit-for-bit anchor remains the SOFA/ERFA vectors above). ANISE is an MPL-2.0 / edition-2024 crate, so the check lives in a standalone, workspace-excluded sub-crate that never touches the published `kshana` dependency graph or any default CI gate. |
| SP3 export round-trip (constellation → SP3-c → re-parse) | `validated` | `tests/sp3_export_roundtrip.rs`: the real Celestrak `gps-ops` snapshot (30 GPS satellites) is propagated with SGP4, exported to SP3-c (`Sp3File::from_propagators` → `to_sp3_string`), re-parsed (`parse_sp3`), and the recovered ECEF positions are compared against the SGP4 truth at every epoch over 24 h — worst residual **< 0.5 m** (the writer's millimetre serialisation; the milestone's 10 m TLE-grade tolerance is met with large margin). The CLI `--export-sp3` path is covered too: the bundled orbit scenario exports a re-parseable 30-satellite SP3, and a non-orbit scenario is rejected. |
| Real GPS constellation geometry from TLEs (SGP4 → ECEF) | `validated` | `tests/igs_real_data.rs`: a genuine **Celestrak `gps-ops` snapshot** (2021-07-28, 30 operational GPS satellites; provenance in `tests/fixtures/celestrak/NOTICE`) is parsed and each satellite propagated through the validated SGP4 core to a common instant, rotated TEME→ECEF. Asserts the full constellation lands on the GPS MEO shell within 1%, and that from a mid-latitude open-sky site the real all-in-view (nine satellites) yields PDOP 1.64 with the vertical dilution exceeding the horizontal (as it must for a ground user). This exercises SGP4 on a real constellation, alongside the SP3 (precise) and RINEX (broadcast) real-data paths. |
| SGP4 mean-element Walker generator | `validated` | `src/walker.rs`: a designed Walker-delta pattern is emitted as SGP4 mean elements (Kozai mean motion chosen so the SGP4 semi-major axis lands on the target shell) and propagated through the validated core. Tests: a 24-satellite GPS-like Walker sits on the MEO shell within 1%, the planes are spaced exactly 90° in RAAN recovered from the orbit normals, and the mean motion reproduces the ~718 min GPS-shell period. |
| Constellation PDOP sweep + coverage/revisit FoMs | `validated` (monotonicity) | `src/walker.rs`: `pdop_sweep` tabulates coverage and median/worst PDOP over a `{planes × sats × inclination}` grid; `coverage_revisit` reports the coverage fraction and revisit gaps at a ground point. Validated by the physical monotonicities a design trade must obey: adding satellites never lowers coverage and strictly lowers the median PDOP (a full 24-satellite design covers continuously at PDOP ~1.7), and densifying a thinned constellation strictly shrinks the worst revisit gap. The mean revisit gap never exceeds the max. These are geometry trade tables, not a certified link/coverage budget. |
Honest framing: this is a deterministic geometry layer (circular orbits, spherical
Earth of mean radius 6371 km, pure line-of-sight). It establishes *availability* and
the *geometric* position accuracy (dilution of precision × a representative range-error
budget) from real geometry, not a precise-ephemeris navigation solution. The
`orbit-gnss-challenged.toml` reference puts a spacecraft inside the GNSS shell: it holds
a fix only ~59% of the day, the quantum clock keeps a 5 ns timing solution through every
gap (availability 1.0) while the chip-scale clock holds ~0.83.
## Numerical (Cowell) propagator & force model
The numerical propagator (`src/propagator.rs`) integrates a hierarchical force model
(`src/forces.rs`) with the adaptive integrators (`src/integrator.rs`). Each term is
validated against analytic truth or a hand-derived closed form, not against another
tool; the perturbations are **off by default**, so the released goldens are untouched.
| Term | Status | Evidence |
|------|--------|----------|
| Cowell propagator vs analytic Kepler truth | `validated` | `src/propagator.rs`: the unperturbed (two-body) orbit reproduces the **exact universal-variable Kepler solution to sub-metre over a 24 h LEO orbit** (a tighter gate than "vs a numerical reference < 10 m"); specific energy and angular momentum conserve to ~1e-9 relative; the J2 nodal regression reproduces the closed-form `j2_secular_rates` to first-order theory (within 2 %, the O(J2²) residual). `solve_kepler_checked` returns `Err` rather than a silently-wrong answer when Newton fails to converge (near-perigee e = 0.999). |
| Adaptive integrators (RK4 step-doubling + Dormand–Prince RK5(4)) | `validated` | `src/integrator.rs`: RK4 integrates `y' = y → e` to < 1e-9, shows the ~16× error drop per halved step (4th-order convergence), and conserves the harmonic oscillator over a period. The DP5(4) embedded error estimate is **O(h⁵)** (halving the step cuts it ~32×), integrates the oscillator over 50 periods conserving energy to < 1e-6, and reaches the same endpoint at the same tolerance in **fewer function evaluations** than step doubling; `propagate_dopri` clears the same sub-metre Kepler gate and agrees with the RK4 path to < 1 m on a J2..J6 orbit. |
| J2–J6 zonal-harmonic field | `validated` | `src/forces.rs`: `zonal_accel` is checked three ways — it **reduces to the 666-vector-validated `j2_accel` to machine precision** when restricted to `[J2]`; it **matches the numerical gradient of its own zonal potential** through the full J2..J6 field (the conservative-field gold standard); and the odd `J3` vs even `J2`/`J4..J6` terms show the characteristic north–south (anti)symmetry under `z → −z`. A propagated J2..J6 orbit conserves total energy to ~1e-8 over a day. |
| Third-body (Sun and Moon) gravity | `validated` | `src/forces.rs` + `src/ephem.rs`: `third_body_accel` **matches the exact gradient of its disturbing potential**, vanishes at the geocentre, and hits the textbook LEO magnitudes (~5e-7 m/s² Sun, ~1.1e-6 m/s² Moon). The low-precision Montenbruck–Gill Sun ephemeris hits hand-derived J2000 anchors (perihelion distance ≈ 1.471e11 m, ~−23° solstice declination, ~1°/day motion); the Moon ephemeris stays inside its perigee/apogee envelope, recovers the ~384 400 km mean distance, never exceeds the 5.3° inclination, and returns to within 1° after one sidereal month. The epoch-driven RHS wiring is **bit-exact** (the RHS term equals `third_body_accel` at the sampled position at t = 0 and t = 1 day), and a quarter-year epoch shift yields a different trajectory. |
| Solar-radiation pressure + conical shadow | `validated` | `src/forces.rs`: `srp_accel` (cannonball `ν·P☉·cᵣ·(A/m)·(AU/d)²·d̂`) is **bit-identical** to the closed form in full sun, pins the **1-AU radiation pressure to ≈ 4.5398e-6 N/m²**, sits in the ~1.36e-7 m/s² LEO band pushing away from the Sun, quarters when the Sun distance doubles (inverse-square), and is **exactly zero** deep in the umbra. `conical_shadow` gives `ν = 1` in full sun and `ν = 0` in total umbra (exact), a smooth monotonic penumbra rising 0 → 1 across the `[b−a, b+a]` band that **extends beyond the umbral cylinder**. |
| Atmospheric drag (first velocity-dependent term) | `validated` | `src/forces.rs`: `atmospheric_density` (Vallado Table 8-4 piecewise-exponential) **anchors at 1.225 kg/m³ at sea level**, clamps below the surface, decreases monotonically through LEO, sits in the ~1e-12 kg/m³ band at 400 km with a physical recovered ~58 km scale height; `drag_accel` opposes the co-rotating relative velocity at the ~2e-6 m/s² LEO magnitude. The signature check: a 300 km orbit loses specific energy **monotonically** and its semi-major axis decays a bounded ~km/day, where the vacuum baseline conserves energy to < 1e-9. |
| Post-Newtonian (Schwarzschild) relativistic correction | `validated` | `src/forces.rs`: `relativistic_accel` (IERS `β = γ = 1` form `a = (μ/c²r³)·{[4μ/r − v²]·r + 4(r·v)·v}`) collapses on a circular orbit to the closed form `3μ²/(c²r³)·r̂` (radial, **outward**, off-axis components exactly zero), shows the textbook **`≈1.9e-9` LEO ratio to two-body**, matches the hand-simplified radial-velocity form `μ(4μ + 3v²r)/(c²r³)` to < 1e-12, and in the propagator **perturbs the orbit while holding the semi-major axis to under a metre/day** — the conservative opposite of drag's decay. |
Honest scope: the released `ForceModel` defaults integrate the zonal field. The
high-degree EGM2008 **tesseral** field (`src/gravity_sh.rs`, degree/order 70), the
solid/ocean/atmospheric **tides** (`src/tides.rs`), and the **Lense–Thirring**
frame-dragging term are now implemented and composed into the precise
ephemeris-fitting force model (`src/precise_od.rs` — see below). The NRLMSISE-00 thermospheric
density (drag is the static Vallado model), solar limb darkening / the oblate-Earth
shadow, DE-grade ephemeris accuracy, and an external GMAT/Orekit cross-validation of a
high-fidelity run remain follow-ons.
## Tides on the geopotential (`src/tides.rs`)
Time-varying corrections ΔC̄_nm, ΔS̄_nm to the Stokes coefficients (IERS Conventions
2010, Chapter 6), validated against the conventions' own published numbers — not against
another tool.
| Contribution | Status | Evidence |
|------|--------|----------|
| Solid Earth tide (IERS Eq. 6.6 / 6.8b / 6.14) | `validated` | `tests/tides_iers.rs`: the permanent (zero-frequency) tide ΔC̄₂₀ lands within 1 % of the IERS-published −4.1736×10⁻⁹ (anelastic Love numbers); the **Step-2 K1 worked example is reproduced bit-for-bit** (sin/cos amplitudes 470.9 / −30.2 ×10⁻¹² to 0.1×10⁻¹²); the normalized Legendre path is pinned by closed-form hand values independent of the recurrence. |
| Ocean tide (IERS Eq. 6.15, FES2004, 8 constituents) | `validated` | The vendored FES2004 subset (`tools/gen_fes2004.py` → `src/fes2004_data.rs`, Doodson→multiplier parse) reproduces the source M2 (2,2) coefficients exactly; the K1 Doodson phase equals θ_g + π; the degree-2 sectorial magnitude is physical and an order below the solid tide. |
| Atmospheric S2 air tide (Ray 2001, NASA GSFC) | `validated` | The vendored Ray harmonics convert through the surface-load relation (Eq. 6.21) to a peak ΔC̄₂₂ ≈ 4×10⁻¹¹ (~10 % of the ocean M2 term). Honest caveat: **no published geopotential-coefficient oracle exists for the air tide**, so it is validated by source-integrity + magnitude, a weaker bar than the solid/ocean tides. |
## Maneuvers & trajectory design (`src/maneuver.rs`)
| Aspect | Status | Evidence |
|--------|--------|----------|
| Impulsive ΔV + covariance propagation | `validated` | A velocity discontinuity with a 6×6 covariance carried forward (deterministic burn ⇒ identity STM; the execution-error covariance rotates from the burn frame — ECI or LVLH — into the velocity block). |
| Finite-burn integration vs Tsiolkovsky | `validated` | Constant-thrust integration over a burn arc (mass as a state) whose achieved ΔV matches the closed-form **Tsiolkovsky** rocket equation to **better than 0.01 %**. |
| Lambert solver (Izzo 2015) round-trip | `validated` | The single-revolution Lambert output (`r1`, `r2`, time-of-flight ⇒ `v1`, `v2`) is **round-tripped through the exact universal-variable Kepler propagator** — it must land back on `r2`. |
| Porkchop sweep vs Hohmann floor | `validated` | The launch × arrival C3 / arrival-V∞ grid's minimum is checked against the **analytic Hohmann-transfer C3 floor** for two coplanar circular orbits. |
Honest scope: no trajectory optimizer, no multi-revolution Lambert branches, and a
synthetic coplanar-circular heliocentric model (no planetary DE ephemeris, so a GMAT
Earth–Mars C3 cross-check has not been run).
## Orbit determination (`src/orbit_determination.rs`, `src/batch_ls.rs`)
| Aspect | Status | Evidence |
|--------|--------|----------|
| Gauss–Newton batch corrector | `validated` | The generic `gauss_newton` solver reaches the exact weighted-least-squares solution on a linear fit, recovers true parameters on a nonlinear `a·exp(b·t)` fit, and solves a 3-D range-multilateration from noiseless ranges. |
| Batch orbit determination from ranges | `validated` | `determine_orbit_batch` recovers `[r, v]` from ground-station range tracking (propagated over the two-body + J2 force model) to **sub-metre / mm·s⁻¹ from noiseless ranges**, and to **~2 m with a post-fit residual at the 5 m noise floor** — the signature of a consistent least-squares fit. |
| Sequential (unscented-filter) OD | `validated` | `determine_orbit_sequential` recursively recovers the state to within tens of metres on the same dynamics and range model. |
Honest scope: range-rate/Doppler and angle measurements and station-visibility masking
are follow-ons for the range-only teaching estimator (the full-force ephemeris-fitting engine
below carries the variational state-transition matrix).
## Force-model validation by ephemeris fitting (`src/precise_od.rs`)
A precise, full-force position-observation batch least-squares estimator: the
EGM2008 tesseral geopotential (evaluated in the Earth-fixed frame through the CIO
reduction) composed with the Sun/Moon third body, SRP, drag, Schwarzschild/Lense–Thirring
GR, and the tides above, fit to a track of inertial position fixes with a variational
state-transition-matrix Jacobian. Validated on **synthetic** data, where the truth is
Kshana's own integrator, so every residual is the estimator's and not the dynamics'.
| Aspect | Status | Evidence |
|--------|--------|----------|
| Precise force model composition | `validated` | `tests/precise_od_synth.rs`: the degree-0 field is the exact point mass at every epoch (the central term is rotation-invariant); degree-8 adds the J2-band oblateness; the Sun third body and the tide term wire in bit-faithfully; a constant radial empirical acceleration is purely radial in RTN. |
| Variational state-transition matrix | `validated` | Every column of Φ over a half LEO orbit (degree-8 geopotential + Sun + Moon) agrees with an independent **whole-arc central finite-difference** re-propagation to **< 1×10⁻⁶ relative** in both the position and velocity response — the documented STM↔FD cross-check. Φ(0) = I; the augmented and plain propagators agree to sub-millimetre. |
| Batch-LS self-recovery | `validated` | A 1-hour Kshana arc (degree-6 + Sun + Moon) observed noise-free is recovered from a 150 m / 0.1 m·s⁻¹ offset to the **epoch state within 1×10⁻² m** with near-zero post-fit RMS; with 5 m white noise the post-fit 3-D RMS settles at the **~σ noise floor**, unbiased. SRP **C_R is recovered from 1.0 to within 1 %** of the 1.4 truth. |
| Outlier editing & RTN reporting | `validated` | A 500 m gross blunder is rejected by 5-σ post-fit editing (exactly one observation), leaving a clean millimetre-level fit; residuals are reported decomposed into radial/transverse/normal as well as 3-D. |
| Empirical-acceleration tier (RTN constant + 1-CPR) | `validated` | The nine a-priori-constrained empirical parameters stay below 1×10⁻⁸ m·s⁻² on empirical-free truth without disturbing the fit; a constant 3×10⁻⁸ m·s⁻² cross-track acceleration injected into the truth is recovered to within 20 %. |
Honest scope: the synthetic wave uses nominal Earth-orientation parameters (UT1 ≈ TT, no
polar motion) — exact for self-recovery, since the same model generates and fits the arc.
The agency-dataset validations layer **real** finals2000A EOP and SP3/SPK truth on top of
this engine through the CIO frame chain (`src/eop.rs`, `tests/agency_galileo.rs`,
`tests/agency_swarm.rs`, `tests/agency_lro.rs`).
### Real agency precise orbits (`tests/agency_galileo.rs`, `tests/agency_swarm.rs`, `tests/agency_lro.rs`)
| Dataset | Status | Evidence |
|---------|--------|----------|
| **Galileo MEO — < 5 m GREEN** | `validated` | Kshana's full-force engine fit to a verbatim slice of **ESA/ESOC's own final orbit** (`ESA0MGNFIN`, ITRF) for Galileo **E11** over 8 h, each ITRF fix rotated into GCRS with real finals2000A EOP: post-fit **3-D RMS 0.132 m** pure force + `C_R` (RTN 0.105/0.067/0.047 m, `C_R` 1.174), **0.070 m** with the empirical tier, from a 78.7 km raw overlap. The full 24 h arc is **0.611 m**. All far inside the 5 m bar. Field gravity-converged by d/o-8 (identical at d/o 8/10/12); the `workflow_dispatch` job runs the full d/o-70. Provenance + SHA-256 in `tests/fixtures/agency/NOTICE.md`; full record in `docs/AGENCY-ORBIT-VALIDATION.md`. |
| **Swarm-A LEO — < 5 m GREEN** | `validated` | Full-force engine (+ atmospheric drag) fit to a verbatim slice of **ESA's own Swarm-A reduced-dynamic precise orbit** (`SW_OPER_SP3ACOM_2_`, L47, ITRF, ~430 km, ~2 cm) over 3 h at d/o-70, real finals2000A EOP through the CIO chain: **dynamic** tier (state-only, static density, `C_R`=1) post-fit **3-D RMS 2.687 m** (RTN 0.925/**2.522**/0.043 m — residual ≈ pure along-track drag), **reduced-dynamic** tier (+ empirical CPR accelerations) **0.098 m** (RTN 0.026/0.092/0.024 m) — ~10 cm against ESA's ~2 cm orbit. Provenance + SHA-256 in `tests/fixtures/agency/NOTICE.md`; full record in `docs/AGENCY-ORBIT-VALIDATION.md`. |
| **LRO lunar — validated, above 5 m (honest)** | `validated` | **Moon-centred** force-model fit (`src/lunar_od.rs`): the GRAIL **GRGM660PRIM** gravity field (d/o 100) in the lunar body-fixed principal-axis frame (`src/lunar_frame.rs`, IAU 2015 mean-Earth + the fixed DE421 ME→PA offset) + Earth/Sun third body, fitted through the *same* generic Gauss–Newton estimator as the Earth datasets to the real **JPL Horizons LRO** (NAIF −85) reconstructed orbit, 4 h / 241 epochs, ~98 km. Honest post-fit: **dynamic 12.6 m** (RTN 3.07/10.19/6.81), **reduced-dynamic 6.6 m** (RTN 3.49/4.56/3.35, 1+2-per-rev empirical), from a 53.8 m raw overlap — **above** the < 5 m bar, not the estimator (identical at d/o 100/150 and `atol` 1e-6/1e-9). A **DE-grade cross-validation** (`xval/anise-lunar-od`: DE440 lunar PA orientation + DE440 ephemeris via ANISE) **corrected** the limiting-factor claim — it improves the dynamic fit (12.6 → 12.0 m) but leaves the **reduced-dynamic floor unchanged** (6.65 → 6.67 m), so the operational floor is **not** the analytic orientation/ephemeris but an empirical-tier-irreducible residual (most consistent with unmodelled LRO non-gravitational dynamics over the short arc); the lean analytic stack already matches DE-grade for the reduced-dynamic orbit. Provenance + SHA-256 in `tests/fixtures/agency/NOTICE.md`; full record in `docs/AGENCY-ORBIT-VALIDATION.md`. **Validated against real agency truth: 3 of 3; meeting < 5 m: 2 of 3.** |
## Gravity-map / alt-PNT navigation (`src/gravimeter.rs`, `src/mapmatch.rs`, `src/particle_filter.rs`, `src/altpnt/terrain.rs`)
| Aspect | Status | Evidence |
|--------|--------|----------|
| Spherical-harmonic gravity-anomaly field | `validated` | `src/gravimeter.rs`: a low-degree, fully-normalised field checked against the closed-form Legendre functions (`P̄₁₁ = √3·cosφ`, `P̄₂₀ = (√5/2)(3sin²φ−1)`, `P̄₂₂ = (√15/2)cos²φ`) and a hand-derived single-term anomaly of 1.897 mGal. |
| Cold-atom gravimeter measurement model | `validated` (model) | The white-noise floor is derived from the CAI accelerometer ASD (`σ = ASD/√τ`), injected as a deterministic seeded sequence (the matcher is never handed noise-free truth, yet the run is bit-reproducible). |
| Sequential-importance-resampling particle filter | `validated` | `src/particle_filter.rs`: the deterministic core is pinned exactly — ESS spanning 1…N, systematic resampling picking indices in proportion to weight, the weighted-mean convex combination, a Gaussian likelihood pulling the estimate onto the measurement, and seeded predict determinism. |
| Map-match likelihood + recovery | `validated` | `src/mapmatch.rs`: `field_likelihood` peaks (= 1) at a perfect match and falls to `e^(−½)` at one sigma; a particle filter over a distinctive synthetic-terrain patch recovers the true position to within 0.1. |
| 60-minute GPS-denied benchmark | `validated` | `run_gps_denied_gravity_nav` (`scenarios/gps-denied-gravity-nav.toml`): a ~700 km / one-hour outage where the inertial solution drifts to **≈ 70 km** is recovered to **≈ 145 m (< 500 m)** by a hierarchical coarse-to-fine matcher — bit-reproducible, stable across noise realisations, and provably refinement-limited (a single coarse grid stalls at ~2 km). |
| SRTM `.hgt` DEM loader + bilinear sample (ORACLE A) | `validated` | `src/altpnt/terrain.rs` / `tests/terrain_nav_validation.rs`: a hand-built 2×2 `.hgt` buffer with corners [100,200;300,400] bilinear-interpolates to **exactly 250.0** at the cell centre (closed-form oracle), the 16-bit-big-endian round-trip is exact, and the row-flip places the northernmost file-row at the highest stored latitude — all against the **GDAL SRTMHGT driver spec** (16-bit signed big-endian, row-major, north row first, void -32768, <https://gdal.org/en/stable/drivers/raster/srtmhgt.html>). A committed 11×11 fixture (`tools/gen_terrain_fixture.py`) exercises the parser in CI; and a **committed real public-domain NASA/USGS SRTM v3 tile** (N36W117, Death Valley, decimated to 6-arc-sec, 722 KB, `tests/fixtures/terrain/N36W117_sub6.hgt`) is read in CI (non-ignored, `real_srtm_committed_badwater_tile_reads_real_relief`), placing **Badwater Basin at ≈ −78 m within the documented [−95,−70] m band** (NGS/NOAA survey ≈ −86 m) with the tile max ≈ 2161 m — source-of-truth is the survey, the DEM is under test (non-circular). Larger published spot-heights (Mount Whitney 4421 m) remain in `#[ignore]`-gated full-tile tests. |
| Terrain-referenced navigation (TERCOM/SITAN) convergence (ORACLE B) | `validated` | `run_terrain_nav` (`scenarios/terrain-nav.toml`): a **hand-derived** injected INS drift (0.5°N, −0.4°E ≈ 70 km at ~12° lat, computed as drift° × M_per_deg × cos lat) is recovered to **< 500 m** (within the grid-resolution floor `search_step/factor² ≈ 140 m`), a > 100× cut over free-inertial — checked against the *injected* number, never the DEM (non-circular), bit-reproducible and stable across noise seeds. |
| Combined gravity+magnetic+terrain fusion gain (ORACLE C) | `validated` | `run_combined_altpnt` (`scenarios/combined-altpnt.toml`): three scalar channels (Δg · |B| crustal-anomaly · elevation) fused as a product likelihood give a **bounded < 500 m** residual over the 60-min outage that, seed-averaged, is **no worse than the best single channel** (information additivity / lower CRLB). Sits in the published TERCOM/TRN "tens of metres" CEP regime (<https://en.wikipedia.org/wiki/TERCOM>; PeerJ 2024 ESKF-TERCOM, <https://peerj.com/articles/cs-3118/>). |
Honest scope: the gravity field is low-degree + synthetic mascons, **not** the full
EGM2008/EIGEN coefficient set; the magnetic channel rides on the smooth IGRF main field plus
**synthetic** crustal-anomaly mascons (a real high-frequency crustal map is a follow-on); the
CI terrain field for the *navigation-fix* benchmarks is the self-contained synthetic DEM, while
the `.hgt` reader itself is now validated in CI against a committed real SRTM v3 tile (above);
the full-resolution real tiles remain `#[ignore]`-gated. A map-representation-error
Monte-Carlo remains a follow-on. All
three navigators are now scenario-engine `kind=` wired (`gravity-map`, `terrain-nav`,
`combined-altpnt`).
## INS / IMU error model — datasheet-referenced validation (`src/inertial/`)
The IMU stochastic error model is validated against **published manufacturer
datasheet / dataset specifications** — never against a value Kshana itself produced.
The bridge from a field-unit spec to the SI noise model is the standard
Allan-deviation identification (Riley, NIST SP 1065 §5; IEEE Std 952): the random-walk
coefficient `N` is the overlapping ADEV read at τ = 1 s on the τ⁻¹ᐟ² slope, and the
bias-instability coefficient is the flat Allan plateau (`tests/imu_allan_spec.rs`,
hermetic — no download).
| Aspect | Status | Evidence / oracle |
|--------|--------|-------------------|
| Unit-conversion layer (the #1 non-circularity risk) | `validated` | Hand-checked: 0.15 deg/√hr = 4.3633e-5 rad/√s; 3.6 µg = 3.5304e-5 m/s²; 2.0 deg/hr = 9.6963e-6 rad/s. |
| Gyro angle random walk | `validated` | ADIS16465 ARW 0.15 deg/√hr (Analog Devices datasheet) recovered from ADEV(1 s) to < 5%. |
| Accel velocity random walk | `validated` | ADIS16465 VRW 0.1 m/s/√hr recovered to < 5%. |
| White-noise Allan slope | `validated` | ADIS16460 white branch log-log slope −0.5 ± 0.05. |
| In-run bias instability plateau | `validated` | ADIS16465 gyro BI 2.0 deg/hr recovered as the flat Allan minimum to < 15%. |
| Second IMU profile | `validated` | NaveGo ADIS16488 (ARW 0.3 deg/√hr, VRW 0.029 m/s/√hr) both recovered to < 5%. |
Honest scope: this validates the **stochastic error model** against datasheet Allan
coefficients. A full strapdown + tightly-coupled EKF cross-check against the i2Nav-WHU
**KF-GINS** vehicle dataset is a follow-on — the published static-RTK segment converges
to the cm-level antenna position, so a meaningful navigation-accuracy comparison there
first requires GNSS/IMU lever-arm compensation and a dynamic free-inertial-divergence
window; until then it would measure a fixed geometric offset, not navigation quality.
## Deep-space & Mars PNT — D0 foundation (`src/body.rs`, `src/mars_frame.rs`, `src/ephem_provider.rs`, `src/radiometric.rs`, `src/timescales.rs`)
The first milestone of the deep-space navigation engine: generalising the Earth-hardcoded core to an arbitrary central body, adding Mars dynamics, an ephemeris-provider seam, sub-microsecond time, and the light-time / relativistic-delay primitives. **Every existing Earth scenario is byte-identical** — the cross-platform reproducibility goldens (`tests/cross_platform_golden.rs`, `tests/golden.rs`) pass unchanged with no regeneration, enforced per commit.
| Term | Status | Evidence |
|------|--------|----------|
| Central-body abstraction (Earth byte-identical) | `validated` | `src/body.rs`: `Body{mu,re,zonals,gravity,rotation,IAU-pole}` with `earth()/mars()/moon()/sun()`. `tests/mars_dynamics.rs` asserts `two_body_accel_body(r,&Body::earth())` and `zonal_accel_body` equal the legacy Earth functions to **exact bit equality** (`assert_eq!`), and a default-`ForceModel` Earth arc is byte-identical — the reproducibility goldens are unchanged (same SHA, no regen). |
| Mars gravity (GMM-3 tesseral, body-fixed) | `validated` (model) | `Body::mars_gmm3(nmax)`: zonals C̄20/C̄30/C̄40 from MRO110 J2/J3/J4 (`C̄n0=−Jn/√(2n+1)`, J2 round-trips to 1.9604e-3) + MRO110B2 sectoral/tesseral C̄22/S̄22, C̄32/S̄32 (Konopliv et al. 2011; tabulated in Liu, Baoyin & Ma 2012), evaluated in the IAU Mars body-fixed frame. `tests/mars_dynamics.rs` checks the round-trip, the J2-scale departure from point-mass, and that the field is genuinely body-fixed (a quarter-Mars-day reorientation changes the inertial acceleration). Higher degree/order loads from a vendored `.gfc` via `gravity_sh::from_gfc`. |
| Mars body-fixed frame (IAU pole) | `validated` | `src/mars_frame.rs`: the IAU 3-1-3 rotation `R3(W)·R1(90°−δ0)·R3(90°+α0)` (WGCCRE/NAIF convention, same as `lunar_frame.rs`). Tests: orthonormal to 1e-12 with det +1, round-trip identity, prime meridian advances at `prime_w_dot` per day. |
| Mars / Sun-central propagation | `validated` (self-consistency) | `tests/mars_propagation.rs`: a Low-Mars-Orbit returns to its start after one Keplerian period (closure); specific energy ε=v²/2−μ/r drifts <1e-9 across the arc; the Mars-J2 secular nodal regression matches the analytic Vallado rate (computed independently ≈ −5.15 °/day at 60° incl.) to 3%; a heliocentric orbit recovers the ~687-day Mars year (vis-viva + independent period recovery from the arc). No external data. |
| DE440 cross-validation (heliocentric Mars) | `kernel-gated` | `xval/anise-mars-od/` (workspace-EXCLUDED; MPL-2.0 ANISE confined to its own Cargo.lock; mirrors `xval/anise-lunar-od`): compares Kshana's Sun-central Mars propagation against JPL DE440 (via ANISE), Horizons truth cited verbatim (provenance + SHA-256). **Reproduced result** (seed JD 2459580.5 TDB, `de440s.bsp` SHA-256 `c1c7fee…b260a49f2`): position residual **137 m @ 1-day arc** (5.97×10⁻¹⁰ of the 2.299×10¹¹ m heliocentric distance), growing to 3.3 km @ 5 d, 87.9 km @ 30 d, 1.71 Mm @ 90 d — the residual *grows* with arc length because a Sun-central two-body model deliberately omits the planetary perturbations (Jupiter chiefly) and the Mars-system internal motion the DE440 barycenter ephemeris carries, so the growth is the honest signature of the unmodelled n-body dynamics, not an integrator error; a short arc staying a tiny fraction of the heliocentric distance confirms the Sun-central machinery is correct. Runs only when the DE440 kernel is present (`$KSHANA_ANISE_DE440S`, or auto-fetched ~32 MB) — a manual / `workflow_dispatch` DE-grade check, **not** a default CI gate, so the published crate and every default gate stay byte-for-byte untouched. |
| Light-time + Shapiro delay | `validated` | `src/radiometric.rs`: the iterative retarded-epoch light-time solution (converges |Δτ|<1e-12 s, capped) and the Shapiro delay Δt=(2μ/c³)·ln[(r1+r2+r12)/(r1+r2−r12)]. Tests: Earth–Sun ≈ 499 s and Earth–Moon ≈ 1.28 s via the builtin ephemeris; a synthetic constant-velocity emitter recovers the analytic retarded solution to <1e-9 s (solver proven independent of any real ephemeris); Shapiro matches a hand-computed value exactly and gives ~100–250 µs for an Earth–Mars superior-conjunction grazing ray. |
| Ephemeris-provider seam | `validated` | `src/ephem_provider.rs`: `trait EphemerisProvider` + kernel-free `BuiltinEphemeris` (Montenbruck–Gill Sun/Moon, geocentric); returns `None` for bodies it has no series for (e.g. Mars), for which the DE-grade ANISE provider is the path — the same builtin-vs-ANISE split as `LunarEnvironment`/`AniseLunarEnvironment`. |
| Two-part time + TT↔TDB | `validated` | `src/timescales.rs`: a `TwoPartJd{day,frac}` with Knuth two-sum accumulation recovers sub-microsecond intervals the single-f64 JD loses near J2000 (>1000× better, tested); `tt_to_tdb`/`tdb_to_tt` use the dominant Fairhead–Bretagnon periodic series (|TDB−TT|≲1.7 ms, hand-verified at J2000). Additive — the existing single-f64 time path is unchanged. |
## Deep-space radiometric observables & CCSDS-TDM — D1 (`src/radiometric.rs`, `src/ccsds_tdm.rs`)
The second milestone of the deep-space navigation engine: the **observable model** that turns the D0 light-time/relativistic-delay kernel into the actual quantities a Deep-Space-Network or ESTRACK tracking pass reports — range, one/two/three-way Doppler, coherent transponder turn-around, regenerative/PN ranging, Δ-DOR, and the propagation media — plus the **CCSDS 503.0-B Tracking-Data-Message** I/O that ingests/emits a standard agency tracking file. **Honest scope:** these are **exact geometric models** following the Moyer / CCSDS conventions; the measurement *noise* is the caller-supplied per-observation `sigma`, not baked into the model. Reproducing a *specific real tracking pass* needs real TDM data plus the deep-space orbit-determination solver — that is **D2**, not D1. Everything below is additive (the reproducibility goldens are unchanged; same SHA, no regen).
| Term | Status | Evidence |
|------|--------|----------|
| Range & 1/2/3-way Doppler (Moyer two-leg solve) | `validated` (model) | `src/radiometric.rs`: `one_way_range`/`two_way_range` compose the D0 retarded (down-leg) and advanced (up-leg) light-time solutions; the full round-trip convention is `ρ₂ = c·(τ_up + τ_down)` (m), documented and consistent. `one_way_doppler`/`two_way_doppler`/`three_way_doppler` form the carrier shift `f_D = −(M·f_ul/c)·(ρ̇_up + ρ̇_down)` (Hz) from the line-of-sight range rate, **sign convention: an approaching spacecraft gives a positive (blue) shift**. Tests: two-way range = sum of legs to <1 mm (~2 AU Earth–Sun round trip); one-way range = `c·τ` exactly; one-way Doppler matches the analytic `−(f/c)·ρ̇` for a radial constant-velocity emitter to ppt; three-way with co-located stations collapses byte-for-byte to two-way. |
| Coherent transponder turn-around (exact DSN ratios) | `validated` | `src/radiometric.rs` `turnaround_ratio`: the exact rational DSN/CCSDS turn-around numbers `M = f_down/f_up` (Moyer §13; DSN 810-005 module 201/214) — S/S 240/221, S/X 880/221, X/S 240/749, X/X 880/749, X/Ka 3344/749, Ka/Ka 3360/3599. Tested for exact equality; an undefined band pair **panics** (fail-loud, never invents a ratio). `two_way_doppler_coherent` applies the band carrier + ratio end-to-end and equals the explicit `two_way_doppler` to machine precision. |
| Regenerative / PN ranging (CCSDS 414; ambiguity) | `validated` | `src/radiometric.rs`: `regenerative_range_ambiguity` = `c/(2·chip_rate)` (the per-chip resolution-scale unambiguous range) and `pn_range_ambiguity` = that × code length (the full PN-code unambiguous range `c·L/(2·f_chip)`), per CCSDS 414.1-B. Tested: a 1 MHz range clock gives ~149.9 m; the CCSDS-414.1 weighted-voting PN code (1 009 470 chips) at 1 Mchip/s gives the ~151 000 km design unambiguous range; doubling the chip rate halves the per-chip ambiguity. |
| Δ-DOR (CCSDS 506 plane-of-sky) | `validated` (model) | `src/radiometric.rs` `delta_dor`: the differential plane-of-sky delay `Δτ = −B⃗·(ŝ_sc − ŝ_quasar)/c` (s) for an interferometer baseline `B⃗` (CCSDS 506.1-B). Tested: matches the exact analytic projection `−B·sin(Δθ)/c` for a known baseline-aligned offset to 1e-15 s, the small-angle magnitude `|Δτ| ≈ B·Δθ/c`, and the **null** (a spacecraft co-aligned with the quasar gives exactly zero differential delay). |
| Plasma 1/f² dual-frequency calibration | `validated` (model) | `src/radiometric.rs`: `solar_plasma_delay` = `K_PLASMA·TEC/(c·f²)` (the dispersive `1/f²` charged-particle group delay, `K_PLASMA = 40.3 m·Hz²·TECU⁻¹`); `coronal_tec_from_sep` the Cassini-class `TEC ∝ 1/sin(SEP)` corona column (analytic stand-in, not a calibrated corona); `dual_freq_plasma_calibration` the `K_disp = (Δt_X−Δt_Ka)/(1/f_X²−1/f_Ka²)` dispersion inversion the DSN uses to remove plasma. Tested: the 1/f² law `d_X/d_Ka = (f_Ka/f_X)²`; the corona rises toward conjunction; and an **injected TEC is recovered to <1 %** (in fact ≪ 1e-9 noise-free) — `dual_frequency_recovers_injected_plasma`. |
| Tropo / iono media (reuse GNSS models) | `validated` (model) | `src/radiometric.rs` `tropo_delay`/`iono_delay`: the deep-space ground-station segment crosses the same Earth atmosphere the GNSS pack already models, so these delegate to `gnss_sim::tropo_delay_m` (Saastamoinen-zenith + Niell-mapping, non-dispersive) and `gnss_sim::klobuchar_delay_m` (broadcast iono). Tested wired + physically signed: tropo positive and larger at low elevation than at zenith; iono a non-negative slant delay (numerical fidelity is the GNSS pack's own datasheet-referenced tests). |
| CCSDS-TDM 503 parse / emit + radiometric bridge | `validated` | `src/ccsds_tdm.rs`: a CCSDS 503.0-B-2 KVN reader **and** writer (header + `META_START…META_STOP` / `DATA_START…DATA_STOP` segments; `RANGE`/`DOPPLER_*`/`ANGLE_*`/`*_FREQ` records). Tests (`#[cfg(test)]` + `tests/fixtures/deepspace/reference.tdm`): the reference fixture parses to the expected named fields; **parse→emit→parse is an equal structure** (semantic round-trip); the bridge `to_radiometric_obs` maps the unambiguous `RANGE`/`DOPPLER_*` records to `RadiometricObs` (way from `PATH` `1,2`/`1,2,1`/`1,2,3`, band from `TRANSMIT_BAND`, km→m); a multi-segment file resets the state machine. **Honest skip-not-guess:** angles/frequencies, unknown bands, day-of-year epochs, and unrecognised time systems are **skipped**, never mis-mapped, and `sigma` is left `0.0` for the caller to weight. |
| DSN/ESTRACK benchmark (model precision ≫ published floor) | `validated (model)` | `tests/radiometric_benchmark.rs`: the observables are exact geometric computations, so the honest benchmark is that the **model's own numerical error is far below the published DSN/ESTRACK measurement floor** — the model is not the accuracy bottleneck, the caller-supplied σ is. Each test computes an observable two independent ways (or recovers an injected truth) and asserts the residual is below the floor: **range** round-trips to <1 mm ≪ the ~1 m PN-ranging floor (DSN 810-005 mod. 214 / CCSDS 414.1-B); **Doppler** matches the exact closed-form retarded range rate `v/(1+v/c)` to ~10⁻⁶ m/s, several × below the ~0.05 mm/s X-band floor (DSN 810-005 mod. 203 / Moyer) — the residual is the f64 range-differencing floor at ~1 AU, a numeric characteristic not a modelling error; **Δ-DOR** reproduces the closed-form projection to an equivalent plane-of-sky angle ≪ 1 nrad ≪ the ~1–10 nrad floor (DSN 810-005 mod. 210 / CCSDS 506.1-B); **plasma** dual-frequency recovers the injected delay to ≪ 1 %; and the **Shapiro** delay sits in the published ~100–250 µs Earth–Mars conjunction band. The published figures are stated and labelled as **order-of-magnitude** DSN specs, not fabricated exact numbers, and the module doc is explicit this benchmarks MODEL PRECISION, not real-pass reproduction (that is D2 + real TDM). |
## Deep-space reduced-dynamic OD — D2 (`src/deepspace_od.rs`, `src/clock_state.rs`, `src/mars_atmos.rs`)
The third milestone of the deep-space navigation engine: the **estimator** that turns the D1 radiometric observables into a recovered trajectory — a numerically-robust Square-Root Information Filter with reduced-dynamic empirical accelerations, a three-state onboard clock, a Mars-drag model, and the radiometric measurement partials that connect range/Doppler to the filter state, closed end-to-end by a synthetic Mars-LMO orbit-determination recovery. **Honest scope:** the recovery is a **synthetic closed-loop** (truth and filter share the same Mars propagator, and the observations are the same geometric observable model the filter inverts, plus injected Gaussian noise) — it validates the **estimator machinery**, not the absolute fidelity of the Mars force model; the DE-grade external cross-check is the kernel-gated `xval/anise-mars-od` crate (D0.8b). Everything below is additive (the reproducibility goldens are unchanged; same SHA, no regen).
| Term | Status | Evidence |
|------|--------|----------|
| SRIF core (square-root information filter) | `validated` | `src/deepspace_od.rs` `Srif`: a hand-rolled Bierman square-root information filter — upper-triangular `R` with `Λ = RᵀR`, info vector `b`, Householder measurement/time updates, bare-`Vec<f64>` arithmetic (no nalgebra). Tests: the sequential SRIF matches the batch weighted-least-squares (`gauss_newton`) solution in **both state and covariance to 1e-9** on a designed linear-Gaussian problem (`srif_matches_batch_on_linear`); the recovered covariance `P = R⁻¹R⁻ᵀ` is **symmetric (1e-12) and positive-definite by construction** after a long measurement/time-update sequence (`srif_covariance_is_spd`, Cholesky succeeds); and information accumulates (trace of `P` decreases monotonically as identical-geometry measurements are folded in). |
| Reduced-dynamic empirical accelerations (cruise↔LMO tuning) | `validated` | `src/deepspace_od.rs` `ReducedDynamicOd`: the nine-state `[r; v; a_emp]` filter with RTN empirical accelerations as first-order Gauss–Markov process states, the JPL/ESOC reduced-dynamic technique exposed as a single `dynamic_tightness` knob (near-dynamic ⇒ smooths noise on a ballistic arc; near-kinematic ⇒ the empirical tier absorbs an unmodelled manoeuvre/drag). Tests: a near-kinematic filter tracks a **stepped (thruster) manoeuvre** with <½ the residual of the near-dynamic one, while the near-dynamic filter **smooths a noisy ballistic arc** better (`reduced_dynamic_tracks_maneuver`); sweeping the tightness moves the post-fit residual **monotonically**, a continuum not a switch (`tuning_is_a_continuum`). |
| 3-state onboard clock (phase/freq/drift, van Loan Q, Allan) | `validated` | `src/clock_state.rs` `ClockState3`: phase + fractional-frequency + frequency-drift error state with the **exact van Loan discrete process noise** and a Joseph-stabilised update. Tests: the discretised `Q` matches the hand-derived van Loan polynomial **to 1e-12** and **reduces exactly to the two-state KalmanClock Q when drift = 0** (a strict superset); a multi-step coast grows the phase variance to the analytic NIST SP 1065 holdover relation through the `T⁵` random-run term; the drift state **recovers a true aging ramp to <5 %**; a **USO-class Allan profile is recovered** (overlapping ADEV at τ=1 s within the calibration gate) and the estimator stays NEES-consistent (>95 % inside 3σ) and PSD through a harsh-`Q/R` run. `ClockClass` carries the cited CSAC/USO/DSAC `σ_y(1 s)` figures and the `Δv = c·σ_y` Doppler floor. |
| Mars atmospheric drag (Mars-GRAM-lite) | `validated (model)` | `src/mars_atmos.rs`: a piecewise-exponential Mars neutral-density profile (surface ~0.020 kg/m³, near-surface scale height ~11 km, the Mars-GRAM 2010 / MCD low-dust mean) and the quadratic drag acceleration taken relative to the **co-rotating** Mars atmosphere, plus a `..._scaled` variant the SRIF can multiply. Tests: the density decreases monotonically and continuously across bands, the recovered near-surface scale height is the physical ~11 km, the drag is **strictly anti-parallel to the relative velocity** (cross-product vanishes to 1e-12) with a sane ~3e-4 m/s² LMO magnitude, and it is linear in the ballistic term. **Honest scope:** a representative engineering atmosphere, not a flight-validated Mars-GRAM/MCD with dust/season/local-time dependence. |
| Radiometric measurement partials (range/Doppler) for the SRIF | `validated` | `src/deepspace_od.rs` `range_observable` / `range_rate_observable`: the line-of-sight range `ρ = \|r_sc − r_sta\|` (∂ρ/∂r = û, ∂ρ/∂v = 0) and range rate `ρ̇ = û·(v_sc − v_sta)` (∂ρ̇/∂v = û, ∂ρ̇/∂r = (v_rel − ρ̇·û)/ρ) partials that connect the D1 observables to the filter state, plus the one-way-Doppler clock-frequency partial `∂ρ̇_obs/∂y = c`. Tests: both partials match a **central finite difference of the observable to 1e-6** (`range_partial_matches_finite_difference`, `range_rate_partials_match_finite_difference`); a single range update **shrinks the covariance in the observed (line-of-sight) direction by ≫1000×** and keeps it PD, and a Doppler update shrinks the LOS **velocity** variance (`radiometric_update_reduces_covariance_in_observed_direction`, `range_rate_update_observes_velocity`). |
| Mars-LMO end-to-end OD from radiometric data | `validated (synthetic closed-loop)` | `tests/mars_lmo_od.rs`: a truth Low-Mars-Orbit arc propagated under `Body::mars_gmm3(4)` gravity generates noisy two-way **range (σ=1 m) + Doppler (σ=0.1 mm/s)** against two tracking stations; the reduced-dynamic SRIF recovers it from a **~2.7 km-perturbed** initial state. Achieved: converged position **RMS ≈ 0.2 m** (sub-metre, far inside the metres-to-tens-of-metres done-criterion) over a 3-orbit, 30 s-cadence arc, with the **factored covariance positive-definite at every epoch** (the SRIF guarantee, asserted throughout). The **cruise→LMO tuning continuum** is exercised end-to-end: with an unmodelled-drag truth the filter template lacks, the reduced-dynamic knob (tightness 1) recovers to ~14 m vs ~677 m for the near-dynamic (tightness 0) run — a 48× improvement from the empirical tier absorbing the drag. **Honest scope:** synthetic closed-loop (truth = the same propagator the filter uses); the DE-grade external cross-check is the kernel-gated `xval/anise-mars-od` (D0.8b), never run here, no network. |
## Deep-space PNT scenario & one-way+two-way fusion — D3 (`src/mars_pnt.rs`, `src/deepspace_od.rs`)
The fourth milestone of the deep-space navigation engine: a runnable **Mars-PNT product scenario** that ties the D0 Mars body / frame, the D1 radiometric observables, and the D2 reduced-dynamic OD into a single `mars-pnt` scenario kind. It models a small **MARCONI-style relay constellation** broadcasting a one-way (clock-coupled) signal-in-space plus relaying a two-way (coherent, clock-free) link to a deep-space station, and navigates three reference users (Mars transfer, Low-Mars-Orbit orbiter, fixed surface point) through the **D3.1 joint one-way + two-way fusion estimator** — a twelve-state SRIF that augments the D2 reduced-dynamic orbit/empirical block with a three-state onboard clock. **Honest scope:** every figure of merit is the estimator's **formal covariance bound** (per-epoch 1σ / 3σ position) and the achieved RMS against a **synthetic closed-loop** truth (truth and filter share the same Mars dynamics and the same geometric observable model, plus injected Gaussian noise) — a simulated navigation FoM, **NOT** an aviation-certified protection level, and **NOT** a flight claim (no certified fault model, no integrity monitor, no real tracking data). Everything below is additive (the reproducibility goldens are unchanged; same SHA, no regen).
| Term | Status | Evidence |
|------|--------|----------|
| Joint one-way + two-way radiometric fusion (fused beats either alone) | `validated` | `src/deepspace_od.rs` `FusionOd`: the D2.2 reduced-dynamic SRIF augmented with the three onboard-clock states (`ClockState3`), ingesting a mixed time series of **two-way** (clock-free, orbit-pinning — zero clock columns in the partial) and **one-way** (clock-coupled — the `∂ρ̇_obs/∂y = c` frequency partial) range/Doppler through `fused_observable`. The calibrate-then-coast structure: two-way passes pin the orbit, one-way data then calibrates the clock and coasts the orbit between passes with error bounded by the clock's Allan stability. Tested in `src/deepspace_od.rs` `#[cfg(test)]`: the **fused** solution beats **either link type alone**; a one-way-only run's coast error grows with the gap between two-way passes consistent with the clock-class Allan bound; the joint covariance stays SPD throughout. |
| MARCONI constellation (areostationary radius, occultation) | `validated (model)` | `src/mars_pnt.rs` `MarconiConstellation`: a five-relay set — three **areostationary** relays at `r = (μ/ω²)^{1/3} ≈ 20 428 km` (the Mars synchronous radius from published μ/spin-rate; tested in the **20 400–20 500 km** band) equally spaced 120° in longitude with 5° inclination, plus two relays in a higher (1.4× areostationary) 60°-inclined circular orbit for plane-of-sky coverage. Per-epoch relay states are forward-integrated under `mars_gmm3` gravity; the **Mars-occultation** visibility test is the chord-clears-sphere geometry (`chord_clears_sphere`, tested: a diametrically-opposite relay is occulted, an overhead relay is visible). **Honest scope:** a minimal representative broadcast-plus-relay geometry built from published Mars constants — not a specific flight constellation design. |
| `mars-pnt` scenario reachable via CLI / Python / WASM / MCP | `validated` | `src/mars_pnt.rs` `MarsScenario` / `run_mars_pnt` is wired as the `kind = "mars-pnt"` scenario across all surfaces (D3.3): the CLI runner, the Python binding, the WASM/playground build, and the MCP server dispatch all route to it (`src/api.rs` round-trip test `mars_pnt_kind_round_trips_through_the_dispatch`). Reference scenarios shipped: `scenarios/mars-pnt-lmo.toml`, `scenarios/mars-pnt-surface.toml`, `scenarios/mars-pnt-transfer.toml`. The result carries an explicit `fom_note` labelling the FoM a covariance bound (not a certified PL), and the `summary` / `to_svg` outputs repeat the honesty label. |
| MARCONI / LightShip targets reproduced: **< 100 m orbiter**, **< 15 m rover** | `validated (synthetic closed-loop, under stated assumptions)` | `tests/mars_pnt_targets.rs`: the published MARCONI / LightShip programme goals (orbiter < 100 m, surface/rover < 15 m) reproduced in simulation. **Shared assumptions:** the five-relay default MARCONI constellation; every in-view relay gives a one-way range+Doppler per epoch; a coherent two-way station pass every **30 min**; DSN-class noise (**range σ = 1 m, Doppler σ = 0.1 mm/s**); the joint reduced-dynamic orbit + 3-state clock fusion filter seeded from a **~2.7 km**-perturbed a-priori state (tightness 0.1); FoM = converged back-half-of-arc 3-D position RMS vs the synthetic truth. **Orbiter** (LMO ~400 km / 60° inc, **USO** clock, 60 s cadence, ~2 h ≈ one-orbit arc): achieved converged RMS **≈ 0.40 m** ≪ the 100 m target (mean 3.1 relays in view, covariance SPD throughout). **Rover** (fixed equatorial surface point, **USO** clock, 30 s cadence, ~2 h arc): achieved converged RMS **≈ 5.1 m** < the 15 m target (mean 3.0 relays in view, covariance SPD). **Both targets met without loosening any assertion** — the USO is the *less* capable of the realistic onboard clock classes (a DSAC would only help) and the 30-min two-way cadence is conservative for a routine relay-network schedule, so the assumptions are physically fair for a MARCONI-class system rather than tuned to pass. **Honest scope:** simulated reproduction under the stated assumptions, NOT a flight claim or certified protection level; synthetic closed-loop truth shares the filter's Mars dynamics and observable model. |
## Lunar coordinate time — relativistic Earth-Moon clock rate (`src/lunar_time.rs`)
A Lunar Coordinate Time scale (LTC/TCL) requires the secular rate at which a lunar-surface clock runs relative to an Earth-geoid (TT) clock. `src/lunar_time.rs` computes it from first principles: a first post-Newtonian identity summing the **self-potential difference** (the IAU `L_G` conventional geoid potential `W0 = L_G·c² ≈ 6.26369e7 m²/s²` minus the Moon's surface self-potential `GM_moon/R_moon ≈ 2.822e6 m²/s²`) and the **kinetic (second-order Doppler) term** `−<v²>/2c²` from the geocentric Moon velocity (finite-differenced from the Montenbruck-Gill lunar series in `src/ephem.rs`). The dominant self-potential term gives **≈ 57.5 µs/day**, the kinetic term **≈ −0.5 µs/day**, for a total **≈ 57 µs/day**. The `lunar-time-offset` scenario reports this rate, the accumulated LTC−TT offset over a horizon, and supports a minimal inverse-variance ensemble (a lunar paper-clock); TT↔LTC conversions round-trip to < 1 ns/day.
| Capability | Status | Oracle / honest scope |
|------------|--------|-----------------------|
| Relativistic Earth-Moon clock rate (LTC/TCL) | `modelled` | `src/lunar_time.rs` `lunar_rate_breakdown`, `tt_to_ltc`/`ltc_to_tt`, `lunar_ensemble`, `LunarTimeScenario`. **Oracle: an internal closed-form relativistic identity** (IAU `L_G` / IERS-conventions geopotential and the GM/r self-potential), **cross-checked against the published lunar-clock-rate band [56, 59] µs/day**. Tests (`lunar_time::tests`): the self-potential term ≈ 57.5 µs/day, the Moon speed ≈ 1 km/s, the kinetic term a small negative, the total in band with the named terms summing to it, the 1-day LTC−TT offset ≈ 57 µs, and a < 1 ns/day TT↔LTC round-trip. **Honest scope:** the headline µs/day figure is **reference-dependent** — it depends on the chosen reference surfaces (Earth geoid `W0` vs a lunar selenoid), the time-averaging window, and the neglected sub-µs/day centrifugal/J₂ corrections — which is why a **band**, not a single certified number, is the published output. This is a `Modelled` self-consistency check, **NOT** an external validation, **NOT** sub-nanosecond absolute accuracy, and **NOT** certified for operational timekeeping. |
## Lunar geodetic VLBI — near-field delay for an Earth baseline observing a lunar beacon (`src/lunar_vlbi.rs`)
A geodetic VLBI observable for two Earth ground stations (a baseline) observing a one-way signal from a NovaMoon-class transmitter on the lunar surface. The Moon is **near-field**, not a plane wave, so the geometric delay is the exact two-range difference `tau_geom = (|r2 − r_B| − |r1 − r_B|)/c` rather than the plane-of-sky projection. The full observable adds the station clock-offset difference and a differenced Earth-potential Shapiro term (reused from `src/radiometric.rs`). Stations are placed in geocentric inertial (GCRS) coordinates via `src/frames.rs` + `src/cio.rs`; the beacon via `src/ephem.rs` (geocentric Moon) + `src/lunar.rs` (selenographic→MCMF) + `src/lunar_frame.rs` (IAU-2015 ME body-fixed→inertial). The `lunar-vlbi` scenario samples the delay, its finite-difference rate, and the near-field correction over a pass.
The **oracle** is the same-codebase plane-wave Δ-DOR observable `delta_dor` in `src/radiometric.rs`: in the far-field limit (a synthetic beacon at 1e15 m) the geometric delay must collapse to `−(B·ŝ_B)/c = delta_dor(r_B, [0,0,0], B)` to machine precision; at true lunar distance the wavefront-curvature near-field correction is non-zero (tens to hundreds of µs). The beacon partials `dtau/dr_B = ((r_B−r2)/|r_B−r2| − (r_B−r1)/|r_B−r1|)/c` (and the station partials) are verified by central finite difference (relative error < 1e-5).
| Capability | Status | Oracle / honest scope |
|------------|--------|-----------------------|
| Lunar geodetic VLBI delay + partials | `modelled` | `src/lunar_vlbi.rs` `geometric_delay_s`, `vlbi_delay_s`, `delay_partials_beacon`/`_station1`/`_station2`, `near_field_correction_s`, `LunarVlbiScenario`. **Oracle: `ReferenceImpl`** — the same-codebase plane-wave `delta_dor` (`src/radiometric.rs`) in the far-field limit, plus **finite-difference partials**. Tests (`lunar_vlbi::tests`): station magnitude ≈ Earth radius, beacon range at lunar distance (356–407 Mm), far-field geometric delay matches `delta_dor` to < 1e-9 s, a non-zero near-field correction at lunar distance, FD beacon/station partials within 1e-5 relative, an exact clock term, and a finite scenario run dispatched through `run_toml`. **Honest scope:** `Modelled`, **NOT** validated against real VLBI data. Polar motion is dropped (`xp = yp = 0`), so station inertial positions carry a few-metre frame error; the beacon mixes a mean-equator-of-date Moon series with an ICRF body-fixed offset (`jd_tdb ≈ jd_tt`), a frame-consistency mismatch below model fidelity but not rigorous; and there is no light-time iteration, Earth-rotation-during-light-time, media (tropo/iono/plasma) or aberration term beyond the differenced Shapiro. No TRL, flight heritage or agency endorsement is claimed. |
## Lunar joint multi-technique OD + clock — simulated closed-loop recovery (`src/lunar_combination.rs`)
A single-epoch (snapshot) batch least-squares fit (`crate::batch_ls::gauss_newton`) that **fuses the geodetic-VLBI, radiometric/lunar-local-ranging and inter-satellite-ranging techniques** on a SIMULATED lunar network to recover, *together*, a lunar surface station's 3-D position, a small constellation's per-satellite positions, and every asset's clock offset. The observables are: (1) Earth-baseline geodetic VLBI delays to the lunar station treated as the beacon (reusing `src/lunar_vlbi.rs`, geocentric inertial) — the headline technique; (2) Earth-station→satellite geocentric radiometric ranges (which multilaterate the constellation from the well-spread Earth stations) and station↔satellite lunar-local ranges (MCI Euclidean distance plus the differenced clock term); (3) inter-satellite ranges (same form); plus a single station-clock sync pseudo-observation that anchors the otherwise-unobservable common clock offset. The state vector is `[station_pos(3), {sat_pos(3)}×N_sat, station_clk, {sat_clk}×N_sat]`. There is **no force-model propagation inside the solver** — the satellites are fixed, distinct, illustrative points on a representative lunar orbit; this is a deliberately clean snapshot fit. Internally, positions are scaled and clocks are estimated in range-equivalent metres so the normal matrix stays well-conditioned in f64; the VLBI and lunar-distance ranges are mean-removed about the nominal geometry so the finite-difference Jacobian does not suffer catastrophic cancellation. The `lunar-joint-od-clock` scenario runs the solve with and without the VLBI legs on the same seed/truth and reports both solutions plus the station-observability improvement factor.
The **headline honest result**: lunar-local ranging from a polar station to a handful of satellites that all sit toward one side of the sky leaves the station's position **weakly observed along one direction**, so the range-only solve is ill-conditioned and the recovered station 3-D error is large (or the solve fails to converge). **Adding the Earth-baseline VLBI delays makes the station's full 3-D position observable**, and the recovered station error collapses to the metre level. The test `vlbi_restores_station_observability` asserts the with-VLBI station error is markedly (≥5×) smaller than the range-only error on the identical injected truth — VLBI restoring the otherwise-unobservable station direction is the phase's point.
The **oracle** is internal consistency: an injected truth state (station ~50 m/axis, satellites ~30 m/axis, clocks ~1e-7 s, with a seeded per-component jitter) is mapped to synthetic observables through the *same* geometry model the solver inverts, seeded Gaussian noise is added (illustrative defaults: VLBI σ ≈ 1e-11 s, range σ ≈ 0.1 m, ISL σ ≈ 0.1 m; weights `1/σ²`), and the estimator must recover the injected truth within the noise from a zeros initial guess. A Monte-Carlo mean NEES `Δxᵀ(HᵀWH)Δx` over seeds (`formal_covariance_nees`) sits within a loose band around `n_params`, a covariance-realism check.
| Capability | Status | Oracle / honest scope |
|------------|--------|-----------------------|
| Lunar joint multi-technique OD + clock | `modelled` | `src/lunar_combination.rs` `estimate`, `formal_covariance_nees`, `LunarCombinationScenario`. **Oracle: `InternalConsistency`** — recovery of an injected simulated truth + NEES covariance consistency. Tests (`lunar_combination::tests`): full-fusion recovers the truth (station < 5 m, sat RMS < 5 m, clocks < 1e-8 s, converged); VLBI restores station 3-D observability (station error ≥5× smaller with VLBI than range-only on the same truth); deterministic (same seed → bit-identical solution, different seed → different-but-recovered); mean NEES within a loose band around `n_params`; finite/guarded reported numbers even when range-only is ill-conditioned; scenario dispatched through `run_toml`. **Honest scope:** this is a **simulated closed-loop recovery** — the truth shares the observation model — **NOT** real-data validation; no force-model propagation inside the solver; the satellite positions and Earth/lunar-station geometry are illustrative. No TRL, flight heritage or agency endorsement is claimed. |
## Lunar reference-frame realisation — 7-parameter Helmert datum fit (`src/lunar_frame_realise.rs`)
Where `src/lunar_frame.rs` *applies* the IAU 2015 WGCCRE lunar body orientation (a forward model), this module *estimates* (realises) a lunar reference frame from a network of estimated point coordinates tied to a datum. The estimation core is a classic **7-parameter similarity (Helmert) transform** — three translation, three small-angle rotation and one scale parameter — mapping points `p_i` (estimated frame) to `q_i` (datum frame) as `q_i = t + (1 + s)·R(θ)·p_i`, with `R(θ) = rz(θz)·ry(θy)·rx(θx)` built from `crate::precession::{rx, ry, rz, matmul, mat_vec}`. The seven parameters `[tx, ty, tz, θx, θy, θz, s]` are estimated from ≥ 3 non-collinear point pairs by weighted least squares through `crate::batch_ls::gauss_newton`: the forward model predicts every `q_i` from its `p_i` and flattens all three components of all points into the observable vector, the observed datum coordinates form `z`, the weights are `1/σ²`, and the solve starts from `x0 = zeros`. Rotation angles are estimated in µrad and scale in ppb so the finite-difference Jacobian stays well-scaled at the lunar-surface coordinate magnitude (~1.7e6 m). A small `icrf_orientation_tie` composes the realised small rotation through the IAU body→ICRF orientation (`transpose(crate::lunar_frame::icrf_to_iau_moon(jd_tdb))`) and returns the realised frame's residual small-angle offset about the ICRF axes — a deliberately simple frame-of-expression change, **not** an independent estimate of the lunar pole.
The **oracle** is internal consistency: a known Helmert transform (translation ~tens of m, rotation ~µrad, scale ~1e-7) is *injected* into a well-spread synthetic point network (varied selenographic lat/lon → MCMF, some points at altitude), seeded Gaussian noise is added, and the fit must recover the injected parameters. On noiseless data the recovery is to ~machine precision (translation < 1e-6 m, rotation < 1e-9 rad, scale < 1e-12); with metre noise the recovered parameters land within a few × the formal σ and the post-fit RMS residual sits near the noise level. The `lunar-frame-realisation` scenario reports the recovered datum, the per-parameter recovery error vs the injected truth, the post-fit RMS residual, and the realised-rotation ICRF orientation tie.
| Capability | Status | Oracle / honest scope |
|------------|--------|-----------------------|
| Lunar reference-frame realisation | `modelled` | `src/lunar_frame_realise.rs` `helmert_fit`, `apply_helmert`, `realise_frame`, `icrf_orientation_tie`, `LunarFrameRealiseScenario`. **Oracle: `InternalConsistency`** — recovery of an injected 7-parameter similarity transform plus the algebraic round-trip identity of `apply_helmert`. Tests (`lunar_frame_realise::tests`): noiseless recovery of the injected transform (translation < 1e-6 m, rotation < 1e-9 rad, scale < 1e-12); with-noise recovery within a few × σ and RMS residual near the noise level; apply-then-invert round-trip; deterministic (same seed → bit-identical, different seed → different-but-recovered); zero-rotation ICRF tie is zero and the tie preserves the rotation magnitude; scenario dispatched through `run_toml`. **Honest scope:** this is a **self-consistency** check — it recovers an *injected* transform on a *synthetic* network — **NOT** a realisation of the lunar reference frame against real tracking / VLBI data, and it claims no absolute frame accuracy. The ICRF orientation tie is a simple change of expression, not an independent pole estimate. No TRL, flight heritage or agency endorsement is claimed. |
## Lunar navigation service volume — DOP / coverage / availability + generalised lunar ARAIM PL (`src/lunar_service.rs`)
This module *composes* already-built pieces into a lunar navigation **service-volume** analysis: it sweeps a selenographic latitude/longitude grid over a time horizon against an **illustrative, public-source Moonlight / LCNS-class** lunar-orbit constellation and reports DOP / coverage / availability plus a generalised lunar ARAIM protection-level envelope. The two reuses are load-bearing and explicit. **DOP geometry** is the **VALIDATED** (vs gnss_lib_py) kernel `crate::orbit::dop` — `service_dop` is a thin elevation-mask filter plus that kernel, and a test asserts the identity (`service_dop` equals a direct `orbit::dop` on the same visible positions). **Integrity** reuses the LunaNet LNIS lunar ARAIM machinery `crate::lunar::lunar_araim` (σ_URE ≈ 30 m, `P_sat` ≈ 1e-4); `lunar_protection_level` generalises the existing south-pole protection level to an arbitrary surface point, and a test asserts it **reduces to the south-pole case** (at the south pole with the same geometry/budget it returns exactly the existing `lunar::lunar_araim` result to < 1e-9).
The constellation parameters are an **illustrative approximation of public ESA descriptions** of the system class (≈ 4 satellites on elliptical lunar frozen orbits favouring south-pole coverage — apolune over the southern hemisphere). They are **not** the real Moonlight/LCNS ephemeris and imply **no** affiliation, endorsement, heritage, certification or TRL. The composition (coverage / availability / integrity) is **MODELLED**: a circular-/elliptical-Keplerian relay set (not a differential-corrected LCNS / 9:2 NRHO ephemeris), a mean-rotation Moon (no physical libration / precessing pole), and published LunaNet LNIS integrity parameters. It demonstrates the service-volume *method*, not an operational availability number. Deterministic (pure geometry; no randomness). The `moonlight-service-volume` scenario reports the coverage percentage, the visible-count and PDOP envelopes, the HPL/VPL envelope over the volume and the protection-level availability against the alert limit.
| Capability | Status | Oracle / honest scope |
|------------|--------|-----------------------|
| Lunar navigation service volume | `modelled` | `src/lunar_service.rs` `LunarConstellation`, `visible_sats`, `service_dop`, `coverage`, `lunar_protection_level`, `LunarServiceScenario`. **DOP geometry validated by reuse** of the gnss_lib_py-validated `crate::orbit::dop` kernel; integrity reuses the published-LunaNet/LNIS lunar ARAIM (`crate::lunar::lunar_araim`, σ_URE ≈ 30 m). Tests (`lunar_service::tests`): `service_dop` is an identity on the validated kernel; `lunar_protection_level` at the south pole equals the existing `lunar::lunar_araim` PL to < 1e-9 (reduces to the south-pole case); coverage is non-decreasing in constellation size; visibility honours the elevation mask; the scenario is deterministic and dispatches through `run_toml`. **Honest scope:** the constellation is **illustrative, public-source, NOT affiliated with ESA** — not the real Moonlight/LCNS ephemeris; the coverage/availability/integrity composition is **MODELLED** (circular-/elliptical-Keplerian relays, mean-rotation Moon, published LNIS parameters), **NOT** an operational Moonlight availability number. No TRL, flight heritage or agency endorsement is claimed. |
## Lunar differential PNT — common-mode cancellation + baseline-growing residual + DGNSS protection levels (`src/lunar_dpnt.rs`)
A lunar DGNSS / SBAS analogue: a **NovaMoon-class reference station** at a *known* selenographic location computes per-satellite differential corrections from an **illustrative, public-source Moonlight / LCNS-class** constellation, and a roving user offset from it by a baseline applies them so the **common-mode** broadcast-ephemeris errors cancel. Each satellite `i` carries an injected common 3-D orbit-error vector `e_i` and a common clock error `c_i`. The reference station's pseudorange residual is the correction `−e_i · û_ref,i + c_i`; the user's raw error is `−e_i · û_user,i + c_i`; the corrected user error is their difference `−e_i · (û_user,i − û_ref,i)`. The **clock term `c_i` cancels exactly** (it is identical in both observations), and the **orbit term collapses to the projection onto the difference of the two line-of-sight unit vectors**, which → 0 as the baseline → 0 (the spatial-decorrelation floor) and grows ≈ linearly with baseline. The per-satellite corrected range errors are mapped through the user geometry by a small dependency-light weighted-least-squares snapshot solve (`[x, y, z, clock]` state via `crate::orbit::invert4`; a constant range bias is absorbed by the clock unknown rather than corrupting the fix) to give the user **position** error. The user **protection level reuses the DO-229E SBAS protection-level machinery** (`crate::sbas::sbas_protection_level`, Precision-Approach K-factors) with the differential residual σ as each satellite's 1-σ error budget.
NovaMoon is referenced **only as a system class** (a public description of a lunar reference station); the constellation reuses the illustrative public-source `crate::lunar_service::LunarConstellation`. These are illustrative parameters for exercising the differential *method*, **not** a real ephemeris, and imply **no** affiliation, endorsement, heritage, certification or TRL. The `lunar-differential-pnt` scenario reports the user position error with vs without corrections, the reduction factor, the error-vs-baseline curve and the protection level. On the default configuration the corrected user error at a 50 km baseline is ≈ 0.008 m against ≈ 24 m uncorrected (a ~3000× reduction), and ≈ 0 at zero baseline.
The **oracle** is internal consistency: the common-mode cancellation is an **exact algebraic identity** (with `noise = 0`, corrected error < 1e-6 m at zero baseline; the clock term cancels to machine precision at any baseline), the residual **grows monotonically with baseline**, and the protection level is the **same DO-229E algorithm** the SBAS stack is pinned against. The spatial-decorrelation residual is a **first-order** geometric model, **not** a fitted decorrelation model from real lunar tracking.
| Capability | Status | Oracle / honest scope |
|------------|--------|-----------------------|
| Lunar differential PNT | `modelled` | `src/lunar_dpnt.rs` `differential_corrections`, `corrected_user_range_errors`, `user_position_error_m`, `lunar_dgnss_protection_level`, `LunarDpntScenario`. **Oracle: `InternalConsistency`** — the differential error-cancellation identity plus reuse of the DO-229E SBAS protection-level machinery (`crate::sbas`). Tests (`lunar_dpnt::tests`): at zero baseline the corrected user error and every corrected per-satellite range error are ~0 (< 1e-6 m) and ≪ the uncorrected error; the clock term cancels exactly at any baseline (orbit-error-zero case); the corrected error grows monotonically with baseline while staying ≪ uncorrected at modest baselines; differential beats standalone by a clear margin (> 2×); the protection level equals a direct `sbas::sbas_protection_level` call and scales with the residual σ; under-determined geometry returns `None`; the scenario is deterministic (seeded) and dispatches through `run_toml`. **Honest scope:** **MODELLED** — the common-mode cancellation is an exact identity but the spatial-decorrelation residual is a **first-order** geometric model, **NOT** a fitted decorrelation model from real lunar tracking, and this is **NOT** real-data validated. NovaMoon is referenced **only as a system class** (illustrative, public-source, **not affiliated with ESA**). No TRL, flight heritage or agency endorsement is claimed. |
## Lunar interoperability export — LunaNet/IOAG-aligned CCSDS OEM + KIF with round-trip conformance (`src/lunar_interop.rs`)
The lunar reference frame, lunar time scale and lunar ephemeris are exported in **LunaNet/IOAG-aligned, CCSDS-based interchange forms**. This phase does **not** invent a wire format: it **reuses** the crate's existing CCSDS OEM 2.0 emitter and parser (`crate::oem::OemFile::to_oem_string` / `crate::oem::parse_oem`) and the Kshana Interchange Format envelope (`crate::interchange::Envelope`), re-tagged for the lunar context. `export_lunar_oem` sets the OEM `REF_FRAME` to the **IAU 2015 WGCCRE lunar body frame** (`MOON_ME` mean-Earth or `MOON_PA` principal-axis), `TIME_SYSTEM` to the **lunar time scale** (`LTC` / `TCL` / `UTC`) and `CENTER_NAME` to `MOON`, over a sample **illustrative, public-source LCNS-class** ephemeris (positions from `crate::lunar_service::LunarSat::position_mci`, velocity by central finite difference). `export_lunar_time_metadata` emits a LunaNet/IOAG-aligned lunar-time descriptor (scale id, secular rate µs/day from `crate::lunar_time::lunar_rate_breakdown`, the published 56–59 µs/day band, reference surface) that round-trips through serde_json. `export_kif_lunar` wraps the OEM + descriptor + frame label in the existing KIF envelope with the MODELLED honesty label.
The **oracle** is round-trip / field conformance. Because `oem.rs` ships *both* directions, the OEM is round-tripped through `crate::oem::parse_oem` (back to an equal segment, with the lunar frame/time and state count preserved) **and** independently checked by `oem_conformance`, which verifies every required CCSDS/LunaNet header keyword is present, that `REF_FRAME` carries a lunar `MOON_*` token and `TIME_SYSTEM` is non-empty, that the `META_START … META_STOP` framing holds, and that the position+velocity data lines are well-formed — returning a structured pass plus the present/missing field list so a broken export (e.g. a dropped `TIME_SYSTEM`) is caught and named. The time metadata round-trips through `parse_lunar_time_metadata`, and the KIF envelope round-trips through `Envelope::parse` to equal artifacts carrying the MODELLED label.
| Capability | Status | Oracle / honest scope |
|------------|--------|-----------------------|
| Lunar interoperability export | `modelled` | `src/lunar_interop.rs` `export_lunar_oem`, `export_lunar_time_metadata`, `parse_lunar_time_metadata`, `export_kif_lunar`, `oem_conformance`, `LunarInteropScenario`. **Oracle: `InternalConsistency`** — deterministic round-trip (CCSDS OEM via `crate::oem::parse_oem`; KIF via `crate::interchange::Envelope::parse`; time metadata via serde_json) plus field-name conformance against published CCSDS OEM + LunaNet/IOAG field semantics. Tests (`lunar_interop::tests`): the exported OEM carries the lunar `REF_FRAME` (`MOON_ME`/`MOON_PA`) and `TIME_SYSTEM` (`LTC`/`TCL`) tokens, `CENTER_NAME = MOON`, the object and well-formed data lines; the lunar-time metadata round-trips (rate/band/reference) and rejects garbage; the KIF envelope serialises and deserialises to an equal value with the MODELLED honesty label present; a deliberately broken OEM (missing `TIME_SYSTEM`) fails `oem_conformance` and is reported missing; the scenario dispatches through `run_toml`. **Honest scope:** the field **names/units** are aligned with **public** standards (CCSDS 502.0-B OEM, IAU WGCCRE 2015 lunar frames, the LunaNet Interoperability Specification / IOAG lunar architecture); the export is **MODELLED** — deterministic round-trip + field-name conformance is the oracle, this is **NOT** a certified interoperability conformance test, the ephemeris is **illustrative, public-source, not affiliated with ESA**, and no TRL, flight heritage or agency endorsement is claimed. |
## Demonstration representativeness & gaps-to-flight ledger (`src/representativeness.rs`)
A simulation result is only useful as evidence if it states, on its face, what it is anchored to, what it assumes, and what still separates it from a flight system. The representativeness ledger makes that a first-class, machine-checked output: a `Representativeness` record carries the demonstration's external `Anchor`s, its modelled assumptions, its `Gap`s to flight (each with the phase/activity that would close it), and the TRL band it is representative for. It is the per-result companion to the capability-level `verification` matrix.
The **oracle** is a set of closed-form invariants, enforced by `representativeness::tests`: a `Validated` record must list at least one `ExternalDataset` anchor (a simulation cannot be validated against itself); a `Modelled` record must list at least one gap-to-flight and cannot claim a representative TRL band above 4 (maturation beyond that needs hardware/flight evidence, which is a gap, not a simulation output); the TRL band must be well-formed; and a `PartnerOwned` record must assert nothing. These invariants operationalise the "representativeness justified + remaining gaps towards flight clearly identified" discipline.
| Capability | Status | Oracle / honest scope |
|------------|--------|-----------------------|
| Representativeness & gaps-to-flight ledger | `modelled` | `src/representativeness.rs` `Representativeness`/`Anchor`/`Gap`, `check`/`is_valid`/`report`/`to_json`. **Oracle: `InternalConsistency`** — closed-form honesty invariants tied to the `verification` status/oracle-kind boundary. Tests (`representativeness::tests`): a Validated record without an external anchor fails; an internal (ReferenceImpl) anchor does not satisfy Validated; a Modelled record without a gap fails; a Modelled record above TRL 4 fails; malformed TRL bands are caught; the record serialises to JSON with the expected fields and renders a readable report. **Honest scope:** the ledger checks the *classification* of a result's evidence — it does not prove the named anchors resolve to live tests (those are curated, exactly like the `verification` matrix); the "machine-checked" claim is scoped to the status/anchor/gap/TRL invariants. |
## Unified quantum-vs-classical trade evidence (`src/qtrade.rs`)
Every quantum-PNT application area answers the same question the same way: fix one comparison frame (scenario + seed + engine version), route a quantum candidate and a classical baseline through one neutral code path, score them on common figures of merit, and report — with a confidence interval and an honest validated/modelled label — where quantum wins and where it does not. `TradeEvidence` gives that answer a single shape so the timing, navigation and anomaly-detection verticals all emit the **same** reproducible, honestly-labelled object. It does not re-implement any trade: per-FoM numbers come from the existing engines (`quantum_trade`, `crossover`, and the vertical modules); this is the contract plus the reproducibility/representativeness wrapper, built on the representativeness ledger and the verification labels.
The **oracle** is closed-form benefit/winner identities plus a faithful wrap of the existing trade output. `qtrade::tests`: the benefit ratio is polarity-correct (oriented so `>1` always means the quantum side is better, for both higher-is-better holdover and lower-is-better error FoMs); an evidence object built from a real `quantum_trade::TradeResult` reproduces its benefit ratios; dishonest evidence is rejected (a Modelled representativeness without a gap, or a `Validated` FoM whose trade record carries no external anchor); JSON is deterministic and carries the frame/seed/engine/FoM/representativeness fields.
| Capability | Status | Oracle / honest scope |
|------------|--------|-----------------------|
| Quantum-vs-classical trade evidence | `modelled` | `src/qtrade.rs` `TradeEvidence`/`TradeFom`/`TradeFrame`/`Winner`. **Oracle: `InternalConsistency`** — closed-form benefit/winner identities + faithful wrap of `quantum_trade::TradeResult`; honesty tied to the representativeness ledger and the verification status/oracle-kind boundary. Tests (`qtrade::tests`) as above. **Honest scope:** the harness standardises *how* a trade is reported and reproduced; the underlying per-FoM physics is whatever the contributing engine provides, carried with its own validated/modelled label — `TradeEvidence` does not upgrade a modelled FoM to validated. |
## Quantum device error-model library (`src/quantum_devices.rs`)
A single place that exposes the quantum-PNT devices the demonstrator trades — optical / trapped-ion / mercury-ion clocks, classical reference clocks, the cold-atom interferometer, and time-transfer links — each as a `DeviceCard` carrying its headline spec and a representativeness record. Most device physics is **reused** (clock stabilities from `holdover::QuantumClockClass` and `clock_state::ClockClass`; the cold-atom interferometer from `quantum_trade`/`crossover`; classical optical/RF links from `timetransfer_adv`). The one genuinely new model is the **entanglement / single-photon time-transfer link**, whose timing precision is shot-limited: `σ_t ≈ jitter/√(R·τ)` in the detected coincidence rate `R` and integration `τ`, degraded by a dark-count penalty `√(1 + dark/R)` and floored by an irreducible systematic.
The **oracle** is the reused, published clock/CAI coefficients plus closed-form shot-noise/loss identities. `quantum_devices::tests`: the reused clock-stability ordering is sane (optical-lattice < mercury-ion < CSAC in σ_y(1 s)); the entanglement precision improves as `1/√τ` (4× integration halves σ) when shot-limited; the detected coincidence rate falls 10× per 10 dB of link loss; dark counts monotonically degrade precision; the systematic floor bounds precision below; every device card is a valid, Modelled representativeness record.
| Capability | Status | Oracle / honest scope |
|------------|--------|-----------------------|
| Quantum device error-model library | `modelled` | `src/quantum_devices.rs` `DeviceCard`, `quantum_clock_card`/`classical_clock_card`, `EntanglementTimeLink`. **Oracle: `InternalConsistency`** — reused published clock/CAI coefficients + closed-form shot-noise/loss identities for the entanglement link. Tests (`quantum_devices::tests`) as above. **Honest scope:** clock/CAI device parameters are reused from the engine's existing published-coefficient models; the entanglement-link model is MODELLED from published quantum-clock-synchronisation behaviour with illustrative public-source pair-rate/efficiency/loss values — it is **not** validated against a measured link, and real source/channel/detector hardware is a stated gap-to-flight. |
## Trusted quantum timing — end-to-end chain, secure dissemination, anomaly + trade (`src/timetransfer_chain.rs`)
The `quantum-time-transfer` scenario composes the timing chain a quantum-PNT demonstrator needs: an end-to-end user-time budget (reference-clock coast error and link timing precision combined in quadrature) for a **quantum** chain (optical-lattice clock + entanglement/single-photon link) versus a **classical** chain (CSAC + RF two-way link); a **timing protection level** (reused from `tpl`) bounding the undetected time error; a **delay/replay-attack security figure of merit** `1 − P_md` at a stated false-alarm rate (reused from `detection`); a **clock-anomaly detection probability + CUSUM change-detection latency**; and the quantum-vs-classical comparison as honest `qtrade::TradeEvidence` carrying a representativeness + gaps-to-flight record.
The **oracle** is a closed-form quadrature budget over reused, separately-validated kernels (ADEV vs Stable32/NIST; the TPL bound; `detection::analytic_pd`/`analytic_pmd`). `timetransfer_chain::tests`: quantum precision improves with integration time; the quantum chain can win *and* can lose the precision FoM (honest — not a universal win; a very lossy link flips the result); the protection level is finite-positive; the security FoM is in [0,1] and grows with attack delay; the anomaly detection probability is monotone in fault magnitude; and the trade evidence is internally honest (`is_honest()`).
| Capability | Status | Oracle / honest scope |
|------------|--------|-----------------------|
| Trusted quantum timing (chain + secure dissemination + anomaly + trade) | `modelled` | `src/timetransfer_chain.rs` `QuantumTimeTransferScenario`/`Report`, `to_svg`; `kind="quantum-time-transfer"`. **Oracle: `InternalConsistency`** — closed-form quadrature budget over reused validated kernels (ADEV/Stable32-NIST, TPL bound, detection analytic Pd/Pmd); honesty tied to the representativeness ledger. Tests (`timetransfer_chain::tests`) as above. **Honest scope:** clock/link parameters are illustrative, public-source; the entanglement link is the MODELLED shot-limited model from `quantum_devices`; this models the *class* of trusted-timing system — it is **not** validated against a measured end-to-end link, and real clock/link/space-channel hardware is a stated gap-to-flight; no TRL, flight heritage or certification is claimed. |
## GNSS-free quantum navigation (`src/quantum_nav_od.rs`)
The `quantum-gnss-free-nav` scenario compares dead-reckoning during a GNSS outage with a **quantum** inertial budget (a cold-atom interferometer accelerometer, reused from `inertial::quantum_imu`/`quantum_trade`, whose CAI white-noise floor is validated to within ~2 orders of magnitude of Freier et al. 2016) against a **classical** navigation-grade INS. It reports the position-error growth over the coast, the holdover time to a position threshold, and the quantum-vs-classical trade as honest `qtrade::TradeEvidence`. Observability is stated honestly: with no external fix during the outage the accelerometer bias is unobservable, so the position error grows without bound (≈ ½·bias·t² plus a velocity-random-walk term); a lower-bias quantum sensor *slows* that growth but does not close the gap — only an external fix does.
The **oracle** is the reused inertial budgets (`QuantumNavBudget`/`ClassicalInsBudget`, both implementing the `PositionDrift` trait) plus closed-form drift identities. `quantum_nav_od::tests`: the quantum budget beats the classical over a long outage (lower error, longer holdover); the error grows with outage duration for both (the observability gap); the advantage is outage-dependent, not a constant win factor (the quantum curve is noise-dominated at short coast, bias-dominated at long); the trade evidence is internally honest.
| Capability | Status | Oracle / honest scope |
|------------|--------|-----------------------|
| GNSS-free quantum navigation | `modelled` | `src/quantum_nav_od.rs` `QuantumNavOdScenario`/`Report`, `to_svg`; `kind="quantum-gnss-free-nav"`. **Oracle: `InternalConsistency`** — reused inertial budgets (CAI validated-by-reuse ~2 orders vs Freier 2016; classical closed-form drift); honesty tied to the representativeness ledger. Tests (`quantum_nav_od::tests`) as above. **Honest scope:** a static dead-reckoning budget comparison, not a dynamic filter on real trajectories; bias unobservable without a fix is stated; device parameters are illustrative, public-source; models the *class* of GNSS-free quantum nav — real cold-atom IMU hardware, a dynamic platform and the flight environment are stated gaps-to-flight; no TRL/flight/certification claimed. |
## Fault / anomaly detection for quantum PNT systems (`src/quantum_faults.rs`)
The `quantum-anomaly-detect` scenario defines a labelled catalog of quantum-PNT faults (clock frequency-jump, drift, lock-loss; sensor bias-step, dropout), scores a detection statistic on nominal vs faulted windows, and reports the detector's discrimination as an ROC AUC and a minimum-detectable fault at a fixed false-alarm rate. The quantum-vs-classical angle: a more stable quantum-clock reference lowers the monitor noise σ, so it detects smaller faults — emitted as honest `qtrade::TradeEvidence` with a representativeness record.
The **oracle** is a closed-form AUC for the Gaussian detection statistic (nominal `N(0,σ)`, fault `N(μ,σ)` ⇒ `AUC = Φ(μ/(σ√2))`) cross-checked against the externally-validated `eval_stats::bootstrap_auc_ci` (validated vs scikit-learn) and `detection`'s analytic thresholds. `quantum_faults::tests`: the analytic AUC matches known values (0.5 at μ=0, →1 for large μ/σ, monotone); the empirical bootstrap-AUC CI brackets the closed form; the quantum monitor has higher AUC and a smaller minimum-detectable fault; the advantage vanishes for huge faults (both detect perfectly — not a universal margin); the catalog has five labelled classes; the trade is internally honest.
| Capability | Status | Oracle / honest scope |
|------------|--------|-----------------------|
| Fault/anomaly detection for quantum PNT | `modelled` | `src/quantum_faults.rs` `FaultKind`, `QuantumAnomalyScenario`/`Report`, `analytic_auc`, `min_detectable_fault`; `kind="quantum-anomaly-detect"`. **Oracle: `InternalConsistency`** — closed-form Gaussian AUC cross-checked against the externally-validated `eval_stats` bootstrap AUC (vs scikit-learn) and `detection` analytic thresholds. Tests (`quantum_faults::tests`) as above. **Honest scope:** the detection statistic is a Gaussian model, not real telemetry; the fault catalog is labelled-synthetic; quantum/classical monitor noise levels are illustrative, public-source; models the *class* of quantum-PNT anomaly detection — real hardware telemetry with ground-truth labels is a stated gap-to-flight; no TRL/flight/certification claimed. |
## Standards conformance — interchange formats, link budget & integrity (external-oracle checks)
These standards are validated by recovering the **verbatim worked examples / reference values published in the standards themselves** (or in an independent reference implementation), not by Kshana's own round-trip. Recovering a standard's own annex example is the strongest available conformance check for an interchange-format parser.
| Standard | Status | External oracle |
|----------|--------|-----------------|
| **CCSDS OEM** (502.0-B Orbit Data Messages) | `validated` | `tests/ccsds_reference.rs`: `parse_oem` recovers the exact epochs and X/Y/Z/Ẋ/Ẏ/Ż state vectors of the KVN OEM example in **CCSDS 502.0-B-3, Annex G §G6, Figure G-11** ("OEM Example with No Acceleration, No Covariance", p. G-10) — both ephemeris segments, including the leading-zero source values (`-063.042`). Fixture `tests/fixtures/ccsds/oem_502_b3_figG11.oem` is reproduced character-for-character from the official Blue Book (the editorial "< intervening data records omitted here >" markers, which are not valid OEM records, are not included). |
| **CCSDS TDM** (503.0-B Tracking Data Message) | `validated` | `tests/ccsds_reference.rs`: `TdmFile::parse` recovers all 41 RANGE records (first 3198.03679519614 km, last 3270.46440460551 km) and the metadata of the KVN TDM example in **CCSDS 503.0-B-2, Annex E, Figure E-9** ("Range Data with TIMETAG_REF=TRANSMIT", p. E-9). Fixture `tests/fixtures/ccsds/tdm_503_b2_figE9.tdm` is verbatim from the Blue Book. (This is the published-Blue-Book oracle alongside the pre-existing hand-authored `reference.tdm` round-trip in `src/ccsds_tdm.rs`.) |
| **CCSDS 133.0** (Space Packet Protocol) | `validated` | `src/space_packet.rs` `primary_header_matches_published_independent_test_vectors`: the encoder reproduces the byte-level primary-header test vectors of the independent **`spacepackets` library** (us-irs/spacepackets-py, `tests/ccsds/test_space_packet.py`) — e.g. `18 02 40 34 00 16` and the all-max `1F FF FF FF FF FF` — plus the canonical all-ones TM case, each round-tripped through decode. Bit-field widths/order per **CCSDS 133.0-B-2 §4.1.3**. |
| **CCSDS 401 / DSN** (link budget, DSN 810-005) | `validated` | `src/linkbudget.rs` `link_equation_reproduces_descanso_galileo_dct`: the FSPL and C/N₀ chain reproduce the published deep-space telemetry design-control table of **J. H. Yuen (ed.), *Deep Space Telecommunications Systems Engineering* (DESCANSO/JPL Pub. 82-76), Table 1-1** — the Galileo X-band (8420.43 MHz) downlink at 6.37 AU: free-space loss **290.54 dB** (<0.05 dB) and received **Pr/N₀ = 54.6 dB-Hz** (<0.2 dB). Band centres sit inside the DSN/CCSDS-401 deep-space downlink allocations (DSN 810-005 Module 201). The *default mission params* remain representative (not a calibrated transponder); the link *equation* is what this pins. |
**GNSS standards — external-oracle status and the honest remaining gaps:**
| Standard | Status & honest caveat |
|----------|------------------------|
| **RAIM / ARAIM** | **Kernel externally validated; PL value not pinned.** The detection *kernel* — the χ² fault-detection threshold, the non-central-χ² missed-detection non-centrality, and the normal-law K-multipliers — now reproduces **SciPy** (`tests/raim_reference.rs`, 171 cases, ≤ 1e-6 rel), and the wrapping geometry reproduces **gnss_lib_py** (DOP). What stays **unpinned** is the protection-level *value*: the one published ARAIM oracle — WG-C ARAIM Reference ADD v3.1 "LPV-200 numerical example" (VPL 18.3 m / HPL 13.45 m) — requires the **multi-constellation, accuracy-weighted, multi-fault-mode** machinery Kshana does not implement (its `GᵀG` is single-clock, equal-weight), so neither the published PL nor its σ_v,acc = 1.3694 m intermediate is reproducible; classic RAIM has **no canonical published input→HPL test vector** at all. So: detection kernel validated, geometry validated, but the PL *number* and the ARAIM MHSS budget allocation remain formula-correct, not pinned to a published value. |
| **SBAS / DO-229E** | **Protection-level algorithm now externally validated.** Given the per-satellite (el, az, σ), kshana's `sbas_protection_level` reproduces the **RTKLIB SBAS-PL fork** (`zsiki/rtklib_ws` `waasprotlevels()`, Siki & Takács 2017, "DO-229D Appendix J") — fed that tool's **own** σ computed from **real EGNOS** broadcast messages + real BUTE RINEX — to **< 2e-3 m HPL** across 6 epochs (`tests/sbas_reference.rs`); ESA gLAB v6.0.0 (`core/filter.c`) confirmed the identical convention. K-factors (`K_H,PA=6.0`, `K_V=5.33`, `K_H,NPA≈6.18`) match DO-229E Appendix J. **What is validated vs what is not:** this pins the PL *algorithm* (σ in → HPL/VPL out); the upstream σ-*modelling* from raw UDRE/GIVE messages is out of kshana's scope — that is where independent *full-pipeline* tools still disagree by metres on real data. K_V matched at the oracle's rounded 5.33 (kshana's exact `Φ⁻¹(1−5e-8)=5.3267` → VPL ~0.06 % smaller; the vertical is validated K-factor-free as `d_U`). |
| **IONEX / Klobuchar** | **Broadcast model externally validated.** The Klobuchar L1 model reproduces **RTKLIB's `ionmodel`** (`tests/klobuchar_reference.rs`, < 1e-4 m) — *implementation parity* with the de-facto open reference, since no official *measured-truth* worked example exists. The remaining gap is the **IONEX TEC-map reader** for real 80-column GIM files (a tracked follow-up), not the broadcast model. See the Klobuchar rows in [Claims vs. reality](#claims-vs-reality--quick-reference). |
| **GPS L1 C/A spreading code (IS-GPS-200)** | **Code generation externally validated.** `src/sdr.rs` (`ca_first_ten_chips_match_is_gps_200_octal`): the G1/G2 LFSR C/A-code generator reproduces the **published code-phase assignments in IS-GPS-200 Table 3-Ia** — the first-ten-chips octal for PRN 1–9 ({1440, 1620, 1710, 1744, 1133, 1455, 1131, 1454, 1626}) — exactly. This pins the **spreading-code chip sequence** against the GPS interface specification; the surrounding modulation, spectral-separation and DLL code-tracking pieces (`navsignal`) stay MODELLED. |
## Operating envelope
Each pack is exercised across its stated input envelope by
`tests/scenario_coverage.rs`, which asserts every numeric output is finite (no
NaN/Inf) and bounded. The table lists the tested input range per pack, the
expected output behaviour, and the covering test.
| Pack | Input swept | Tested range | Expected output | Covering test |
|------|-------------|--------------|-----------------|---------------|
| clock | `threshold_ns` (timing spec) | 1 – 500 ns | finite holdover/timing/security FoMs | `clock_pack_covers_the_spec_threshold_envelope` |
| inertial | `accel.bias` | 1e-7 – 1e-2 m/s² (cold-atom → crude MEMS) | finite, bounded position RMS/p95 | `inertial_pack_covers_the_accel_bias_envelope` |
| orbit | `mask_deg` (elevation mask) | 5° – 30° | finite DOP/availability; bounded | `orbit_pack_covers_the_elevation_mask_envelope` |
| spoof | `attack.rate_ns_per_s` | 0.1 – 50 ns/s | finite P_md / security, bounded | `spoof_pack_covers_the_attack_rate_envelope` |
| hybrid | `position_spec_m` | 10 – 1000 m | finite timing/position holdover | `hybrid_pack_covers_the_position_spec_envelope` |
| orbit (real) | real Celestrak `gps-ops` TLEs | 30-satellite snapshot, checksum-strict | loads only with valid checksums; bounded geometry | `real_gps_constellation_scenario_loads_with_valid_checksums_and_bounded_output` |
| clock (flicker) | `flicker_floor` on/off | 0 vs 1e-12 | enabling the 1/f floor **worsens** the timing-p95 coast | `flicker_fm_floor_degrades_the_clock_holdover_when_enabled` |
| fusion (realism) | `accel.bias` 0 vs 5.88e-7 m/s² | zero vs realistic non-zero | filter still converges, within 3× the zero-bias error | `fusion_filter_converges_with_a_realistic_non_zero_bias` |
The flicker-FM and fusion-bias rows close two specific realism gaps: the noise
terms are **off by default but demonstrably affect output when enabled**, and the
joint fusion filter **does not depend on biases being zeroed** — it converges with
a realistic cold-atom-grade residual bias too.
## Known limitations
- Quantum and classical runs now use independent RNG seeds (classical seed = seed + 0x9e3779b97f4a7c15) so their noise realizations are uncorrelated — fixed after review.
- `holdover_s` is segment-aware: outage timelines are split into contiguous segments at GNSS re-acquisition and the reported value is the worst-case (shortest) coast across them. It remains bounded by the time-grid resolution (a lower bound).
- ISL time-transfer re-sync models the residual link uncertainty as fresh zero-mean jitter per measurement step plus re-anchoring at the configured interval.
## Claims vs. reality — quick reference
A hostile reviewer's checklist. For each term that could be read as more than it is:
what Kshana does today, and where the real version sits on the roadmap.
| Term you may see | What it actually is today | What it is **not** (yet) |
|------------------|---------------------------|--------------------------|
| "hybrid quantum/classical PNT simulator" | a classical stochastic simulator driven by published quantum-sensor Allan/noise coefficients | first-principles quantum physics (Mach–Zehnder phase, projection noise, systematics) — see [`QUANTUM-MODELS.md`](QUANTUM-MODELS.md) |
| "joint Kalman fusion" / fusion pack | the runnable `fusion` pack observes the clock and position **separately** (a direct time fix and a direct position fix), for which the optimal estimator is genuinely block-diagonal — two two-state filters with a combined FoM | a *coupled* filter for the **pseudorange** case now **exists** as `fusion::coupled::CoupledPntFilter` (4-state stacked `[pos,vel,phase,freq]`, non-zero cross-block covariance, NEES-validated) but is **not yet wired into the runnable pack**, and the pack is **1-DOF** — the 3-D 8-state extension is future work |
| Security FoM (`spoof` kind) | `1 − P_md` of a **stochastic time-spoof detector**: a two-sided χ²₁ / Neyman–Pearson test on a clock-aided monitor statistic, threshold set from a target `P_fa`, `P_md` evaluated **at the spec magnitude** both closed-form and by Monte-Carlo (which agree to a few ×1/√N) | a multi-satellite RAIM/ARAIM detector — see [`INTEGRITY.md`](INTEGRITY.md); the monitor statistic is a Gaussian model with the clock's PSD-derived σ, not a signal-level correlation |
| "clock-aided spoof-detection" | a **single-clock** time-discrepancy monitor (Gaussian statistic, σ from the clock PSD), NOT multi-SV RAIM | range-domain multi-SV RAIM with protection levels |
| spoof-detector P_md (closed-form vs Monte-Carlo) | the analytic `Φ`-based `P_md` and the empirical `mc_runs`-trial estimate match within sampling error (≈ 1/√N); validated in `src/spoof.rs` / `src/detection.rs` tests at a stressed deflection (μ/σ = 2) | a measured detector ROC against recorded spoofing data |
| `gnss-sim` pseudorange model | a **forward simulation**: `ρ = geometric + clocks + Klobuchar iono + Saastamoinen·Niell tropo + noise + multipath`. Each atmosphere model is unit-tested against hand values (ZHD ≈ 2.3 m, Niell = 1 at zenith, Klobuchar obliquity), and the full Klobuchar broadcast model is **cross-checked against RTKLIB's `ionmodel()`** reference implementation to sub-millimetre (`klobuchar_agrees_with_rtklib_reference_implementation`, 6.1278 m for the RTKLIB default-coefficient case — no official published Klobuchar test vector exists, so this is implementation parity, not measured truth); a **zero-noise run reproduces geometry + corrections to sub-millimetre** (RAIM residual ≈ 0) | a position **solution** from real signals (PPP/RTK), dual-frequency iono-free combination, or carrier-phase ambiguity resolution — not modelled |
| jamming model (`jamming` kind) | a **link-budget** chain: J/S from free-space path loss + per-direction antenna gain → effective C/N₀ via the anti-jam equation (despreading gain × spectral-separation `Q`) → loss of lock at a threshold (Kaplan & Hegarty §9.4) | multipath, terrain shadowing of the jammer, near/far AGC, adaptive nulling, acquisition-vs-tracking hysteresis; `Q` is a representative per-type constant, not derived from the jammer's measured PSD |
| Integrity FoM | filter **self-consistency** (fraction of outage samples inside its own k·σ bound) | HPL/VPL, integrity risk / P_HMI, alert limits, DO-229E/316/ED-259A |
| legacy `inertial` scenario **pack** FoM | a **single-axis (1-DOF)** accelerometer/gyro error budget (VRW/ARW, bias-instability) with a truth-snap reset | the *legacy* pack is 1-DOF; the **3-axis path ships in the `gnss-ins` pack** (next row) |
| 3-axis strapdown **library** + `gnss-ins` **pack** (`src/inertial/`, `src/fusion/pack.rs`) | a verified quaternion/NED navigator with a deterministic IMU error model — **scale-factor, misalignment, g-sensitivity, quantization, rate-ramp** modelled (IEEE Std 952-1997 §A.2; Groves 2013 §4.3) — now **driven by the `gnss-ins` scenario pack** and **configurable per sensor from TOML** (`[imu_*.error_model]`) | **not modelled:** vibration rectification error, temperature-gradient drift. The fused-coast error is floor-limited by hand-over attitude error (tilt/accel-bias weakly separable), so per-bias calibration is not claimed |
| "Hybrid PNT integration" | open-loop dead-reckoning bracketed by truth-snap GNSS resets | a coupled (loose/tight) GNSS–INS Kalman blend |
| "Positioning Performance" (clock packs) | clock-phase **timing** RMS in ns (`timing_rms_ns`) | a position-domain RMS/CEP/SEP/DOP-weighted accuracy |
| inertial position FoM (`pos_rms_m`) | a **single-seed, single-axis** position RMS/p95 in metres | an ensemble CEP / 2DRMS distribution (roadmap: Monte-Carlo CEP) |
| "validated to ~2%" | a *typical* observed agreement; the **enforced gate is 20–25%** relative error | a guaranteed 2% accuracy bound |
| "reproducible / bit-identical" | bit-identical re-run on the **same OS + pinned toolchain** | a committed cross-platform golden-hash check (roadmap: reproducibility milestone) |
| SGP4 GPS scenario (`orbit-sgp4-gps.toml`) | synthetic Walker TLEs (placeholder NORAD IDs) for geometry demonstration | the real `gps-ops` constellation — drop in a Celestrak snapshot (see [`REAL_TLE_GUIDE.md`](REAL_TLE_GUIDE.md)) |