# LOTUS plugin versus NASA/JPL flight trajectories

Validation study date: **15 November 2026**. Flight-reference research and the preserved Horizons release remain dated **10 November 2026**. This study replaces the earlier private-helper/coast demonstrations on the examples page.

## What is being compared

The gold curve is computed by the actual local LOTUS plugin HTTP API: `POST /orbital/analyze`, followed by the job endpoint and native `.lotus_analysis` export. The requests, raw jobs, progress logs, two independent run exports and independent validation reports are retained. This pipeline is the product's native deterministic mission planner, not a plot fitted to JPL spacecraft positions.

The blue curve is the preserved geometric spacecraft ephemeris distributed by **NASA/JPL Horizons**. BepiColombo's navigation trajectory is supplied by ESA; it is not an independent JPL orbit determination of that ESA mission. Psyche's trajectory is supplied by JPL navigation. Horizons states and project kernels are reference products, not publicly established covariance or exact truth. [JPL Horizons manual](https://ssd.jpl.nasa.gov/horizons/manual.html), [Horizons API conventions](https://ssd-api.jpl.nasa.gov/doc/horizons.html).

The benchmark is a **fixed historical planet encounter timing design comparison with withheld spacecraft state history, not a forecast**. Publicly researched event dates and planet ephemerides are mission design inputs. No spacecraft initial, interior or terminal state enters the native solve. No translation, rotation, scaling, trajectory warping or endpoint snap is applied afterward. The actual native cost ranking is preserved.

Both published primary candidates **fail physical screening under the declared nominal hardware/Isp scenario**. The nominal dry mass floors are explicit assumptions, not empirically established flight-mass lower confidence bounds. These candidates must not be advertised as a successful reproduction of the flown missions. Solver determinism, numerical consistency, physical feasibility and flight-path agreement are separate questions.

## Cases and selection rule

| Mission | Designed phase | Fixed epochs, UTC | Historical prograde winding |
|---|---|---|---|
| BepiColombo | Earth return → first Venus encounter | 10 Apr 2020 04:24:57.237 → 15 Oct 2020 03:58:31.604 | No complete revolution |
| Psyche | Earth launch → Mars encounter | 13 Oct 2023 14:19:43 → 15 May 2026 19:28:02.786 | No complete revolution |

These are complete designed phases, with no seven-day trimming to remove encounter effects. They do not represent complete missions or resolve the close flybys. Encounter dates come from independently checked direct planet-centered Horizons queries; launch time and spacecraft separation time remain distinct. All encounters and sources are in the [mission reference report](../reference/mission-reference-report.md).

Primary selection is the lowest native rank for a **low-thrust solve with the historical encounter epochs and independently established trajectory winding**. Unwrapped JPL ecliptic longitude establishes BepiColombo's approximately 4.63477 rad sweep and Psyche's approximately 6.07579 rad sweep, each less than one full revolution. This is an explicitly disclosed mission topology constraint, not a search for the smallest spacecraft position residual.

The initial Psyche unconstrained grid allowed `Nrev=[0,1,2]` and selected a cheaper one-revolution shape. It differs from the historical topology. That baseline is preserved; the primary request fixes `Nrev=[0]` before its residuals are evaluated. Ballistic approximations and failed multi-encounter chains are retained as additional evidence. They do not silently replace the requested low-thrust benchmark because they happen to have a smaller position error.

The full BepiColombo launch → Earth → Venus → Venus → Mercury chain was also submitted with exact event times and with ±1-day event windows. No accepted full chain was returned. A no-candidate result has no predicted curve and cannot inherit an imported flight-reference path.

## Determinism

Every case is submitted twice through the actual API, with native low-thrust and graph execution fixed to one thread, parallel search disabled, and persistent stage caching disabled. The entire scientific result payload is serialized canonically and hashed. All constraints, options, encounter states, leg states and controls, flyby records, trajectory costs, candidate order and categories must match exactly.

Only the top-level `diagnostics` object is explicitly excluded from this hash because it contains runtime, worker and profiling observations. This exclusion is recorded in both reports. No state, control, cost or constraint is removed from the scientific payload. The raw files also retain their own distinct SHA256 hashes, since timings and export timestamps can differ.

A deterministic repeat requires exact scientific hash equality, not merely rounded RMS agreement. Raw sampled position arrays are compared separately. A tampered-state regression test proves the hash check rejects changed trajectories; the same test preserves constraint changes and ignores only the declared diagnostic timing exclusion.

## Frames, epochs and reference coverage

The actual requests use **SPICE J2000, observer SUN, aberration correction NONE**, i.e. Sun-centered ICRF equatorial vectors. This matches the packed Horizons reference frame directly; no ecliptic rotation is applied to these actual outputs. All model/reference pairing uses continuous **TDB seconds since JD2451545.0**. UTC labels are for people and are converted using the stored time offsets. Default TDB epochs are never relabelled as UTC.

The spacecraft reference is evaluated independently using cubic Hermite interpolation of position and velocity, with no extrapolation. Source gaps are excluded. Model positions at comparison epochs are evaluated with cubic Hermite interpolation of the raw native profile's `r` and `v`. Raw `u` values have declared units km/s², regardless of magnitude; a small magnitude never triggers an AU/day² unit guess.

BepiColombo's source fit cutoff is 22 October 2026; Psyche's is 27 October 2026. Both selected primary phases precede their cutoffs and are labelled reconstructed flight references. Future provider predictions are not called flown results. Source names, headers, hashes and documented discontinuities remain attached to the reference report.

The packed Psyche reference begins at **13 October 2023 17:00 UTC**, after the native Earth launch endpoint at 14:19:43. The unsupported first comparison epoch is excluded. The full native solve time remains displayed; the report does not pretend its missing launch reference sample was measured. Source gaps that fall between plotted samples are recorded as path breaks rather than silently connected.

The preserved reference's approximately 100 m largest tested interpolation errors describe the reference packaging. They do **not** describe LOTUS prediction accuracy, JPL flight covariance or the model's much larger position separation. [Reference audit and event checks](../reference/import-independent-review.json).

## Error metrics

Each raw native leg is compared at 4001 uniformly spaced TDB epochs; unsupported/gap epochs are excluded and counted. Position separation is the Euclidean **3D** distance in kilometres. RMS is `sqrt(mean(separation²))` over all published valid paired epochs, including valid endpoints. Maximum, arrival separation, optional interior-only RMS and velocity separation are retained separately. No plotting projection enters any metric.

Planet-center design endpoints are not the flown spacecraft's closest-approach states. For example, endpoint separation naturally includes approximately 19064 km at BepiColombo's Earth encounter, 16774 km at Venus and 8000 km at Psyche's Mars encounter. These values are reported without snapping. Small endpoint residual relative to a prescribed planet position is a boundary condition, not spacecraft forecast accuracy.

The optional normalized RMS divided by mean heliocentric radius is an orbit-scale ratio, not a percent of mission accuracy. Absolute kilometre errors dominate the public comparison.

## Independent numerical and physical checks

Native job completion alone is never a pass. The independent validator imports **no LOTUS shape or propagation helper**. It consumes the unmodified exported arrays and applies:

- Finite, aligned, strictly increasing profile epochs; explicit frame, center and unit checks.
- Time-of-flight closure to **1 ms**, endpoint model position closure to **1 km** and velocity closure to **0.0001 km/s**. These are numerical tolerances, not flight tolerances.
- Independent trapezoid integration of acceleration magnitude, checked against reported integrated control delta-v.
- **DOP853 initial-value replay** of solar gravity plus a cubic-spline interpolant of the raw returned inertial control. It starts only from the first raw model state; no terminal or spacecraft reference state enters the replay. Maximum replay separation must be at most **10 km**. Numerical integration uses `rtol=2e-12`, `atol=1e-11` and a maximum step of phase duration/2000. The sampling resolution and measured residual are retained. A resolution failure remains a failure until a finer actual plugin run passes; the tolerance is not raised to hide it.
- A separate derivative residual for `dv/dt = -mu*r/|r|³ + u`, reported as a diagnostic rather than a navigation error.
- Solar clearance at native profile nodes.
- Force and ideal propellant bounds, including variable mass, assumed hardware mass floor and stated propulsion limits.
- Where raw flyby records exist, independent asymptotic speed/turn/periapsis constraints. Unresolved close approach, full power, throttle, duty, attitude and reserve constraints remain **not evaluated**.

`solverConverged` in the downloadable evidence means every selected leg passes the stated numerical closure and independent replay tests. It does not claim a global optimum or physical mission feasibility. The page calls this numerical solve consistency.

For constant assumed Isp, `m(t)=m0*exp(-integral(|u|dt)*1000/(g0*Isp))` and `T(t)=1000*m(t)*|u(t)|`. The thrust screen also records `1000*mDry*max(|u|)` and the constant launch-mass upper estimate. A necessary ideal propellant estimate is `mDry*(exp(deltaV*1000/(g0*Isp))-1)`. These are conditional screening bounds under declared mass/Isp assumptions, not reconstructed telemetry.

BepiColombo uses nominal launch mass 4043.2 kg, a 3600 s illustrative effective Isp, 581 kg xenon and the published 0.25 N two-engine envelope. The independent hardware-mass screen uses the research report's nominal dry hardware sum of **2632.3 kg**, rather than treating all non-xenon launch mass as dry hardware (the stack also carries chemical propellant). Psyche uses launch mass bound 2747 kg, assumed 1800 s Isp, 1085 kg xenon and **0.24 N** from the cited press kit. The report's 1616 kg Psyche nominal dry specification is an explicit screening assumption; the press kit states “up to,” so it is not a measured lower confidence bound. The reports retain assumptions and do not silently promote these values to flight mass histories. [BepiColombo research sources](../reference/mission-reference-report.md), [JPL Psyche spacecraft specifications](https://www.jpl.nasa.gov/press-kits/psyche/mission/spacecraft/).

The native planner currently does not enforce hard SEP thrust, power or propellant feasibility in these mission solves. Passing a mass model to the plugin does not establish that those bounds constrained the generated shape. Independent physical failures therefore remain visible beside the plot.

## Reproduction and downloads

The repository's `scripts/mission-examples/plugin-validation/` contains the actual API runner, topology audit, independent validator, validator regression tests and publication script. `audit/2026-11-15/plugin-validation/` preserves all input requests, raw API jobs, two native exports per case, independent reports and the solver capability audit.

The public page offers the exact request, full independent report, and a `.lotus_analysis.gz` download. Decompress the latter before importing into LOTUS. The gzip has a fixed zero timestamp and no host filename; its decompressed bytes match the untouched native export SHA256. The public plot data records raw analysis/request/reference/report hashes, chosen trajectory rank and leg IDs, interpolation and selection rules. Failed cases remain inspectable.
