JPL data and Swiss Ephemeris comparisons
Cosmic Ephemeris calculates planetary positions using JPL DE440 and DE441 data through Skyfield. Selected calculations are checked against recorded Swiss Ephemeris/Moshier reference results.
The comparisons below identify the tested samples, coordinate conventions and measured differences. Planetary data, house systems and sidereal conventions each contribute to the chart your application receives.
Retained fixture-backed comparisons
- Tropical planets
- 70 comparisons across 7 epochs; maximum observed difference 1.2434436 arcseconds, below 1.4
- Chart angles
- 81 cases and 162 Ascendant/MC comparisons; maxima 8.063712 and 2.0986956 arcseconds respectively, below 9 and 2.2
- Sidereal planets
- 210 declared-frame comparisons across 7 epochs and three modes; maximum observed difference 1.3214296 arcseconds, below 1.4
The comparisons use one committed independent Swiss Ephemeris/Moshier fixture as offline reference data. Planet samples cover 1900–2050; the angle cases combine isolated geometry with dated locations. Sidereal values are aligned to Cosmic Ephemeris's declared retained-nutation frame; the larger raw difference from a conventional sidereal frame is a convention difference, not declared-frame error. Swiss software and data are not runtime dependencies. These figures describe the enumerated samples, not a universal error bound or a ceiling for ancient dates.
Historical audit: 1000 BCE–2500 CE
The October 15, 2026 audit completed 1,071 standard engine tests and all five optional timeline tests: 1,076 distinct passes. Included were 924 endpoint scenarios across all 77 date-bearing operations, both calendars and three offsets, with expected invalid-window rejections. The 78th operation, place search, has no date. Successful boundary handling is separate from numerical accuracy.
- Planetary longitudes
- 3,440 comparisons at 344 sampled epochs; maximum observed difference 0.446514 arcseconds
- Fixed stars
- 456 cases; maxima 1.122700 arcseconds longitude and 1.488251 latitude
- Ancient chart angles: before correction
- 2,064 comparisons per angle; R281 Ascendant maximum 47.185418 arcseconds, Midheaven maximum 21.061749 arcseconds
- R282 angle correction
- Same grid: Ascendant maximum 1.490923 arcseconds; Midheaven 0.901749; Equal-house cusps 1.630710. These sampled differences are not universal accuracy ceilings.
Astropy 7.1.0 / ERFA 2.0.1.5 and Skyfield 1.53 used matching TT instants and shared pinned JPL data. These are independent implementation comparisons, not independent observations. The original planet/angle audit was ad hoc; its baseline results remain recorded separately from the older fixture-backed evidence above. The corrected angle grid now has the committed regression described below.
The source correction uses ERFA's matrix-based sidereal-time formulation with the engine's existing planetary frame. The full ERFA reference retains a small model difference. A committed 344-epoch regression now checks the angle grid. Skyfield-managed topocentric sunrise calculations are not changed or newly certified by this correction. Historical date support does not imply uniform precision, and some derived tools still lack independent full-range references.
The corrected birth-chart true node uses the same geometric osculating model as the lunar-node operation, with seven additional Horizons orbital-element references spanning the historical range. Mean Rahu/Ketu remain unchanged; existing saved snapshots are not rewritten. Modeled Delta-T, uncertain ancient timekeeping, star extrapolation, undefined polar geometry and sampled-search limits still apply.
Historical derived-result checks
A retained Astropy/ERFA reference set now checks 90 epochs spanning 1000 BCE–2500 CE in both calendars: 900 planetary positions, 4,050 aspect pairs, 900 composite midpoints, 90 Panchang snapshots, 2,520 D2/D3/D9/D12 planet mappings and 90 Vimshottari birth balances. A separate Moon-phase scan and bisection checks all 37 events in eight 35-day windows; the largest sampled timing difference was 0.506 seconds when rounded upward.
These checks use independent coordinate calculations with the same JPL kernels and modeled time conversion. They do not certify every Varga, nested dasha, sunrise, search or historical instant. A D9 sample near a subdivision boundary is retained with an explicit uncertainty check. The machine-readable evidence records the fixture hash, measured differences and limitations; supported dates are not a uniform accuracy guarantee.
What customers can verify
- The public evidence file publishes aggregate comparison counts, sample designs, observed maxima, published ceilings, and fixture hashes, plus one runnable golden vector with input, expected values, and tolerance.
- The raw independent comparison fixtures are not publicly distributed; their committed hashes, counts, and executable replay remain operator-verifiable while provenance and licensing review remains open.
- Calculation responses state their zodiac, ayanamsa, house model, frame, and other route-specific conventions rather than implying interchangeability with another engine.
POST /api/astro/natal-contextproduces deterministic JSON, XML, or Markdown fact bundles with stable fact IDs, source paths, a calculation hash, and artifact hashes.- Place and time resolution identifies GeoNames snapshot
2026-10-02, IANA tzdb2026d, and Pythontzdata 2026.4.
Interpret the evidence correctly
A deterministic result can still differ from another product because of ephemeris, frame, Earth-orientation, historical timezone, house-system, boundary, or rounding choices. The evidence supports reproducibility of Cosmic Ephemeris's declared model; it is not scientific certification, practitioner approval, or predictive validation.