NatalChart.AI accuracy test set, version 2026.11.21.2 Run 2026-11-21T23:05:52.828Z. Engine files of commit 861cd1c; their SHA-256 are in report.json (environment.engineSourceFiles). WHAT THIS IS 13 synthetic birth charts (none is a person) and 4 conversion-only time cases, computed by the production engine and compared with references the engine does not use. Four tests run separately, so an error in one part cannot hide in another: 1. positions engine vs JPL Horizons, and vs the Swiss Ephemeris library 2. angles and houses engine vs an implementation that shares no code with the library 3. time zones engine vs Python zoneinfo, and the zone found from coordinates 4. signs and aspects engine vs signs and aspects recomputed from Horizons Inputs, tolerances and pass rules are in fixtures.json. They were fixed before the run and are never changed to fit a result; a case outside its tolerance is published as a failure. ENVIRONMENT OF THIS RUN engine sweph 2.10.3-b-1 (Swiss Ephemeris 2.10.03), flags SEFLG_SWIEPH | SEFLG_SPEED, tropical zodiac, Placidus, files in public/ephemeris: seas_18.se1 sha256 5fd9c2aa1654e37c09a6aeb558076e795409b7dc4bd948ebc0faa7d4a7686b5b sefstars.txt sha256 18b0dcafbe5b7240773daba2c038a325f5b3fc4163f61e0a7f4e92abd4f517c6 semo_18.se1 sha256 ecfa54dbf5bc0b5a9bc3e04ed28629a821e98625eacae38f4070593bba0e2980 sepl_18.se1 sha256 0b7e416e3c1be9e6a0dd1d711dae7f7685793a0e7df13f76363a493dc27b6ea1 host Node 22.22.0, ICU 77.1, TZDB 2025b, TZ=UTC (as Vercel) time reference Python 3.14.4 zoneinfo, TZDB 2026f Horizons API 1.2, DE441, JPL#171, JPL#48, JPL#74, JPL#144, JPL#36, EOP file eop.261120.p270216 HOW TO RERUN Inside the site's source code (the engine is not a published package): npm ci npx tsx --tsconfig tsconfig.json scripts/generate-accuracy-bundle.ts The script needs the network and python3 (3.9 or later, standard library only), sets TZ=UTC for itself, and writes public/evidence/accuracy// and src/lib/demo/accuracy-bundle.json. Horizons may have updated an orbit since this run; differences in the last digits are then expected. Without the source, every reference value can be recomputed independently and set against the engine's values, which are in report.json and the CSV files: the Horizons requests are listed in report.json (horizons.requests) and described in step 2, the house reference's formulas are in step 3, and the time zone reference is the Python program in the appendix. STEP 1: THE ENGINE For each vector, calculateNatalChart({date, time, latitude, longitude, timezoneId}) from src/lib/astro/calculate-natal-chart.ts, under TZ=UTC. Its Julian Day (UT) is the moment every later step compares at, so a time zone error cannot leak into the position or angle tests. STEP 2: POSITIONS One Horizons request per body with all Julian Days in TLIST: COMMAND 10 Sun, 301 Moon, 199 Mercury, 299 Venus, 4..9 system barycentres of Mars..Pluto (what the JPL files and the Swiss Ephemeris tabulate), 2060; Chiron, 1; Ceres, 2; Pallas, 3; Juno, 4; Vesta CENTER 500@399, QUANTITIES 2,31, ANG_FORMAT DEG, TLIST_TYPE JD, TIME_TYPE UT, CAL_FORMAT JD, CSV_FORMAT YES, EXTRA_PREC YES Example (Saturn, vector V01): https://ssd.jpl.nasa.gov/api/horizons.api?format=text&COMMAND=%276%27&OBJ_DATA=%27NO%27&MAKE_EPHEM=%27YES%27&EPHEM_TYPE=%27OBSERVER%27&CENTER=%27500%40399%27&TLIST_TYPE=%27JD%27&TIME_TYPE=%27UT%27&TLIST=%272448506.1458333335%27&QUANTITIES=%272%2C31%27&ANG_FORMAT=%27DEG%27&CAL_FORMAT=%27JD%27&CSV_FORMAT=%27YES%27&EXTRA_PREC=%27YES%27 Compared: the engine's longitude (stored to 0.0001°) minus ObsEcLon, folded into ±180°, in arcseconds. Also the library called directly (swe_calc_ut, SEFLG_SWIEPH | SEFLG_SPEED) minus the engine. STEP 3: ANGLES AND HOUSES The reference is in scripts/generate-accuracy-bundle.ts and shares no code with the Swiss Ephemeris: sidereal time IAU 2006 GMST (IERS Conventions 2010, eq. 5.32) from the Earth rotation angle, plus the equation of the equinoxes, Δψ cos εA + 0.00264096″ sin Ω + 0.00006352″ sin 2Ω obliquity IAU 2006 mean obliquity (eq. 5.40) + Δε nutation IAU 1980, the 63 terms of Meeus, Astronomical Algorithms, table 22.A (about 0.01″ from the IAU 2000 model the library uses) MC atan2(sin θ, cos θ cos ε), θ = sidereal time + east longitude Ascendant atan2(cos θ, −(sin θ cos ε + tan φ sin ε)) Placidus each of cusps 11, 12, 2, 3 is the ecliptic point whose distance east of the meridian in right ascension is 1/3 or 2/3 of its diurnal semi-arc (11, 12), or 60° + 2/3 and 120° + 1/3 of it (2, 3); solved by scanning and bisection, so a missing root shows up as "undefined" instead of a number Porphyry each quadrant divided into three equal arcs. Above the polar circle (|latitude| at least 90° minus the true obliquity) the engine returns Porphyry cusps and reports houseSystem "Porphyry"; there its cusps are compared with this computation, and Placidus is not computed. Every chart also has one case for the label: Placidus below the polar circle, Porphyry above it. Both sides read the engine's Julian Day as UT1, as the library does. Before use the reference reproduces Meeus examples 12.a and 22.a (see fixtures.json). Cross-checks, published but not pass/fail: sidereal time Horizons quantity 7 at CENTER coord@399, SITE_COORD 0,0,0 (on the equator, where polar motion does not move the meridian): https://ssd.jpl.nasa.gov/api/horizons.api?format=text&COMMAND=%2710%27&OBJ_DATA=%27NO%27&MAKE_EPHEM=%27YES%27&EPHEM_TYPE=%27OBSERVER%27&CENTER=%27coord%40399%27&COORD_TYPE=%27GEODETIC%27&SITE_COORD=%270%2C0%2C0%27&TLIST_TYPE=%27JD%27&TIME_TYPE=%27UT%27&TLIST=%272448506.1458333335%27&QUANTITIES=%277%27&CAL_FORMAT=%27JD%27&CSV_FORMAT=%27YES%27&EXTRA_PREC=%27YES%27 Before 1962 Horizons reads the time as UT1, as the engine does; after 1962 as UTC, and then the difference is UT1 − UTC. obliquity the rotation between Horizons' apparent RA/Dec (quantity 2) and ecliptic longitude/latitude (quantity 31) of the same bodies STEP 4: TIME ZONES For each vector and probe, birthTimeToUTC() from src/lib/astro/timezone.ts, against Python zoneinfo for fold 0 and fold 1 (PEP 495), with the zone named in the fixture. A local time is "normal" when both folds agree, a "gap" when it does not exist, an "overlap" when it happens twice. The expected answer is fold 0 in every case: in a gap and in an overlap that is the engine's stated rule (RFC 9557 "compatible"). The zone from coordinates is resolveTimezone() from src/lib/timezone.ts (geo-tz), against the expected IANA zone after following the TZDB's Link lines in tzdata.zi. The same conversions are repeated with the host at UTC, America/New_York, Europe/Kyiv, Asia/Tokyo, and every local minute from a day before to a day after each of 8 clock changes (Europe/London 2021-03-28, Europe/London 2021-10-31, Europe/Kyiv 2021-03-28, Europe/Kyiv 2021-10-31, America/New_York 2021-03-14, America/New_York 2021-11-07, Australia/Sydney 2021-04-04, Australia/Sydney 2021-10-03) is converted by both sides, again against fold 0. Every host zone must give the same UTC as the UTC host. STEP 5: SIGNS AND ASPECTS Signs of the 15 bodies from the Horizons longitudes, of the MC and Ascendant from the reference. Aspects between those 15 bodies from the Horizons longitudes with the engine's own table (src/types/astro.ts): Conjunction 0° orb 8°, Sextile 60° orb 6°, Square 90° orb 8°, Trine 120° orb 8°, Opposition 180° orb 8°, inclusive, compared with the engine's list. TOLERANCES (fixed before the run) positions, Sun to Pluto 1″ positions, Chiron, asteroids 2″ engine vs library 0.19″ MC, Ascendant, cusps 1″ (Porphyry cusps above the polar circle: 1″) house system label exact: Placidus, or Porphyry above the polar circle local time to UTC 0 s, fold 0 inside a gap or an overlap zone from coordinates exact, after TZDB links signs exact aspects same list, orbs within 0.01° The reasoning for each is in fixtures.json under "tolerances". PREVIOUS RUN Run 2026.11.21 (2026-11-21T21:42:01.481Z) stays published as it was, in /evidence/accuracy/2026.11.21/, as the evidence of the failures it found. Fixed since then, on 2026-11-21: tzSeconds, tzWrongInstant, polarPorphyry (src/lib/astro/timezone.ts, src/lib/astro/sweph-init.ts, src/lib/astro/houses.ts, src/lib/api/chart.ts). report.json (fixes) sets each one before and after. FILES AND HASHES manifest.json lists the SHA-256 of every other file in this folder. To check: shasum -a 256 report.json positions.csv angles-houses.csv timezones.csv \ signs-aspects.csv fixtures.json protocol.txt and compare with manifest.json. The hash of manifest.json itself is published on https://www.natalchart.ai/llm-hub/accuracy-report. APPENDIX: THE TIME ZONE REFERENCE Run with python3 -I -c '' and a JSON object on standard input: {"zones": [zone names to resolve through tzdata.zi links], "cases": [{"zone": "Europe/Paris", "local": "1900-05-15T14:00"}, ...], "scans": [{"zone": "Europe/Kyiv", "date": "2021-10-31", "minutes": 4320}, ...]} It prints, for every case, the UTC of fold 0 and fold 1, the offset in seconds, and whether the local time is normal, in a gap or in an overlap. import json, sys from datetime import datetime, timedelta, timezone from zoneinfo import ZoneInfo, TZPATH req = json.load(sys.stdin) out = {"python": sys.version.split()[0], "tzpath": list(TZPATH)} def tzdb_version(): for p in TZPATH: for name in ("+VERSION", "tzdata.zi"): try: with open(p + "/" + name) as f: line = f.readline().strip() return line[len("# version "):] if line.startswith("# version ") else line except OSError: pass return None def tzdb_links(): for p in TZPATH: try: links = {} with open(p + "/tzdata.zi") as f: for line in f: if line.startswith("L "): _, target, name = line.split() links[name] = target return links, p + "/tzdata.zi" except OSError: pass return {}, None out["tzdb"] = tzdb_version() links, links_file = tzdb_links() out["linksFile"] = links_file def canonical(z): seen = set() while z in links and z not in seen: seen.add(z) z = links[z] return z out["canonical"] = {z: canonical(z) for z in req["zones"]} EPOCH = datetime(1970, 1, 1) def convert(zone, naive): z = ZoneInfo(zone) folds = [] for fold in (0, 1): aware = naive.replace(tzinfo=z, fold=fold) off = aware.utcoffset() utc = naive - off back = aware.astimezone(timezone.utc).astimezone(z).replace(tzinfo=None) folds.append({"utc": utc.strftime("%Y-%m-%dT%H:%M:%S"), "epoch": int((utc - EPOCH).total_seconds()), "offsetSeconds": int(off.total_seconds()), "exists": back == naive}) if folds[0]["epoch"] == folds[1]["epoch"]: kind = "normal" elif folds[0]["exists"] and folds[1]["exists"]: kind = "overlap" else: kind = "gap" return {"kind": kind, "fold0": folds[0], "fold1": folds[1]} out["cases"] = [convert(c["zone"], datetime.fromisoformat(c["local"])) for c in req["cases"]] scans = [] for s in req["scans"]: start = datetime.fromisoformat(s["date"]) - timedelta(days=1) f0, f1 = [], [] for i in range(s["minutes"]): r = convert(s["zone"], start + timedelta(minutes=i)) f0.append(r["fold0"]["epoch"]) f1.append(r["fold1"]["epoch"]) scans.append({"zone": s["zone"], "date": s["date"], "fold0": f0, "fold1": f1}) out["scans"] = scans print(json.dumps(out))