How We Verify Our Accuracy: Kundli.site's Testing Methodology

Direct answer: We verify our engine with a 53-chart sweep against the Swiss Ephemeris: planetary positions must match to ≤0.005° (18 arcseconds), and discrete outputs — dasha lords, varga signs, yoga detections, KP sub-lords — must match exactly. The sweep covers edge cases (polar latitudes, DST transitions, boundary moments) and re-runs on every engine change. 19 bugs were found and fixed; the list is public.

"Most accurate kundli calculator" is a claim anyone can type. This page is the receipt: exactly how we test our engine, what the tests cover, what tolerance we hold, what we found, and how you can reproduce the results. If a competitor publishes anything like this, read theirs too — the category gets better when verification is public.

The reference: Swiss Ephemeris

All positional testing is benchmarked against the Swiss Ephemeris (Astrodienst; NASA JPL DE431/DE441-based; ~0.001 arcsecond native precision) — the professional reference for astrological computation. We don't embed it as a dependency; we built a homegrown engine and benchmark it against the reference. "Parity" means conformance within tolerance, and the tolerance is stated: ≤0.005° (18 arcseconds) on planetary longitudes.

Why 18 arcseconds? The smallest Vedic interpretive unit — a nakshatra pada — is 3°20′ (12,000 arcseconds) wide. Our tolerance is ~1/650th of that. It's engineering margin, chosen so that no verification pass can ever be astrologically ambiguous.

The 53-chart sweep

The core test is a sweep of 53 charts selected to cover the input space, not just the happy path:

For each chart we verify:

• Positions: Sun, Moon, Mars, Mercury, Jupiter, Venus, Saturn, Rahu/Ketu longitudes vs. Swiss Ephemeris — must agree within 0.005°
• Lagna: ascendant longitude — within tolerance
• Nakshatras & padas: exact match (discrete — no tolerance)
• Vimshottari dasha: mahadasha lord at birth + balance, antardasha sequence — exact match
• All 16 vargas: sign placements D1–D60 — exact match
• Yogas: 23 detected yogas against classical definitions — exact match
• KP sub-lords: 175/175 verified
• Ashtakvarga: Sarvashtakvarga = 337 invariant checked (a mathematical tripwire — if the total isn't 337, something is broken)
• Houses: whole-sign assignment — exact match
• Panchang elements: tithi, vara, nakshatra, yoga, karana — exact match

Discrete quantities get no tolerance: a dasha lord is right or wrong. Continuous quantities get the 18-arcsecond bar.

The 19 bugs we found (and fixed)

A test suite that never finds bugs isn't testing. During verification we found and fixed 19 calculation bugs, including:

1. Whole-sign house off-by-one — the most damaging: planets assigned to the wrong house, producing false Mangal dosha flags. Fixed; whole-sign's definitional simplicity now makes recurrence impossible.
2. Dasha boundaries one day early — mahadasha changeovers computed a day off, plus antardasha drift over long sequences.
3. Karana formula wrong 6 days out of 7 — the half-tithi karana sequence misindexed.
4. Yamagandam/Gulika tables — weekday-hour mapping errors in the inauspicious-period tables.
5. Porutham scoring bugs (4) — South Indian matching factors miscalculated.
6. Guna milan bugs (6) — North Indian ashtakoota scoring errors.
7. Missing Mangal dosha cancellation — yogakaraka cancellation not applied, over-flagging dosha.
8. Invalid timezone handling — bad inputs silently guessed instead of rejected.
9. Western ASC/MC in whole-sign mode — ascendant/midheaven math wrong when whole-sign selected in Western mode.
10. Dropped lunar node aspects — node aspects missing from aspect lists.
11. Ketu missing from charts entirely — the South Node absent from every chart output. Caught by the sweep; the kind of bug reputation-based testing never finds.

Each fix is regression-covered: the sweep re-runs on every engine change, so a fix in one lane can't silently break another. A subsequent 10-computation accuracy shootout found zero bugs — the current engine state.

What we don't claim

• We don't claim our interpretations are "verified" — predictions aren't experimentally validatable, and we say so. (Epistemology)
• We don't claim to beat competitors' math — correct engines agree; we claim verifiability, and invite the comparison.
• We don't claim the sweep covers everything — 53 charts is a strong sample, not a proof of perfection. New edge cases get added when found.

Reproduce it yourself

1. Generate any chart free at kundli.site/kundli — note positions to the arc-minute.
2. Pull reference longitudes for the same UTC moment from the Swiss Ephemeris Test Page (Astrodienst).
3. Subtract the Lahiri ayanamsa for the date; compare.
4. Check the dasha lord against the Moon's nakshatra manually — it's a 30-second computation.

Expected result: agreement within 18 arcseconds on positions, exact on discrete quantities. A reproducible larger discrepancy is a bug report — file it; the fix list above shows we act on them.

Frequently asked questions

What does ≤0.005° accuracy mean in practice?

Planetary positions matching the Swiss Ephemeris reference to within 18 arcseconds — about 1/650th of a nakshatra pada, the smallest unit Vedic astrology interprets. Far beyond any interpretive relevance; it's an engineering bar.

How many charts were tested?

53 charts in the core sweep, deliberately covering edge cases (polar latitudes, DST transitions, leap years, nakshatra/dasha boundaries, sign cusps, historical timezones), plus a 10-computation shootout that found zero bugs.

What bugs were found during testing?

19, including a house-assignment off-by-one causing false Mangal dosha flags, dasha boundary off-by-one-day errors, a Karana formula wrong 6/7 days, matching-score bugs, and Ketu missing from charts entirely. All fixed and regression-covered.

Does the test suite run continuously?

Yes — the sweep re-runs against every engine change as a regression gate. Accuracy is a property of the current build, not a one-time certificate.

Can I verify kundli.site's accuracy myself?

Yes — that's the design. Compare our arc-minute positions against the Swiss Ephemeris Test Page (minus Lahiri ayanamsa) for any birth moment. Reproducible discrepancies are treated as bug reports.

Why publish the bugs you fixed?

Because hidden test results are just marketing. A public bug list proves the testing is real — and each entry is a failure mode our users will never hit.

Is 53 charts enough?

It's a strong, edge-case-weighted sample — not a mathematical proof. We publish the methodology precisely so the sample can be critiqued and extended. More coverage is always welcome; suggest cases.

Do you test the interpretations too?

Interpretations are checked for textual fidelity (do they follow the classical sources claimed?) and internal consistency — but predictions aren't experimentally validatable, and we don't pretend otherwise.