Headline results from fifteen locked tests: confirming sound alignments, refuting artifacts, distinguishing sibling monument classes with opposite sky targets, and — most importantly — detecting the engine's own analytic errors.
Correspondence: scalarflower.com. Manuscript type: Methods / validation study. With computational collaboration by the Hermes agent (Nous Research).
Background. Archaeoastronomical alignment claims are notoriously vulnerable to three failure modes: privileged-target selection (fitting a monument to a favored sky object), multiple-comparisons inflation (many candidate targets guarantee a "hit"), and post-hoc goalpost-shifting. We ask whether a deliberately minimal positional-astronomy engine, constrained by a locked pre-registration protocol and a formal self-audit, can adjudicate such claims without succumbing to these failure modes — and, critically, whether it can detect errors in its own pre-registered analysis.
Methods. A first-principles engine computes solar, lunar, and (via IAU precession) stellar rising/setting azimuths and declinations for arbitrary epoch, latitude, and horizon altitude. Every test was pre-registered: hypotheses, null models, pass/fail declination and azimuth bands, and sanity anchors were frozen before computation. Population orientation was tested in azimuth space using the Rayleigh statistic with a Monte-Carlo uniform-orientation null (N = 105). A four-tag epistemic protocol ("the membrane") separated measured results REAL, structural correspondences ARCH, interpretation INTERP, and falsified hypotheses NULL. Fifteen tests spanned nine single-site claims, three monument populations, and three tests of a human-outcome hypothesis.
Results. The engine reproduced well-supported solar alignments to arcminute precision (Göbekli Tepe 1.4′; Serpent Mound 0.04°) and correctly refuted two published/popular numerological claims (an eight-site "Regulus at azimuth 90°" claim; Thom's Megalithic Yard) on locked criteria alone. On three monument populations analyzed identically, it recovered three different targets: the Scottish Recumbent Stone Circles cluster on the southern major lunar standstill (R = 0.935, p = 2×10−4); the neighbouring Clava cairns cluster on the midwinter Sun (R = 0.911, p = 1×10−5); and the &Gcaron;gantija-phase Maltese temples cluster on southern stars of the Crux/Centaurus complex (R = 0.913, p = 9×10−5), independently reproducing a peer-reviewed 2025 result via a separate pipeline. The protocol caught five pre-registration or data-integrity errors — including a sign-dropping OCR corruption in a source table that would have inverted the Malta verdict — and, in one adversarial round, identified a blind spot in the engine itself. Three confound-controlled tests of the program's own interpretive hypothesis (n up to 32,507) returned clean nulls; in one case a nominally significant raw effect (p = 0.014) was shown by the locked control to be 98.9% a birth-year artifact.
Conclusions. A minimal, pre-registered, self-auditing engine can serve as a disciplined adjudicator of archaeoastronomical claims: confirming sound alignments, refuting artifacts, distinguishing sibling monument classes with opposite targets, and detecting its own analytic errors and structural limits. We advocate pre-registration and formal self-audit as standard practice for alignment studies.
Claims that ancient monuments encode astronomical alignments range from the rigorously demonstrated (Newgrange's winter-solstice passage; Chankillo's solar towers) to the statistically dubious (universal metrological units; multi-site single-star alignments). The field's central methodological hazard is that the sky is crowded: with enough candidate targets, latitudes, epochs, and horizon assumptions, almost any azimuth can be "explained." Robust adjudication therefore requires (i) targets fixed a priori, (ii) an explicit random-orientation null, and (iii) protection against post-hoc rationalization.
We report a validation program built around these requirements. The instrument is intentionally minimal — a first-principles positional-astronomy engine with no free fitting parameters — and is wrapped in two disciplines borrowed from clinical trials and preregistered psychology: a locked pre-registration for every test, and a four-tag epistemic protocol that quarantines interpretation from evidence. We treat the engine's ability to catch its own mistakes as the primary validation criterion.
2.1 The engine. Rising/setting azimuth A for a body of declination δ at geographic latitude φ and horizon altitude h is obtained from the standard horizontal–equatorial transform, sin δ = sin φ sin h + cos φ cos h cos A, inverted for the target declinations. Solar and lunar targets (equinox δ = 0; solstice δ = ±ε; lunar standstills δ = ±(ε ± i), with obliquity ε and lunar orbital inclination i = 5.145°) are computed at the epoch-appropriate obliquity. Stellar targets are precessed from catalog (ICRS/FK5) positions to the construction epoch using IAU precession plus proper motion; ephemeris-based observation is avoided for deep antiquity (no standard ephemeris spans ~3300 BCE), so fixed-star declinations are obtained by precession alone.
2.2 Population statistics. Monument-axis azimuths were tested for non-random clustering with the Rayleigh statistic (resultant length R, Z = nR2), and significance was assessed by a Monte-Carlo uniform-orientation null (N = 105), with p = (r + 1)/(N + 1). Clustering was tested in azimuth space only; declination was used solely to interpret which sky target a confirmed cluster matched. This convention was adopted after an early diagnosis that the azimuth→declination transform concentrates density at the southern declination extreme under a uniform-azimuth null, manufacturing spurious "southern" peaks if significance is tested in declination space.
2.3 Pre-registration and the four-tag protocol. For each test, the hypothesis, null model, pass/fail bands, decision tree, and ≥3 sanity anchors (reproductions of independently known values) were written and locked before computation. Auto-scored misses against locked bands were reported as misses. A guard against three canonical failure modes was applied: C1 (privileged constant / circularity), C2 (multiple comparisons / selection), C3 (moved goalpost / retrodiction). Every reported statement was tagged REAL, ARCH, INTERP, or NULL.
2.4 Human-outcome tests. The program's interpretive hypothesis — that the lunar node's declination amplitude ("vertical breath") at birth carries meaning about a person — was tested on large athlete datasets with exact birth dates and objective outcomes (Olympic medaling; age at death). Designs controlled the birth-year confound either by within-year stratification or by using a lunar variable with low birth-year variance (daily declination, R2 = 0.061 on birth-year, vs. the breath variable's 0.989). A stated honest prior of <5% probability of a real effect was recorded before analysis.
Table 1. Single-site tests.
| # | Site | Claim | Verdict | Precision / note |
|---|---|---|---|---|
| 34 | Newgrange / Stonehenge | Winter-solstice passage | HIT + UNRESOLVED | Inside surveyed window (134.12°); exact single-figure reproduction unresolved (source horizon convention) |
| 35 | Giza | Sphinx due-east; pyramid solstice sunset | HIT (weak) | Geometry-guaranteed hits; caveats reported |
| 36 | Göbekli Tepe | Enclosure solstice sunrise | HIT | 1.4′ match (tightest solar hit); self-caught band error |
| 37 | 8-site "Regulus 90°" | Intentional stellar alignment | REFUTED | Astronomically inevitable; C1/C2 artifact |
| 38 | Chankillo | Non-linear solar "hover" | HIT | Rate ratio 6.96 reproduced (curve shape) |
| 39 | Callanish | Major lunar standstill azimuth | MIXED / MISS | ~5° miss; engine omits lunar parallax (reported limit) |
| 40 | Stonehenge Station Stones | Lunar intent | REAL geom, WEAK evidence | Standstill ≈ perpendicular-to-solstice at 51°N (near-forced); engine blind to cremation evidence |
| 41 | Megalithic Yard | Universal 2.72-ft unit | REFUTED (artifact) | 79 units fit Thom's data better; phantom-unit mechanism demonstrated |
| 42 | Serpent Mound | Solstice head; deep-time date | HIT (0.04°); Hancock refuted | Fit best at 1070 CE, worst at 10000 BCE |
Table 2. Population tests (identical machinery).
| # | Population | n | Rayleigh R | Monte-Carlo p | Target |
|---|---|---|---|---|---|
| 43 | Scottish Recumbent Stone Circles | 37 | 0.935 | 2×10−4 | Southern major lunar standstill |
| 49 | Clava cairns (SW classic) | 11 | 0.911 | 1×10−5 | Midwinter Sun |
| 50 | Maltese temples (&Gcaron;gantija phase) | 9 | 0.913 | 9×10−5 | Southern stars (Crux/Centaurus) |
Two geographically and culturally adjacent Scottish classes (43, 49) resolved to opposite targets (Moon vs. Sun) under identical rules; Malta (50) resolved to a third class (stars), independently reproducing Silva & Lomsdalen (2025).
Table 3. Human-outcome tests (interpretive hypothesis).
| # | Design | n | Result | Verdict |
|---|---|---|---|---|
| 46 | Breath-state vs. Olympic medaling | 11,538 | Raw p = 0.014 shown to be 98.9% birth-year artifact | NULL |
| 47 | Daily |δ| vs. medaling (de-aliased) | 11,538 | Flat, sign-wrong, no signal | NULL |
| 48 | |δ| vs. longevity (continuous) | 32,507 | +0.31 yr across full range, p = 0.34 | NULL |
3.1 Self-audit outcomes. The protocol caught five errors during the series: (1) a mis-defined band floor at Göbekli Tepe (pillar-orientation vs. sunrise target); (2–3) two over-tight locked declination bands in the recumbent-circle population test; (4) an impossible 71.7° declination from a pole-rotation bug at Chaco, caught by a sanity anchor; and (5) a sign-dropping OCR corruption in the Malta source table, caught pre-analysis because south-facing temples at 36°N cannot have positive declinations. In the Stonehenge adversarial round the audit identified a limitation of the engine rather than the data: it cannot represent non-geometric archaeological evidence (burial clustering, stone provenance) that bears on intentionality.
The engine behaves as a disciplined adjudicator rather than an advocate. It confirms sound solar alignments at arcminute precision, refutes two well-known non-astronomical claims on criteria fixed in advance, and — the strongest evidence of instrument validity — distinguishes sibling monument classes that point at different sky targets without being tuned to any of them. The Malta result demonstrates the engine extends beyond Sun/Moon to precession-dependent stellar targets and reproduces a recent peer-reviewed finding through an independent route.
We emphasize three honest limitations. First, the Malta stellar result replicates a minority-but-rising position in a contested field; it validates a method, not a settled archaeological consensus, and cannot distinguish deliberate stellar targeting from a terrain/wind-driven southern bias the stars happened to occupy. Second, the engine is solar-derived and omits lunar horizontal parallax, producing a real ~5° azimuth miss at Callanish that we report rather than conceal. Third, geometry alone cannot resolve intentionality where an alignment is near-forced by latitude (Stonehenge Station Stones).
The human-outcome nulls are, by our pre-registered prior, the expected outcome and do not bear on the astronomy engine. Their methodological value is the demonstration that a nominally publishable effect (Test 46, p = 0.014) can be entirely a calendar artifact, exposed only by a locked control — a cautionary result for any correlational program that touches birth dates.
Pre-registration, an explicit random-orientation null, azimuth-space testing, and a formal self-audit together let a minimal positional-astronomy engine adjudicate alignment claims with a low false-positive rate and, crucially, the capacity to catch its own errors. We recommend these practices as a standard for archaeoastronomical alignment studies.