Extended architectural and selection analyses
Supplementary Material S9
Lucas Giovani Ribeiro and Marcos André Simonssini
Companion to Stonehenge v1.44 | 30 October 2026
This supplement provides source definitions, algebraic derivations, constructive tests, regional comparisons and phase controls supporting the main results. Citations use the main bibliography. Supplementary Notes belong to S1, while S5–S8 retain their own section numbering. All comparisons are retrospective; their candidate families and retained assumptions are specified at the point of use. Section S9.8 records information order, execution scope and the numbering crosswalk.
S9.1 Sources, measurements and drawing proxies
The primary objects are the Aubrey circuit, the four Station positions, the Sarsen outer envelope, the transverse Trilithon-horseshoe span and the Y- and Z-hole circuits. The Sarsen outer envelope is distinguished from the published inner-face diameter [23]. The four inventory counts are A=56 Aubrey positions, B=30 outer sarsen uprights, C=10 trilithon uprights and D=4 Station positions. These refer to the reconstructed inventory across the monument's history, including absent settings; the four counts do not describe an inventory present at one date. Table S9.1 defines the notation and distinguishes the evidential status of the inputs.
Table S9.1. Operational notation and evidential status. Published measurements, approximate summaries, digitized proxies and model assumptions enter different operations; their sources and boundaries are specified in S9.1–S9.2 and S9.5. The reconstructed inventory spans construction phases.
Symbol or input
Meaning and use
Evidential status
N, s
Aubrey and Station outer counts; N=A and s=D
Published archaeological reconstruction; includes absent settings.
a, b, c
Primitive short leg, long leg and hypotenuse
Derived frame selected within the declared comparison class.
A, B, C, D
Aubrey, sarsen, trilithon and Station inventory
Reconstructed counts across phases, not a simultaneous surviving inventory.
r, g, j, v
Radial, axial, longitudinal and latitudinal seeds
Derived outputs of the specified count map.
Q
Ordered eight-value register
Deterministic expansion of the seeds.
W
Ordered temporal triple
Derived numerical target; its timing interpretation is stipulated.
u
Common terrestrial scale
Parameter fitted only to the declared exterior dimensions.
Exterior measures
Published Station sides and Aubrey diameter
Published measurements with stated feature boundaries and source conventions.
Interior measures
Trilithon span, Sarsen outer envelope, Y/Z diameters
Approximate published dimensional summaries; excluded from scale fitting.
Plan coordinates
Cleal symbols and Rees labels or hatch centroids
Digitized drawing proxies; no measured per-feature survey uncertainty.
Roles and timing scale
Outer 10:24:26; inner 4,10,12,16; minute-per-numerical-unit convention
Declared model choices. The geometric roles were recognized with the plan available.
Perturbation bounds
Side factors, interior ±20% factors and ±0.25 m coordinate scenarios
Stipulated sensitivity distributions or bounds, not measured archaeological errors.
The published Station centre-to-centre distances are 34.17 and 32.70 m for the short sides and 79.93 and 80.26 m for the long sides [24–25]. Stations 91 and 93 survive; 92 and 94 are archaeologically located absent-stone settings. The primary Aubrey diameter is the 86.44 m Thom survey value tabulated by MacKie [25]. Alternative published conventions span 85.832–87.0 m [25–26]. Inner inputs are the approximately 13.7 m Trilithon span, 33.0 m Sarsen outer diameter, 39.0 m Z-hole diameter and 54.0 m Y-hole diameter [1]. Hole diameters pass through hole centres; the Sarsen measure concerns its occupied outer envelope. The inner-face convention [23] is retained as a sensitivity. The ditch [27] and Heel Stone are excluded from calibration and the inner ladder. Supplementary Note 2 records boundaries, sources and precision classes.
For the contextual plan in main Figure 1B, 56 Aubrey symbols and four Station proxies were digitized from Cleal et al.'s published Plan 1, using a 20 m printed scale bar spanning 357 pixels [1]. One raster pixel corresponds to about 0.056 m, which describes sampling rather than archaeological accuracy. Stations 91 and 93 use projected standing-footprint centres; 92 and 94 use socket-outline proxies, with incomplete excavation at 94. A leaning stone's footprint need not locate its original socket centre. The coordinates are local drawing coordinates, not georeferenced survey observations. No per-feature survey uncertainty is supplied by this extraction, and scanner precision is not assigned to these proxies. The primary dimensional fit uses the separately cited measurements above, not this new illustration (S5 Appendix, section 6). Main Figure 1C separately plots the archived Rees click-plan labels used in the interior directional diagnostics, identifying the entrance pair 1–30 and complete inner pairs 51–52, 53–54 and 57–58. They are label or hatch-centroid proxies, including separate fragment labels, rather than stone outlines or newly surveyed centres. Panels B and C have different source coordinates and are not registered together.
We estimate one scale from the exterior and apply it to the four inner roles without refitting. Calibration uses only the mean Station short side, mean long side and Aubrey diameter, with roles k=(10,24,26). For observed dimensions xᵢ and fixed model lengths kᵢ, the least-squares scale through the origin is û=Σ(kᵢxᵢ)/Σ(kᵢ²). The inner roles are k=(4,10,12,16). We report signed relative residuals and their root-mean-square value. The inner dimensions are held out of scale estimation; their model roles were recognized with the plan available. Recovery denotes retrospective agreement under these declared roles. Predicted values elsewhere in the paper are conditional model outputs; the holdout concerns scale estimation, not blind selection of the roles. Numerical outputs in the narrative are rounded for readability; tables and supplements retain reproduction precision. Printed digits do not estimate archaeological uncertainty.
Primitive Pythagorean triples with hypotenuse at most 100 define the metric comparison class. On an ideal N=56 circle of diameter Φ, chord k is Φ sin(πk/N). The joint diagnostic compares the 16 primitive frames with 13 unequal-leg complementary step pairs. Independently perturbing each published side by uniform factors of ±1%, ±2% or ±5% tests ranking stability, with 200,000 deterministic-seed trials per level. The perturbations are stipulated sensitivity scenarios. Complete estimators and comparison classes appear in Supplementary Notes 2–4 and 51.
A separate metric-compression control independently multiplies each of the four interior summaries by a uniform factor in [0.8,1.2], holds the exterior scale fixed, and assigns each perturbed dimension its nearest integer multiple of that scale. It evaluates 1,000,000 trials with seed 20260916 (Supplementary Note 9; S2, stonehenge_ladder_null.py). This local perturbation control does not repeat the selection of geometric constructions.
Plan comparisons and constructive operations
The archived digitization of Cleal et al.'s Plan 1 uses its printed scale bar [1]. A registered Rees-based click-plan provides alternative proxies [28]. We distinguish continuous angular gaps between Stations, assignment to a numbered Aubrey-anchored grid, and shape comparison after fitting one common rotation about the Aubrey centre. All 14 discrete rectangular templates k|(28−k)|k|(28−k), k=1,…,14, receive the same rotation fit. A continuous 5:12:13 rectangle and a freely proportioned rectangle are comparators. A separate free circle through the four Station proxies tests centre and scale compatibility (Supplementary Notes 20–23).
Constructive tests ask how the frame can supply the proposed lengths and axes through explicit operations. For a right frame with short side x, long side y and diagonal z, copying x, halving y, subtracting x from z and constructing the inradius (x+y−z)/2 supply the four inner lengths. The pivot transformation is T(x,y,z)=(z−x,(x+y+z)/2,x+y). We test its algebra, circle consequences and candidate physical realizations. An opening axis is the normal to the line between support proxies; axial similarity is the squared inner product of unit directions. Path resultants, pivot rays, all 220 trios of a declared 12-proxy set and a 14-shape orientation census are evaluated under disclosed rules. These diagnostics use drawing proxies and do not replace the primary survey inputs (Supplementary Notes 31–43 and 51). A separate solar audit compares the ideal long chord with obliquity, including a Laskar-model sensitivity [29].
For the path decoder, a cord with equal terminal tensions defines a resultant from the sum of its two inward-pointing terminal unit tangents. The path length defines its numerical channel, while the normalized resultant defines its axis. We compare that axis with each opening normal using K(u,n)=(u·n)². The directional score gives equal weight to J against the entrance normal 1–30 and V against the mean of the three complete inner opening normals; exchanging J and V leaves the targets and weights fixed but supplies a complementary description for orthogonal outputs, not an independent control.
We examine two specifications of the route choice. A joint enumeration applies the 34/62 nearest-integer length labels and orthogonal output axes to four symmetric route classes (Supplementary Note 54). An alternative entrance-based selector chooses between the two contour families by alignment with the entrance normal and between the two crossed families by alignment with its perpendicular. It then inspects mean lengths across each selected pair of mirrored routes; the inner-opening normals do not enter this selector (Supplementary Note 40).
S9.2 Count map, inverse and stellar register
The material map converts the architectural counts into four ordered seeds: r=A−B−C, g=C, j=B+D and v=A+C−D. A seed is a role-specific output of this map; the register is their deterministic expansion. With q=j−g and d=v−j, the expansion is Q=(r,q,d,j,d+q,v,3q,v+j), and the temporal word is W=(d,3q,r). The map is inverted without changing coefficients when the longitudinal and latitudinal roles are exchanged. A bounded census compares alternative maps; a separate algebraic analysis solves r=D², d=A/2 and p²=g²+q² jointly, with p=A−B, over all strictly positive integer inventories (Supplementary Notes 5, 28–29).
The stellar module defines g as the Deneb–Sadr longitude step, j as the Deneb–Albireo longitude span and v as the mean latitude of Deneb and Fawaris. Coordinates are expressed in a fixed mean J2000 ecliptic frame. One nearest-integer rule is applied to each descriptor before differences are formed; r=16 is supplied by the material map. Historical catalogues [30], the ecliptic of the ancient date and continuous contrast-based recovery provide separate comparisons. Longitude span and mean latitude are retained as different types throughout inversion (Supplementary Notes 7–8, 24 and 27–34).
Alongside its measured lengths, the monument supplies four architectural counts. These counts generate the ordered seeds through one map:
r=A−B−C=16; g=C=10; j=B+D=34; v=A+C−D=62.
With q=24 and d=28, their expansion is:
Q=(r,q,d,j,d+q,v,3q,v+j)=(16,24,28,34,52,62,72,96).
The inverse recovers the four counts from the ordered seeds:
A=(r+j+v)/2; B=(−r−2g+j+v)/2; C=g; D=(r+2g+j−v)/2.
Exchanging j=34 and v=62 changes the recovered Station count from four to 32. The unchanged inverse returns (56,30,10,4) in the specified order and (56,30,10,32) under exchange, leaving the other three counts unchanged (main Figure 3A). The ordered stellar roles therefore have an explicit consequence within the reconstruction.
The three specified closures admit exactly one strictly positive integer inventory within the fixed compiler, with no upper bound imposed on inventory size: (56,30,10,4). The conditions are r=D², d=A/2 and p²=g²+q², where p=A−B. The same solution supplies the doubled frame (g,q,p)=(10,24,26) and W=(28,72,16). Within these conditions, both the inventory and its overall size are fixed. The factorization proof and condition-removal controls are in Supplementary Note 29; the frame and word follow from this conditional solution (main Figure 3B).
The relation can also be read directly from the outer counts and primitive frame. With N=56, s=4 and (a,b,c)=(5,12,13), the identities N=2s(b−a) and N+s²=s(c+a) give W=s(b−a,c+a,s)=(N/2,N+s²,s²). The outer inventories yield 56−30=26 and 30−26=4. These equivalent expressions locate the same frame and word in the chosen material system; their full algebraic and quadratic Gram forms are preserved in Supplementary Notes 17–19 and 52.
Two of 40,960,000 ordered maps return the target seeds, and both are invertible. The comparison is complete and bounded, with each coefficient restricted to −1, 0 or 1. A separate enumeration over ordered count quadruples with entries from 1 to 60 finds that 2.55% can express all four targets within this coefficient class. These are conditional counts of formula and input flexibility; the classes and wider-coefficient sensitivity are specified in Supplementary Notes 5 and 11.
The primary J2000 stellar descriptors are g=10.488°, j=34.078° and v=62.160°, rounding to (10,34,62). Propagating stellar positions to 2965.2 BCE while retaining the fixed J2000 ecliptic frame gives (10.485°,34.075°,62.141°), with the same rounded tuple. With material r=16, these descriptors supply the same Q and inverse. Fixed-plane stability is retained across the declared 10,000-year window under both archived motion prescriptions (Supplementary Notes 24 and 33).
A second, explicitly conditioned route uses contrasts. Given D=4, q=24 and d=28, the compiler gives A=q+d+D=56 and B−C=q−D=20. Adding r=D²=16 gives B+C=40, hence B=30 and C=10. This recovers the inventory for both tested absolute tuples, (10,34,62) and (11,35,63). Without rounding the ancient contrasts, the same inventory uniquely minimizes the constrained squared error: RMS 0.128°, compared with 0.686° for the next inventory (Supplementary Notes 28 and 33).
Historical catalogues provide a further portability comparison. Their Cygnus figures retain an approximately 16° minimum-cover radial scale; the Ulugh Beg values also round to (10,34,62), whereas Ptolemy's values produce a non-target inverse [30]. The detailed catalogue readings, modern coordinate definitions and provenance of the prior grammar are preserved together in Supplementary Table S53.1 and Notes 7–8. Here the radial seed remains a material input.
S9.3 Dimensional recovery and phase structure
The 5:12:13 rectangle ranks first among the 16 primitive candidates, with 1.57% side RMS against 6.64% for the next candidate, 36:77:85. Combining metric and cyclic comparisons again selects 5:12:13 with the 7|21|7|21 partition, first among 208 combinations. It remains first in every side-perturbation trial at ±1% and ±2%; at ±5%, it remains first in 96.72% of metric-only and 97.63% of joint trials. Main Figure 1 locates the archaeological setting and the exterior positions used by the reconstruction.
The discrete construction gives hypotenuse-normalized legs 4.975 and 12.010, with 0.36% RMS against 5 and 12. At the independently published 86.44 m diameter, its chords are 33.08 and 79.86 m, differing from the observed short- and long-side means by −1.06% and −0.29%. The side ratio therefore supports 7|21|7|21 as an approximate discrete shape (Figure S9.1).

Figure S9.1. Idealized 56-fold Aubrey construction. Model gaps 7|21|7|21 yield chords close to the reduced 5:12:13 legs. The labels 0, 7, 28 and 35 are relative indices within the ideal model. Supplementary Notes 20–23 give the shape comparisons and the separate anchored-position test.
Among all 14 discrete templates, 7|21|7|21 gives the smallest shape residual in each archived plan convention after fitting a common rotation. The Cleal continuous gaps, 7.178|21.064|7.127|20.631 steps, round to the same sequence. In Cleal its angular RMS is 0.86°, compared with 2.81° for the next discrete template. The continuous 5:12:13 comparator gives 0.79°. At common centre, orientation and diameter, corresponding vertices of the two ideal models differ by 0.09 m. Supplementary Notes 20–23 and 32 distinguish these shape comparisons from registration with individual Aubrey positions.

Figure S9.2. Four inner dimensions predicted from the outer Station–Aubrey calibration. Each prediction uses the same scale without refitting; the combined relative RMS is 2.02%.
In the declared metric-compression control, 72,285 of 1,000,000 perturbed configurations reach or improve the observed 2.02% RMS, a rate of 7.23%; the median simulated RMS is 4.12%. These results quantify the gain over the median and the frequency of comparable nearest-integer compression in this local ensemble. The frequency describes this stipulated perturbation process, not a probability for the full reconstruction or prehistoric intention (Supplementary Note 9).
The same length-transfer operations select 7|21|7|21 from all 14 discrete templates at fixed Aubrey diameter, with 2.12% RMS against 8.75% for the next template. Cutting the 10u by 24u rectangle along its 26u diagonal gives a right triangle with inradius 4u; half its long side is 12u, and diagonal minus short side is 16u. These operations express the existing Trilithon, Z-hole and Y-hole roles, respectively (main Figure 2). Together with copying the short side, they specify how each of the four inner lengths can be constructed from the outer triangle (Supplementary Notes 31–32).
The free circle fitted to the four Cleal Station proxies approximately matches the Aubrey centre and scale: its diameter is 86.49 m, and its centre lies 0.68 m from the separately fitted Aubrey centre. The 56 Aubrey proxies have radial RMS 0.63 m against that Station-derived circle. Alternative sources, every Station omission and their sensitivity scenarios are retained in Supplementary Note 21.
Construction phases
Table S9.2. Construction-phase classification of inputs and held-out dimensions. Darvill et al.'s intervals are identified by event and source [2] (Table 3); the broader English Heritage dates are contextual summaries. These are external archaeological ranges, not uncertainty intervals from this model. Residual partitions retain the same outer-only scale (main Table 1).
Feature and model role
Archaeological phase
Consequence for the model
Aubrey circuit; A=56, diameter
Initial enclosure: 3000–2920 cal BCE in Darvill et al. [2] (Table 3); revised ditch ranges compared separately [18].
Only securely early count among A, B, C and D; its diameter alone does not calibrate the Station frame.
Station positions; D=4, short and long sides
Insecurely dated; assigned to Stage 2 (2620–2480 cal BCE) in the reconstruction of Darvill et al. [2] (Table 3).
Required for the 5:12:13 calibration and the four-count compiler; a date near the initial enclosure is unproved.
Sarsen circle and trilithons; B=30, C=10; two inner dimensions
Stage 2: 2620–2480 cal BCE [2] (Table 3); commonly summarized as c. 2500 BCE.
Two held-out roles: 2.05% RMS under the scale calibrated from Station–Aubrey dimensions.
Y and Z hole circuits; two inner diameters
Y/Z excavation: 1630–1520 cal BCE [2] (Table 3); broader heritage summary, 1800–1500 BCE.
Two held-out roles: 1.99% RMS under that same scale; their agreement cannot independently date an early plan.
The earthwork and Aubrey circuit belong to the early monument around 3000 BCE. Darvill et al.'s chronology assigns the major sarsen settings and Stations to Stage 2, 2620–2480 cal BCE, and Y/Z excavation to 1630–1520 cal BCE [2] (Table 3). The Station dating remains insecure, so their placement within that reconstruction is not an independent determination for each setting [1–3]. English Heritage summarizes the central stones as about 2500 BCE and the late holes more broadly as 1800–1500 BCE [35]. Thus A=56 is the only securely early count among the four inputs at the nominal astronomical minimum. The Station-dependent calibration and all four inner comparisons cross phases (Table S9.2; Figure S9.3). These dates constrain structures; they do not date an otherwise undocumented design.
The temporal word alone also lacks a unique material inverse. Its map is W=(A−B+C−2D, 3(B+D−C), A−B−C), a rank-three linear map with kernel generated by (1,0,1,1). Every positive integer inventory (56+t,30,10+t,4+t), t≥−3, gives the same W. With W fixed, the additional closure r=D² selects D=4 and t=0. This does not contradict the uniqueness theorem under all three closures; it identifies the extra assumptions that provide uniqueness. The theorem supplies conditional mathematical identification, while independent archaeological evidence is needed to assign those assumptions to an early phase. A free common timing scale further makes proportional words indistinguishable under E (S5 Appendix, section 5).

Figure S9.3. Construction-phase dependency of the model roles. The Aubrey circuit belongs to the earliest monument; the sarsen settings and perhaps the Stations are later, and the Y/Z circuits later still. The figure locates features used in the mathematical reconstruction and does not imply that they were constructed together or that the proposed information was transmitted between phases.
S9.4 Constructive triangles, paths and directions
The pivot transformation converts the Station frame into the doubled 8:15:17 triangle through three length operations:
T(10,24,26)=(26−10,(10+24+26)/2,10+24)=(16,30,34).
Requiring both input and output of this transformation to be right triangles forces the input ratio y/x=12/5. Its operations transfer the diagonal-minus-short-side difference, the semiperimeter and the sum of the legs. The input ratio is therefore required by the simultaneous right-triangle conditions, linking the two primitive families through one specified construction (Figure S9.4; Supplementary Note 36).

Figure S9.4. One transformation connects the two primitive triangles. The 10:24:26 right triangle supplies three operations that produce 16:30:34. Its pivot angles round to 28° and 62°, and their difference rounds to 34°. The inverse exradii recover the initial legs. All dimensions are model units, and both diagrams use local schematic axes. Approximate readings for the 7|21|7|21 chord model are retained separately.
The resulting acute angles are 28.072° and 61.928°, and their difference is 33.855°. They round to 28°, 62° and 34° under the original rule. The ideal 7|21|7|21 version retains the 62/34 angular pair when its actual transformed triangle is used. A circle-based inverse returns the initial legs: the exradii opposite sides 16 and 30 are 10 and 24. Equal-area relations, Gaussian constructions and the distinct uses of 34 as a length and an angle label are derived in Supplementary Notes 36 and 52.
These constructions produce further quantities that can be compared with the plan. The four incentres of the Station rectangle's triangular halves have a circumdiameter of 53.72 m at the existing scale, close to the published 54 m Y-ring summary. The ordered paths J=x+y and V=2z+x realize 34 and 62; their two positive rectangular solutions are (10,24,26) and (78/7,160/7,178/7). Only the first returns integer inventories in the fixed module, and its external metric RMS is 0.26% against 7.93% for the other solution (Supplementary Notes 37 and 39).
A separate inventory-angle comparison gives arctan(30/56)=28.179°, only 0.096° from the recomputed stellar difference v−j=28.083°. The Aubrey–sarsen pair ranks first among all six inventory pairs under this same function. The derived ditch centreline is 32.12–32.19u under its declared boundary convention. These auxiliary comparisons, the complete tangent ranking and the modern constellation-boundary diagnostics are retained in Supplementary Notes 15–19 and 52.
Under the same equal-tension rule, the J route joins the midpoints of opposite short sides via one long edge, giving 5+24+5=34 in the ideal frame. The V route follows diagonal–short side–diagonal, giving 26+10+26=62. Under equal terminal tensions, these paths produce resultants along the short and long axes, respectively. Each channel thus has a length and a direction supplied by the same route. Applied to the Rees proxies, their axial azimuths are 52.15° and 140.04°: oblique axes with east–west and north–south predominance in the grid. The two channels therefore have distinct directional consequences under one stated operation.
The entrance can select these route families within the alternative directional protocol. J maximizes alignment with the normal to the 1–30 opening; V maximizes alignment with that normal’s perpendicular. In both Rees conventions, the selected families have ideal-frame lengths 34 and 62 when their lengths are inspected after selection. This gives the entrance an explicit role in choosing which routes are read (Supplementary Note 40).
The two specifications exchange inputs and outputs. Under the entrance-based selector, the J–entrance alignment is the selection criterion and the ideal-frame lengths 34 and 62 are its outcome; under the length-based enumeration, the length labels and orthogonality are inputs and the alignments reported below are outcomes. Both read the same plan geometry and are not independent confirmations.
Under the length-based enumeration, the J path, labelled 34, gives an axis only 0.45° from the entrance normal in the primary digitization (axial Gram 0.99994). The V path, labelled 62, has mean Gram 0.93891 with the three available inner opening normals. These axes come from the specified outer-frame path decoder and are compared with the plan without fitting a rotation. In the indicative Station 91 branch the scores are 0.99988 and 0.93802 (main Figure 5; Supplementary Note 54). The inner-opening values average the three individual squared alignments; they do not score a single averaged normal.
The compact cross-Gram retains scores 0.0684 for J against the inner normals and 0.00084 for V against the entrance. The assigned and exchanged aggregate scores are 0.9694 and 0.0346. For exactly orthogonal channels, the exchanged score is identically 1 minus the assigned score. The digitized axes are nearly orthogonal: the two scores sum to 1.0040 in the primary branch and 1.0014 in the indicative branch. Their contrast therefore restates the alignment rather than providing an independent control (Supplementary Note 54).
The two channels also restrict each other within the declared route family. Two of its 12 ordered assignments recover the 34/62 length labels; only the short-side-midpoint J route paired with the diagonal–short-side–diagonal V route also produces orthogonal resultants. The alternative route of length 34 produces an axis parallel to V. In the digitized primary branch, the J–V axial Gram is 0.00136 for the selected pair and 0.99995 for that alternative. The ideal 7|21|7|21 form retains the selected pair, giving lengths 33.971 and 61.950, which round to 34 and 62 (Supplementary Note 54).
The three complete trilithon openings have normals closer to the Station long-side family than to its short-side family. Squared axial similarities are 0.984, 0.952 and 0.883 for the long sides, against 0.029, 0.032 and 0.146 for the short sides. No rotation is fitted, and the preference survives both Station 91 conventions and the stipulated ±0.25 m coordinate perturbations (Supplementary Note 35).
The Station-only construction predicts an axis 0.61° from the entrance normal, placing that pair first among 19 consecutive proxy pairs. With grid east fixed, the pivot rays approximate two trilithon-opening normals at 1.18° RMS; the discrete model gives 1.13°. References, reflections and assignments were selected retrospectively and are retained in full (Supplementary Notes 38–41).
The exhaustive 220-triangle proxy census finds 52–54–60 closest to 16:30:34 at 0.69% RMS with its own fitted scale. Across seven primitive comparators, 5:12:13 and 8:15:17 give the two smallest best-fit discrepancies, 0.56% and 0.69%. A different census retains 541 of 252,000 shape–orientation configurations under all six criteria, every survivor having the 7|21|7|21 form. The criterion-removal analysis shows that four restrictions select the same shape; direction adds a restriction, while the horizontal/vertical label adds none after it (Supplementary Notes 42–43).
These material and astronomical relations raise a further question: which parts of the reconstruction survive alternative assignments, inputs and scales? The next sections test stellar selection, early-versus-later phase preference, continuous metric targets and joint spatial–temporal performance. The closing sensitivity analysis identifies which physical readings depend on the source or convention.
S9.5 Regional stellar and word comparisons
Temporal operator and candidate definitions
The fixed angular benchmark compares the role-ordered absolute right-ascension intervals Deneb–Sadr, Sadr–Albireo and Aljanah–Fawaris, in the equatorial frame of date, with θ=(7°,18°,4°). This is exactly the earlier calculation against W=(28,72,16) in sidereal minutes: multiplying degree intervals by four multiplies every RMS by four and leaves its minimizers, ranking and runner-up ratio unchanged. Thus the time-unit expression is an equivalent coordinate conversion. The fixed target still assumes one architectural numerical unit corresponds to one quarter degree. Since D=4 in the reconstructed inventory, θ can also be written as (W/D) degrees; this is a retrospective material parametrization of the chosen scale, not an independent derivation of its ancient use. The primary archived astrometry is propagated with the long-term Vondrák–Capitaine–Wallace model and corrigendum [31–34]. We minimize the unweighted angular RMS across astronomical years −8000 to +2000, refining every bracketed local minimum. A six-dimensional space-motion variant and channel omissions assess sensitivity. Earlier tables retain their explicitly labelled sidereal-minute scores for reproducibility. Ecliptic descriptors do not redefine the target during optimization.
A material word library is formed from N=A, s=D and the selected primitive frame. Its eight distinct values, {4,16,28,32,40,68,72,100}, generate 512 ordered triples. Each receives the same full-window evaluation. Archaeological intervals enter afterward, as comparisons or expressly identified census filters. The supplementary solar audit instead compares a dimensionless ideal chord value with obliquity expressed in degrees, under two disclosed models [29,32–33]; that numerical angle convention is an additional assumption (Supplementary Notes 9–10, 25–26 and 51).
Stellar censuses test whether the stated rules recover the proposed configuration when the stellar identities are no longer supplied. The within-region stellar search assigns all five roles to distinct Hipparcos entries within each modern IAU constellation region. It uses one uniform catalogue-epoch convention, V≤5 as the principal magnitude limit, and nested V≤4 and V≤6 comparisons. Candidates first satisfy the fixed spatial tuple, then receive full-window temporal optimization. A temporal pass has RMS no greater than the nominal graph under the same inputs. Ancient contrasts (24,28) and the 3000–2920 BCE initial-enclosure interval within Stage 1 are evaluated separately and jointly after fitting. Role-by-role brightness is compared without fitted weights. A three-star ablation removes the latitude descriptor and third temporal channel. A flat-horizon calculation at 51.18° N tests whether all selected stars can be above the horizon during astronomical night, allowing season and sidereal time to vary (Supplementary Notes 44–48). This construction interval, summarized in Darvill et al.’s construction chronology [2], is the window labelled “Stage 1” in the archived filters; it is narrower than the full duration of that stage.
Crossed selection and a free timing scale
A further panel selects the five brightest catalogue entries in each of 88 modern constellation regions from the archived V≤6 table, breaking magnitude ties by HIP identifier. Each set receives all 120 temporal-role assignments and all 512 library words. The resulting 5,406,720 labelled descriptions reduce to 2,703,360 classes after exchanging the final unsigned pair. No spatial-tuple, chronology, visibility or shape filter precedes fitting. Full-window searches use 1-year and 0.5-year grids with local refinement and endpoints; the nominal threshold is recalculated from the same catalogue. These modern groupings delimit a reproducible panel, not an exhaustive set of prehistoric asterisms. The earlier unrestricted within-region stellar census is broader in star subsets but conditions on the spatial tuple (S5 Appendix, sections 1–3).
Two controls preserve the original material assumptions within this same bright-star panel. First, an independently implemented absolute-RMS search holds W=(28,72,16) fixed across all 5,280 role classes, using a 1-year grid, local refinement and endpoint evaluation. Second, the 512 words are classified by whether any strictly positive integer inventory generates them under the unchanged compiler. For W=(d,m,r), let q=m/3 and h=(d−r)/2. Feasibility requires integer q and h and B=q+h>0; then D=t>max(0,−h), C=t+h and A=d+q+t provide every positive integer solution. This is weaker than imposing all three original closure equations. Both controls were formulated retrospectively after examining the unrestricted panel; their constraints come from the original map, and they do not test alternative compilers (S5 Appendix, section 3.1).
A separate test retains W=(28,72,16) but removes its prescribed minute-per-integer timing scale. For temporal widths Tᵢ(t), set zᵢ=Tᵢ/Wᵢ and E(t)={mean[(zᵢ/mean(z)−1)²]}½. E is invariant under a common unit conversion; mean(z) is reported as the fitted common scale. This relative-dispersion objective changes channel weighting and is not the original absolute-RMS objective with its original ranks. It is evaluated for the primary and six-dimensional inputs and every bright-panel assignment. It frees only the temporal scale; the fixed-J2000 spatial tuple and the auxiliary degree-based solar relation remain separate conventions (S5 Appendix, section 4).
In an additional retrospective diagnostic, we retain W, the five stellar roles, timing scale, astrometry and equal channel weights, and minimize the same absolute RMS separately inside 3000–2920, 2980–2720, 2940–2800 and 2620–2480 BCE. The first three intervals were already used as archaeological comparisons; the last represents the Stage 2 construction interval. Each bounded search samples one-year intervals, refines bracketed local minima and evaluates both boundaries. We distinguish membership of the unconstrained optimum in an interval from a new optimization restricted to that interval. Restricted ranking is checked against the archived global minima of all 5,280 fixed-word classes. Error ratios are descriptive model contrasts, not likelihood ratios or chronological confidence levels (S5 Appendix, section 3.3).
A further check minimizes each absolute channel residual separately over 2620–2480 BCE on 1-year and 0.25-year grids, refines bracketed local minima and evaluates both boundaries. We compare these minima with the three residuals at the nominated early optimum. Componentwise inequalities then determine whether the phase preference depends on channel weights (S5 Appendix, section 3.4).
Spatially filtered stellar assignments
All seven temporal survivors in the principal V≤5 census lie in Cygnus. The expanded Hipparcos search supplies neither a named anchor nor Cygnus as input: 1,625 eligible stars across 77 regions yield 2,990 spatial candidates, of which seven reach or improve on the nominal temporal benchmark. Applying the ancient-contrast and Stage 1 filters jointly retains only Deneb–Sadr–Albireo–Aljanah–Fawaris in the declared roles. At V≤6, 1,072,375 spatial candidates yield 903 temporal passes and six joint passes, again all in Cygnus and all retaining Deneb and Sadr (Figure S9.5).

Figure S9.5. The joint stellar solutions concentrate in Cygnus and the nominal configuration is geometrically observable. Panel A reports the nested magnitude censuses after spatial and temporal filtering and after the additional joint contrast and Stage 1 filters. Panel B shows the full interval of magnitude limits for which the nominal quintet or ablated core is uniquely retained under its respective test. Panel C shows simultaneous stellar altitudes in an admissible astronomical-night configuration. The census uses uniform Hipparcos astrometry, and the night test allows season and sidereal time to vary.
The nominal quintet is brighter than every other joint survivor in at least one differing role and is never fainter in a corresponding role. Its role-ordered magnitudes are (1.25,2.23,3.05,2.48,2.86). It remains the unique assignment satisfying the unchanged joint criteria throughout the magnitude-limit interval 3.05≤V<5.06 (Supplementary Notes 45–46).
Removing the latitude descriptor and third temporal channel leaves the Deneb–Sadr–Albireo core as the only temporal survivor at V≤4. This core uniqueness spans limits 3.05≤V<4.59 before ancient-contrast or horizon filters. There are five temporal survivors at V≤5 and 131 at V≤6. The core minimum is near 2848.4 BCE, compared with 2965.4 BCE for the full graph in the uniform catalogue convention, showing the third channel's contribution to the full phase (Supplementary Note 47).
All five selected stars can be simultaneously above 63° during astronomical night at the nominal uniform-catalogue phase. The flat-horizon annual-night calculation finds this configuration while the solar centre is below −18°. The same rule excludes two V≤5 core competitors containing stars that never rise at the site at their fitted epochs. This establishes a geometrically possible observing configuration and a physical distinction among candidates (Supplementary Note 48).
Earlier anchored censuses, both historical membership lists and every alternative assignment remain in the supplement. The expanded comparisons use Hipparcos positions propagated from catalogue epoch J1991.25, with nominal minimum 2965.4 BCE and RMS 0.390 minute. These are common-input comparisons; the primary archived specification remains 2965.2 BCE and 0.419 minute (Supplementary Notes 18, 27, 34 and 44–48).
Fixed word and released assumptions
With W=(28,72,16) and its original timing scale fixed, nominal Cygnus ranks first among all 5,280 role classes in the declared five-brightest regional panel, without spatial-tuple or archaeological chronology filtering. Its RMS is 0.390 minute; runner-up Reticulum gives 0.554 minute at 2848.6 BCE, 1.42 times the Cygnus RMS. This is sole first place within that panel, not a pass against an independently specified absolute threshold or a search of all stellar subsets (S5 Appendix, section 3.1). Expressed directly in degrees, the same Cygnus minimum has RMS 0.097532° against θ=(7°,18°,4°). The first-place ranking and the 1.42 runner-up ratio are identical under the angular formulation.
The crossed bright-star panel yields 336 qualifying classes after exchanging the final pair, equivalent to 672 labelled role-word descriptions across 57 regions. The common-catalogue Cygnus benchmark is 0.390 sidereal minute near 2965.4 BCE. Six classes have their best qualifying word-specific minima in 3000–2920 BCE: four Cygnus descriptions and two first/third-role reversals in Lacerta. The Lacerta example has W=(32,32,28), RMS 0.138 minute and minimum near 2921.5 BCE. The best timing fit anywhere in the panel is a Scutum description with W=(16,4,4), RMS 0.019 minute near 3237.2 BCE (Table S9.3). These alternatives are numerical comparators, not proposed ancient identifications. Their words, however, cannot arise from the original integer compiler: neither 32 nor 4 in the second channel can equal 3q for integer q. They are alternatives under word freedom, not counterexamples under the unchanged material map.
Of the 512 words, 55 admit some positive integer inventory under that map, defining 290,400 role-word classes. Algebraic feasibility retains 26 of the 336 timing passes. Only nominal Cygnus has its best qualifying minimum in 3000–2920 BCE. In the broader 2980–2720 cal BCE 95% ditch interval [18], three feasible classes remain: Cygnus and two Vela assignments with words (32,72,68) and (40,72,68), minima at 2905.8 and 2884.3 BCE, and RMS 0.365 and 0.276 minute. Both Vela assignments lie within the 2940–2800 cal BCE 68% interval, which excludes the unconstrained Cygnus optimum. Vela's r=68 cannot satisfy r=D² for integer D, so these are not full three-closure alternatives. All three chronology windows are reported rather than choosing the one with the strongest selectivity.
The free-scale primary Cygnus solution occurs near 2986.8 BCE with relative dispersion E=0.682%; the six-dimensional variant gives 2987.5 BCE and E=0.682%. In the common-catalogue comparison, Cygnus gives E=0.649% near 2985.5 BCE. Thirty-eight classes improve that value, placing Cygnus 39th; the 39 matching-or-better classes span 25 regions; Draco gives E=0.0206% near 3052.0 BCE. Both grid resolutions recover identical qualifying candidate sets. The primary branch survives this scale change, but its exclusivity does not. Neither this panel nor the earlier separately conditioned censuses supplies a global chance probability or an equal-freedom comparison of complete archaeological models (S5 Appendix, sections 3–4).
Table S9.3. Comparison universes, fixed assumptions and released choices. All searches retain their stated stellar definitions and propagation model. Global optimization covers astronomical years −8000 to +2000. Absolute RMS and relative dispersion E are distinct objectives; their scores and ranks are not interchangeable. A “pass” uses the corresponding Cygnus benchmark, not an independent significance threshold. Interval membership after fitting and optimization within an interval are distinguished below.
Comparison and universe
Held fixed and varied
Objective and chronology
Result and location
S4 principal censusWithin-region subsets, V≤5
Fixed spatial tuple, W and timing scale; stellar identities and role assignments vary.
RMS; global epochs first. Ancient-contrast and initial-window filters applied afterward.
2,990 spatial candidates; seven temporal passes, all in Cygnus. Joint filters retain the nominal class (S9.5).
Fixed-word panel5,280 role classes
Fixed W=(28,72,16) and timing scale; roles vary among five brightest stars in each of 88 regions.
RMS; no spatial-tuple or archaeological chronology filter.
Cygnus rank 1: 0.390126 min, 2965.4 BCE. Reticulum: 0.554285 min, 1.42× (S9.5).
Integer-feasible words290,400 classes
Same bright-star panel and timing scale; 55 words vary under unchanged compiler feasibility, without all three closures.
RMS; global fitting, then interval membership. No spatial-tuple filter.
26 passes; one in 3000–2920 BCE and three in 2980–2720 BCE (S9.5).
Unrestricted library2,703,360 classes
Fixed bright-star panel and timing scale; roles and all 512 library words vary. Compiler feasibility not imposed.
RMS; global fitting, then interval membership. No spatial-tuple filter.
336 passes in 57 regions; six classes in 3000–2920 BCE (S9.5).
Free timing scale5,280 role classes
Fixed W and bright-star panel; role assignments vary and the common timing scale is fitted.
E; no spatial filter. Global search and separate interval-constrained searches.
Global: 39 classes match or improve Cygnus in 25 regions. Cygnus ranks 2, 2 and 4 in the three constrained windows (S9.5–S9.6).
Nominated-input sensitivityThree astrometric branches
Fixed Cygnus identities, roles and W; timing scale fitted. Primary, six-dimensional and common-catalogue inputs.
E; global search without archaeological chronology bounds.
Optima 2986.8, 2987.5 and 2985.5 BCE; E=0.6819%, 0.6817% and 0.6490%. These inputs do not rank a stellar panel.
Phase comparisonSpecified Cygnus model
Fixed word and roles. Original scale with channel-weight checks; separate tests fit scale under specified losses.
Global versus interval-constrained minima. Chronology explicitly bounds the latter.
Stage 2/original global RMS: 5.33× in the common catalogue. Free-E ratios: 15.31–16.06× across inputs (S9.6).
Metric-word extensionFive targets, same panel
Fixed construction and calibration coefficients; targets derive from all five declared exterior conventions. Roles vary.
Fixed-scale RMS and free-scale E, separately; global and constrained searches.
Fixed-scale global rank 1 in all five variants, with optima in 3000–2920 BCE. Free-scale global ranks 10–49 (S9.6).
Joint space and time5,280 classes; 10,560 typed assignments
Fixed S, W, reference and panel. Independent positive spatial and temporal scales; both final-pair spatial orientations.
Spatial CV or minimax with temporal E; no aggregate weight. Global and constrained comparisons.
No class dominates Cygnus. Global frontiers contain eight or six classes; no unique winner (S9.6).
The three constrained comparison windows are 3000–2920, 2980–2720 and 2940–2800 BCE, in that order; Stage 2 is 2620–2480 BCE. “Global” results do not use those intervals as fitting bounds. The complete panel definitions and adverse outcomes remain in S4 and S5.
S9.6 Chronology metric targets and adverse controls
Original target search and chronological comparisons
The inventory-derived word W=(28,72,16) ranks first by temporal fit among all 512 words in the declared material library, before applying an archaeological interval. Across the full 10,000-year search, its minimum occurs at 2965.2 BCE, with executed vector (27.785,72.653,16.233) sidereal minutes and RMS 0.419 minute, or 0.105°. The runner-up, (28,68,28), gives RMS 0.318° near 1744.1 BCE, approximately three times the primary RMS.
Three library words have minima in the declared 3000–2920 BCE initial-enclosure window; only W also has a fit at least as close as the primary result. The six-dimensional space-motion variant gives 2965.9 BCE. Main Figure 4 displays the complete search window and the archaeological comparisons (Supplementary Note 9).
The astronomical minimum falls within published radiocarbon-based intervals for the early monument, although those intervals do not enter the numerical minimization. It lies within Darvill et al.'s 3000–2920 cal BCE range for initial enclosure construction within Stage 1 [2] and Greaney et al.'s revised 2980–2720 cal BCE 95% ditch interval, but precedes the latter's 2940–2800 cal BCE 68% range [18]. These chronological models provide an external comparison with an early phase, rather than a date for the full input inventory. Any material system producing the same word under the specified operator would receive the same astronomical phase.
A solar comparison concerns the ideal cyclic construction. Its normalized long chord, 26 sin(21π/56)=24.0209, is close numerically to the 24.0163° obliquity at the stellar phase. Under the degree-identification convention, solar inversion gives 3016.7 BCE in the primary model and 2999.4 BCE in the Laskar sensitivity model. Of 416 disclosed cyclic arms, three intersect the obliquity range in the full window; only the selected arm gives a solution near the stellar phase (Supplementary Notes 10 and 25).
Phase comparisons
The fixed Cygnus model favours the early epoch over the later sarsen phase: the best fit in that phase has 5.33 times the RMS of the unrestricted minimum. In the common catalogue, the chronology-free optimum is 2965.4 BCE at 0.390 minute RMS. Restricting the same model to 2620–2480 BCE gives its lowest RMS at the oldest boundary, 2620 BCE: 2.080 minutes. The primary and six-dimensional inputs give corresponding ratios of 4.97 and 4.96 (Figure S9.6A). These input variants test astrometric sensitivity of one model, not independent archaeological confirmations.
The early epoch also outperforms Stage 2 channel by channel. At 2965.4 BCE, the common-catalogue absolute residuals are (0.211,0.605,0.213) minutes; the respective minima found anywhere in 2620–2480 BCE are (0.742,0.728,3.449), all at 2620 BCE. Every channel therefore favours the early reference date. Given these numerical interval bounds, any nonnegative channel weights with positive total preserve its lower weighted RMS, including when a channel is omitted. The primary and six-dimensional inputs give the same ordering. This establishes robustness of the phase comparison to weights, while allowing the weighted optimum to move to another epoch; it does not establish a new stellar ranking under changed weights (S5 Appendix, section 3.4). In the common catalogue, the narrowest margin is the Sadr–Albireo channel, 0.605 versus 0.728 minute; across all three input branches, the narrowest ratio is 1.040 in the six-dimensional branch.
The nominal optimum remains outside Greaney et al.’s 2940–2800 cal BCE 68% ditch interval. A separate constrained-epoch test gives Cygnus RMS 0.418 minute at its oldest boundary, 2940 BCE, only 7.20% above the unrestricted score (Figure S9.6B). Cygnus remains first among the same 5,280 fixed-word classes: every other class has an archived unrestricted minimum of at least 0.554 minute, and restricting its epoch cannot lower that minimum. Reticulum attains this bound at 2848.6 BCE, inside the interval, so it remains second with 1.33 times the constrained Cygnus RMS. This follows from the archived global scores and the new Cygnus bounded fit; it does not place the unrestricted optimum inside the narrower interval.

Figure S9.6. Temporal discrimination under the unchanged word and timing scale. Both panels use the common-catalogue absolute-RMS objective. (A) The unrestricted optimum falls in the initial-enclosure interval; the smallest error permitted in Stage 2 is 5.33 times the unrestricted minimum. (B) Restricting the epoch to the 68% ditch interval moves the Cygnus optimum to 2940 BCE while retaining a score below the best unrestricted score of every competing class. Reticulum’s minimum lies within that interval, fixing its second-place score. Shaded bands show archaeological intervals, not probability densities. These retrospective constraints do not replace the chronology-free search in main Figure 4.
The constrained ranking shows that an optimum can fall outside an archaeological interval while retaining comparative fit inside it. It does not relocate the original optimum or remove the free-word and free-scale alternatives in section S9.5 (S5 Appendix, section 3.3).
The preference for the early branch also survives fitting a common timing scale under the tested losses. Across the primary, six-dimensional and common-catalogue inputs, the best Stage 2 relative dispersion E is 15.31–16.06 times the global minimum. In the common catalogue, even the best Stage 2 epoch and minimax scale leave at least one channel discrepant by 11.06%, whereas an epoch near 2998 BCE permits all three channels to agree within 0.794%. These results compare phases within the specified Cygnus model. They do not preserve its first-place temporal ranking when scale is freed: interval-constrained free-scale searches rank it second, second and fourth in the initial-enclosure, 95% and 68% windows, respectively (S5, sections 9–10).
Continuous exterior measures
A metric-word diagnostic applies the existing constructive relations g=x, J=x+y, V=2z+x and r=z−x to the published exterior measures. It evaluates Wmetric=(2z−y,3y,z−x)/u, retaining u=(10x+24y+26z)/1352. Here x=33.435 m and y=80.095 m are the reported side means. Four previously declared Aubrey diameter conventions supply z=86.44,86.87,87.00 and 85.83168 m; a fifth branch uses the computed Euclidean diagonal √(x²+y²), not a measured diagonal. Each target receives the same fixed-scale RMS and free-scale E comparisons in the common catalogue. No branch is selected by its result. Target-change bounds applied to the archived global minima identify every class that could equal or improve the new Cygnus score; all eligible classes are refitted on 1-year and 0.5-year grids with local refinement and endpoints. The ranking therefore inherits the numerical scope of the archived minima (S5, section 12).
The fixed-scale selection does not require the target to equal the exact integer word. Applying the specified construction to continuous exterior measures leaves Cygnus first among the same 5,280 role classes in all five declared variants (Table S9.4). The principal Aubrey convention gives Wmetric=(27.852,72.127,15.911), RMS 0.323 minute and an optimum at 2975.9 BCE. The five global optima span 2992.6–2960.6 BCE, all within 3000–2920 BCE, without using that interval in the global ranking (S5, section 12).
Table S9.4. Fixed-scale stellar selection under all five continuous metric conventions. Every row uses the common catalogue and the original numerical timing convention. The Aubrey diameters serve as diagonal proxies in the declared construction; the Euclidean diagonal is calculated from the two reported side means. Rank is within the 5,280-class panel. The range across variants is a sensitivity, not a confidence interval.
Exterior convention
Optimum BCE
RMS min
Rank
Aubrey 86.44 m
2975.885
0.322947
1
Aubrey 86.87 m
2964.122
0.424256
1
Aubrey 87.00 m
2960.577
0.459085
1
Aubrey 85.83168 m
2992.625
0.257887
1
Computed diagonal 86.79348 m
2966.211
0.404460
1
The five variants share measured inputs and the original construction. Their free-scale global ranks range from 10 to 49. All five fixed-scale global optima lie outside the 2940–2800 BCE interval; restricting to that interval reduces the MacKie variant to second place. The result establishes stability under the examined metric changes at the original timing scale, not under every timing convention.
Joint spatial and temporal scales
A further diagnostic preserves S=(10,34,62) and W=(28,72,16) but permits separate positive spatial and temporal scales. It evaluates both assignments of the final temporal pair to the spatial latitude role: 10,560 typed assignments corresponding to 5,280 temporal classes. Spatial error is assessed separately by relative dispersion and by the minimum largest fractional residual; temporal error is E. A candidate dominates Cygnus only if it does not worsen either error and improves at least one. No weighted spatial–temporal sum or chronological filter enters the global comparison. All classes, sign conventions and complete frontiers are retained (S5, section 11). The additional phase audit also profiles a common timing scale under least squares and a minimax fractional loss. S5, sections 9–10, reports those sensitivities and the distinct consequences of imposing the original spatial constraints.
Cygnus remains nondominated in the joint spatial–temporal comparison when the two positive scales are fitted independently. Its spatial relative dispersion is 2.146% and its spatial minimax fractional error is 2.267%; its global temporal E is 0.649%. Cygnus belongs to frontiers of eight and six classes, respectively. It remains nondominated under both spatial criteria in each of the three previously examined archaeological windows (S5, section 11).
The spatial criterion alone ranks Cygnus second: a Vulpecula assignment improves spatial dispersion to 1.064% but has temporal E=4.075%. Conversely, candidates with lower temporal E worsen the spatial fit. This trade-off is the result of the joint comparison; nondominance does not identify a sole winner or supply a historical probability. The targets, stellar roles, J2000 reference and regional star selection remain fixed, and the diagnostic was formulated retrospectively after the separate rankings were known.
Source and model controls
The Aubrey-anchored assignment in the primary Cleal branch gives 8|21|7|20; exact registration with individual Aubrey positions remains unresolved. The continuous-gap and fitted-shape results above concern different observables. The tested eight-sector construction procedure and seven-hole residual model also fail to improve cross-validation over the regular-circle baseline (Supplementary Notes 20 and 30).
In the ancient mean ecliptic of date, the absolute stellar tuple is (11,35,63), and the unchanged inverse with r=16 returns (57,30,11,5). The rounded contrasts remain (24,28) at the tested phase. Fixed-J2000 stability therefore does not establish the same absolute register relative to the prehistoric ecliptic (Supplementary Notes 24 and 27). Changing the ideal cyclic solar input to the exact 5:12:13 ratio or published mean-side ratio gives 2784.8 or 2713.2 BCE, respectively (Supplementary Note 25).
Several physical realizations are source-sensitive or unsupported. The four-incentre assignment to the 30 digitized Y symbols has 2.87 m RMS, with 97.13% of tested rotations matching as well or better. The path's equivalent force line crosses Stone 1 in the primary branch and clears the entrance only with indicative Station 91. The best 16:30:34 point triangle has a predominantly north–south long side, contrary to the proposed horizontal-long-side assignment. The ditch centreline is derived from width summaries; alternative boundaries do not preserve its integer reading (Supplementary Notes 15, 37, 40 and 42).
Selection depends on the declared comparison universe. Releasing Deneb in the historical contrast-only census leaves two assignments common to both lists. Among arbitrary positive integer words, 179 fit the nominated sky graph as closely as W. Its rank within the 512-word material library conditions on that graph. S5 separately reports unrestricted words, words feasible under the original integer compiler, and the original word held fixed. The fixed-word comparison ranks Cygnus first across its bright-star role panel without an archaeological chronology filter. These observations concern different candidate universes; none supplies a global historical chance probability (Supplementary Notes 9, 44–46 and 52; S5 Appendix).
S9.7 Regional geometric and dimensionless comparisons
A supplementary diagnostic transfers the existing role selector from a separate stellar study [47, section 2.9] to the same five-brightest-star sets used here. Apparent V magnitude defines membership and brightness order. For each quintet, angular distances select the unique four-arm minimum spanning tree hub when present; otherwise the angular medoid is used. The brightest non-hub star defines the first endpoint, and the remaining star with the most opposite tangent bearing at the hub defines the third role. The remaining unordered pair supplies the third interval. The rule does not use W, temporal residuals, archaeological dates, the ecliptic tuple or a Galactic target. S6 specifies tie handling, the strict tree-margin test and the independent implementation.
The rule is applied to all 88 Hipparcos quintets at J1991.25 and J2000, and to all 88 archived HYG quintets. The selected Cygnus clock is then recomputed with the original word, scale and search window. Six archived kinematic scenarios test the common trio/quintet median, tree and role persistence at the existing clock epoch and at annual samples from 3000 to 1500 BCE. Catalogue membership and brightness ranks remain fixed. A subsequent scope check evaluates all 88 Hipparcos groups at three specified historical epochs. These are retrospective model checks; S6 records their order and complete results.
Dimensionless timing and a local horizon test
Let w=(7,18,4) denote the ratios of W=(28,72,16). For each of the 88 J2000 geometric assignments from section S9.7, let dᵢ(t) be the three absolute wrapped right-ascension differences relative to the equator of date. Define qᵢ(t)=dᵢ(t)/wᵢ and E(t)=sd(q₁,q₂,q₃)/mean(q₁,q₂,q₃), using population standard deviation. E vanishes exactly when the three intervals share the target proportion. Multiplying every interval by any positive unit-conversion factor, or every target component by a common positive factor, leaves E unchanged. The metric is the existing S5 free-scale relative dispersion, now applied to one geometrically selected assignment per group. It requires neither the fixed-J2000 ecliptic tuple nor a minute-per-count convention. The count map, regional membership, apparent-brightness selection, interval topology and relative-error loss remain specified model choices.
We search astronomical years −8000 to +2000 without archaeological chronology constraints, using 0.5-year and 1-year grids with refinement of every bracketed minimum and both endpoints. Every selected role is checked against S6 and every unrestricted score against S5. A common positive rescaling of the angular units or target is checked over all sampled configurations. Removing scale leaves two independent interval ratios; the fitted epoch is still a model output. This calculation is a recombination of existing components, not an independent replication of their discovery.
The local-observation branch imposes the same necessary horizon condition on every group. At latitude φ=51.18° N, a star of declination δ has geometric upper culmination h=90°−|φ−δ|. All five stars must have h>0 at the candidate epoch. We identify every admissible epoch segment on both grids, refine its boundary crossings and reoptimize E within those segments, including boundary infima. Thus a group is not discarded merely because its unrestricted optimum is invisible. This condition is necessary for observing the complete configuration locally; it does not model darkness, extinction, terrain or transmission of knowledge from another latitude. We separately test simultaneous altitude feasibility for Cygnus and Tucana at their unrestricted optima. S7 records the complete search, input provenance and numerical scope.
Cygnus is the only four-arm star-shaped minimum spanning tree in each 88-quintet catalogue panel. The geometric rule selects Deneb–Sadr–Albireo and leaves Aljanah–Fawaris as the remaining pair, recovering the existing temporal roles. At J2000 in the common Hipparcos catalogue, Albireo makes a 168.702° bearing angle from Deneb at Sadr, compared with 107.449° for Aljanah and 84.895° for Fawaris. No selected role changes between J1991.25 and J2000 in the 88 Hipparcos groups. Recomputing the clock with the recovered roles gives 2965.425 BCE and 0.390126 sidereal minutes RMS, reproducing the common-catalogue result in section S9.5. The original 5,280-class comparison and its 1.42 runner-up ratio remain the fixed-unit temporal benchmark.
At that epoch, all six kinematic scenarios retain Sadr as a certified equal-weight spherical median of both quintet and selected trio, the four-arm tree and the same role assignment. The proportional change of the trio’s angular shape relative to J2000 is 0.0394–0.0472%. Across 1,501 annual epochs per scenario from 3000 to 1500 BCE, all 9,006 sampled model cases preserve both median certificates, the tree and the selected continuation; the minimum tree margin is 0.3990°. The additional 88-group checks at 3000 BCE, the existing optimum and 1500 BCE also leave Cygnus as the sole four-arm case. S6 gives the certificates, scenario definitions, complete comparison panels and secondary Galactic-axis diagnostic.
With one geometrically selected assignment per group and no horizon or chronology restriction, Cygnus ranks second among 88 groups. Its minimum is E=0.649026% near 2985.482 BCE. Tucana ranks first at E=0.134236% near 3237.025 BCE, followed by Norma at 2.801116% near 7314.476 BCE. The Cygnus right-ascension intervals are (6.954903°,18.170633°,4.006198°); their common numerical scale is discarded. Every one of the 88 unrestricted scores reproduces its S5 free-scale value, and both search-grid resolutions recover the same minima. Table S9.5 distinguishes this panel from the unrestricted 5,280-role comparison in section S9.5.
Table S9.5. Leading dimensionless fits after the same geometric role rule. The last column reports the best score after reoptimization over epochs satisfying the necessary horizon condition. Scores are percentages; smaller is better. The search uses no archaeological date filter.
Group
Global E (%)
Global epoch (BCE)
Horizon E (%)
Tucana
0.134236
3237.025
None
Cygnus
0.649026
2985.482
0.649026
Norma
2.801116
7314.476
2.801116
Aquila
4.360427
2496.611
4.360427
Draco
4.623273
1299.828
4.623273
Seventy-one groups have at least one admissible epoch under the necessary horizon condition. Cygnus ranks first after reoptimization within all such epoch segments; Norma is second. Tucana has no admissible segment. Four of its five stars remain below the geometric horizon throughout the numerical search, with refined maximum upper culminations of −7.336°, −15.700°, −13.178° and −3.835°. At its best temporal fit, all five stars have negative upper culminations. Cygnus, in contrast, permits all five stars simultaneously above 64.719° near its optimum, at an appropriate sidereal time. The local-observation ranking therefore follows a common physical requirement rather than a chronology selection.
The dimensionless Cygnus optimum lies within the 3000–2920 BCE initial-enclosure interval, although it is just older than the separate 2980–2720 cal BCE 95% ditch interval. Within the later 2620–2480 BCE sarsen phase, its smallest E is 10.422720%, 16.059 times the global minimum. This reproduces the S5 free-scale phase contrast. The 2985.482 BCE optimum and the earlier fixed-unit 2965.425 BCE optimum are outputs of different objectives, not independent dates or uncertainty bounds. Across the full 5,280-role panel, free-scale alternatives remain as reported in section S9.5; the geometric selector specifies which one role assignment per group is retained here. The model still requires archaeological justification of the count map and any proposed transmission across phases.
S9.8 Information order execution and navigation
The public archive dated 26 August 2026 (Zenodo 22103786), with its Calanais extracts and checksum ledger retained in S5, records the stellar grammar, role order, rounding rule, W and temporal operator. The Stonehenge map, dimensional ladder and subsequent diagnostics were developed with relevant data available. Archaeological dates do not constrain the original optimization or fixed-word ranking; explicitly labelled phase tests use them. S5 fixed-word and integrality checks followed its unrestricted panel, followed by phase and robustness diagnostics. S6 transferred the geometric selector after both stellar and Stonehenge results were known. S7 combined geometric roles, ratios and uniform horizon reoptimization after preliminary rankings and Tucana’s horizon failure were known. S8 invariance and loss checks preceded the spherical axis audit, then the field census and kinematic recensus. The complete fixed-scale census and its matching free-scale comparison were executed after those four-arm results were known. Both field definitions and the angular target remain retrospective choices in this application. Radius calibration was formulated after the fixed-scale and free-scale census results were known; its two radius conventions and subsequent catalogue/state checks are additional retrospective analyses. These are retrospective tests, not preregistered discoveries. Supplementary Notes 1, 14 and 50–54 and the S5–S8 protocols retain their execution order and adverse outcomes, including rank losses and unsuccessful inverse constructions.
OpenAI ChatGPT assisted with literature retrieval, drafting, code construction and the additional S5 computational audits in September 2026; its exact model version was not exposed in the authors’ working record. It also assisted with the October 2026 manuscript revision, S6–S10 computational diagnostics and reproduction checks. The code was executed against archived inputs, with checks and outputs supplied in the accompanying package. No archaeological survey or catalogue observation was generated by the service. Scientific interpretation and responsibility remain with the authors.
The complete fixed-scale and radius-calibrated field results are supplied in S8.9 and S8.10. The broader contextual report on Bulford [45] is retained in S1 as source context; it is not a numerical input to the Stonehenge model. Revision v1.41 reorganizes the narrative and relocates extended analyses. It introduces no new numerical search or change to the archived candidate outputs.
Portable code, pinned inputs and complete numerical outputs support reuse [46]. The S5 README preserves the earlier execution history. For revision v1.35, the outer-scale calculation, local perturbation control and conditional integer audit were rerun, and the S5 checks recomputed reported temporal scores and verified archived ranks and robustness summaries. No new data or new full S2–S4 execution is claimed. Revision v1.36 adds the executed S6 diagnostic and its reproduction and integrity checks; the earlier S1–S5 archives and figure files are preserved unchanged. Revision v1.37 executes S7 and integrates its dimensionless comparison. All 88 roles and unrestricted scores are checked against S6 and S5, and both horizon-constrained and unrestricted minima agree across the two search grids. The earlier S1–S6 and figure archives remain unchanged. Revision v1.38 integrates the subsequently executed invariance, axis and field audits as S8. Their original numerical files and clean-extraction reproduction records are retained; all 26 field-census output files reproduced byte for byte before integration. Revision v1.39 independently enumerates the complete 19,939-group Hipparcos J2000 family and executes fixed-scale, free-scale and fixed-scale horizon searches on 1-year and 0.5-year grids. The earlier 03_Fields module was also rerun from clean extraction, reproducing all 26 outputs byte for byte. S1–S7 and figures remain unchanged. The packaging checks document claims and file integrity; no new full execution of S1–S7 is claimed. Revision v1.40 integrates the executed radius-calibration audit and refits all four-arm candidates in four nominal and six targeted state panels under both radius conventions, with unrestricted and horizon-constrained searches on both grids. It checks candidate membership and roles, cross-grid scores, vector expressions and original-file integrity. The earlier modules and S1–S7 remain unchanged. Revision v1.42 adds the common-epoch dominance inference from archived complete-census minima and new Cygnus evaluations, including agreement across both grids and an independent vector check. It restores the dated public antecedent, separates the phase comparison in main 3.4, and updates the navigation of retained appendices. No new complete rival optimization or archaeological observation is claimed.
Revision v1.43 integrates the subsequent complete-pattern and unit-independent inverse analyses as S10 and main section 3.7. The added midpoint relation followed inspection of Cygnus. Both new complete censuses retain the archived field family and roles. S10 also reports the inconclusive construction-error diagnostic. The earlier S1–S8 numerical results and common-epoch dominance scope remain unchanged; the new objective is not added to that dominance proposition.
Revision v1.44 adds the matched free-midpoint/free-scale ablation and the equal-opportunity search over the previously archived 66 signed channel/divisor descriptions. These retrospective controls are reported in S10.3 and main section 3.7. Their ranks are 119 and 16 in the complete family; the 66-description comparison retains first place within the existing 283-field four-arm subset. Earlier numerical outputs are preserved.

Figure S9.7. Dependencies of the original fixed-unit analysis. The exterior scale, count map, stellar search and chronological comparison occupy distinct steps. The 5,280-class panel shown here is the original regional comparison; the complete 19,939-field census and radius-calibrated extension are documented in S8.9–S8.10 and main Table 2.
Crosswalk for retained supplementary references
Earlier supplements retain their original numbering and numerical records. Use the following crosswalk when an earlier file cites a main-text location. Reference numbers in the shared bibliography are unchanged. The main figures and the seven S9 figures together preserve all twelve earlier figures; S8 contains the detailed field tables. The crosswalk covers the earlier numbering through v1.40; v1.41 main 3.4 and 3.5 are now main 3.5 and 3.6. The new phase subsection is main 3.4, and the dominance proposition is S8.11.
Table S9.6. Navigation from the expanded v1.40 numbering.
Earlier location
Current location
Methods 2.1–2.2
Main 2.1; S9.1 and S9.4
Methods 2.3–2.6
Main 2.2–2.3; S9.2, S9.5–S9.6
Methods 2.7–2.10
Main 2.2–2.5; S9.7–S9.8; S6–S8
Results 3.1–3.3
Main 3.1–3.2; S9.2–S9.3
Results 3.4–3.5
Main 3.3–3.4 and 4.1; S9.3 and S9.6
Results 3.6–3.7
Main 3.6; S9.4
Results 3.8–3.13
Main 3.4 and 3.6; S9.5–S9.6
Results 3.14–3.15
S9.7; S6–S7
Results 3.16
Main 3.3 and Table 2; S8.1–S8.3 and S8.9
Results 3.17
Main 3.6; S8.4–S8.7
Results 3.18
Main 3.5 and Table 2; S8.10–S8.11
Tables 1, 3, 4, 5, 6
S9.1, S9.2, S9.3, S9.4, S9.5 respectively
Table 2
Main Table 1
Tables 7–9
Main Table 2 summarizes; S8.8, S8.1 and S8.9 retain detail
Figures 1, 3, 5, 6, 9
Main Figures 1–5 respectively
Figures 2, 4, 7, 8, 10, 11, 12
Figures S9.1–S9.7 respectively
