Geometry Before Language

1.Introduction: An Uncomfortable Observation

The Chronometer was not built to prove anything.

It was built because I wanted a browser-native orrery that did not phone home to an ephemeris server — a single file, privacy-by-design, computing Kepler's equation and perturbation series locally. The astrological layer came later, and the language layer later still.

The uncomfortable observation arrived unbidden.

Given a chart and a constraint to reason only from computed geometry, the system produced interpretations that were internally consistent across mythological, theological and biographical registers — and, more surprisingly, that refused to produce interpretations when the geometry did not support them. It declined to infer biography from aspect. It repeatedly distinguished what it had calculated from what it was suggesting.

That refusal is more interesting than any individual reading.

This paper is an attempt to take the observation seriously without overclaiming from it, and — equally — without performing the ritual dismissal the topic usually attracts. Both failures are available. Both are boring.

2.The Evidential Situation, Stated Honestly

2.1 The null results are real

Any paper approaching this subject that does not front-load the negative literature is not worth reading. So:

Shawn Carlson's double-blind test, published in Nature in 1985, asked astrologers to match natal charts to California Personality Inventory profiles under conditions designed in consultation with astrologers themselves. Carlson reported no effect (Carlson, 1985).

David Voas analysed the astrological pairings of ten million married couples from the 2001 census of England and Wales — the largest test of astrological compatibility ever conducted. He found no relationship between sun-sign combination and the probability of marrying, or of staying married (Voas, 2008).

Geoffrey Dean and Ivan Kelly examined over two thousand "time twins" born in London within days of one another in March 1958, tracked across more than a hundred measured characteristics. They reported that the test conditions were maximally favourable and the results uniformly negative (Dean & Kelly, 2003).

The Comité Français pour l'Étude des Phénomènes Paranormaux attempted to replicate Michel Gauquelin's "Mars effect" — the reported excess of Mars in key sectors at the births of eminent athletes. Of 1,120 champions, 209 (18.66%) fell in a key sector, against a computed base rate of 17.70%. The Committee concluded that no effect had been demonstrated, and attributed Gauquelin's original finding to selection bias in his sampling (Benski et al., 1996).

Dean, Mather, Nias and Smit's later survey of hundreds of studies reaches a broadly negative overall verdict (Dean et al., 2016).

This is not a small pile. It should not be waved away, and this paper does not wave it away.

2.2 Some of them are contested, and the contest matters

What is less often reported is that the most famous of these results has been substantially disputed on statistical grounds — and disputed by someone with no stake in astrology's success.

Suitbert Ertel, emeritus professor of psychology at Göttingen and a critic of sloppy work on both sides of the question, reanalysed Carlson's published data. He argued that the study's design was unfair to its participants and that its analysis violated ordinary statistical norms: it partitioned the sample into sub-groups rather than testing totals, and neglected power, effect size and sample size. On reanalysis of the aggregate, astrologers matched CPI profiles to natal charts above chance at marginal significance (three-way forced choice, p = .054), with the positive result replicating under a different assessment method (ten-point rating, p = .04) (Ertel, 2009).

Ertel's own conclusion is the model of the register this paper wants to occupy, and it is worth stating precisely: the results are insufficient to regard astrology as empirically verified, but sufficient to regard Carlson's negative verdict as untenable.

That is not a vindication. It is a demotion of a supposedly decisive result to an inconclusive one.

Kenneth McRitchie has argued similarly against the Dean and Kelly time-twins study — principally that the test never defined its criteria of resemblance, and that predicting psychological test profiles does not distinguish permanent personality dimensions from transient states, which is precisely the distinction astrological claims turn on (McRitchie, 2016).

I do not know who is right in these disputes. I have read the reanalyses and the rebuttals and I am not equipped to adjudicate them. That admission is itself the point: a field where the flagship negative result has been credibly challenged, and where the challenge has been met with silence rather than re-testing, is not a settled field. It is an abandoned one.

2.3 What has been tested, and what has not

Look at what those studies actually operationalised. Sun-sign compatibility. Chart-to-inventory matching. Personality-factor similarity between time twins. Gross positional frequencies in a single sector.

Now look at what a practising astrologer actually claims to use: applying versus separating aspects; orb tightness as a weighting on relational strength; angularity; the whole weighted graph of mutual relation, evaluated in temporal motion.

The gap is not subtle. The tested hypotheses have been overwhelmingly the ones easiest to encode in a spreadsheet, not the ones anyone practises. Sun sign is to natal astrology roughly what "does the patient look ill" is to differential diagnosis — a real signal in principle, a strawman in practice.

I want to be careful here, because this is exactly where special pleading lives. The astrological literature has a long and disreputable habit of retreating into complexity whenever a simple claim fails: ah, but you need the whole chart. Voas anticipated this and answered it well — with ten million couples, even a one-in-a-thousand effect would show.

So the argument cannot be "the tests were too simple." It has to be sharper than that.

Here is the sharp version. Voas's sample size defeats the retreat-to-complexity move for sun signs specifically. It does not test aspect geometry, because sun-sign combination is not a proxy for aspect geometry — it is a coarse projection of one variable pair onto a twelvefold grid, discarding orb, discarding applying/separating direction, discarding every other body. A null on the projection does not entail a null on the structure. This is an ordinary point about information loss under projection, not a metaphysical escape hatch.

The honest summary, then, is not "astrology has been disproved." Nor is it "astrology has been vindicated." It is: the simple versions have been tested and have failed; the structural version has never been tested at scale, because until recently nobody could compute it, blind it, and control it cheaply enough to try.

3.Mechanism Is Not the Bottleneck

The standard objection is not really evidential. It is prior. It runs: there is no known physical mechanism by which planetary position could influence human affairs, therefore the correlation cannot exist and need not be examined.

This is the weakest link in the sceptical chain, and it is weak for a reason the history of science makes embarrassingly plain.

Alfred Wegener proposed continental drift in 1912 and elaborated it in 1915 on the basis of a striking correlation — coastline fit, matching fossil assemblages, continuous geological strata across oceans. He was ridiculed for half a century, and the central objection was always the same: no plausible mechanism could move continents. The correlation was real the entire time. Plate tectonics arrived in the 1960s and supplied the mechanism. The evidence had been waiting fifty years for an explanation it did not require in order to be true (Wegener, 1915).

Ignaz Semmelweis observed in the 1840s that puerperal fever mortality collapsed when physicians washed their hands in chlorinated lime. He had a robust correlation, a clean intervention, and no theory whatsoever — germ theory was decades away. He was dismissed, and he died in an asylum. The correlation was real (Semmelweis, 1861).

The pattern is consistent: absence of mechanism has never been a reliable guide to absence of phenomenon. It has been, repeatedly, a guide to the limits of contemporary theory.

But I want to be exact about the disanalogy, because a hostile reader will find it immediately and they will be partly right.

Wegener and Semmelweis both had a replicated correlation begging for explanation. Astrology's difficulty is more severe: what is missing is not only the mechanism but a robustly replicated effect for a mechanism to explain. That is a real and material difference and I am not going to paper over it.

What follows from it, though, is not what sceptics assume. It does not follow that the question is closed. It follows that the question is at an earlier stage than either side pretends — the stage of establishing whether a correlation exists at all, under conditions that would actually detect one. Astrology has been treated as a hypothesis awaiting a mechanism when it is in fact a hypothesis awaiting a measurement.

And so: a mechanism unknown is not a mechanism unknowable. If a structural effect were found, candidate mechanisms would not be scarce — one would look first to whether chart geometry inadvertently encodes seasonal, photoperiodic, geomagnetic or gestational variables that correlate with birth timing, none of which require any force acting at a distance and none of which are exotic. The claim of this paper is deliberately smaller than "the planets do something." It is that we do not currently know whether the relational structure carries information, and that this is a determinable fact rather than a permanent mystery.

4.The Chronometer Architecture

Nothing mystical happens in the computational layer.

Orbital elements. Kepler's equation. Perturbation series. Coordinate transforms. Aspect geometry as angular relation, with orb as a continuous weight rather than a binary threshold. Applying and separating states derived from relative angular velocity.

Everything is arithmetic, and it is deterministic. The same input produces the same graph, every time, on any machine.

Only after this stage does a language model enter — and it enters constrained. It cannot introduce a body that was not computed. It cannot introduce an aspect that was not measured. It is instructed, structurally, to distinguish computation from interpretation. It is not asked open questions.

The difference between these two prompts is the whole architecture:

"Tell me about Saturn."

"Saturn is applying to a square with the natal Moon, orb 0.6°, separating from a trine to Mercury. Interpret these two relations and nothing else."

The first invites improvisation. The second permits only description.

Geometry first. Language second. Relationship first, narrative afterwards. The language no longer invents the world it describes — even when that world is symbolic.

5.The Jesus Case: Illustration, Not Evidence

A chart was cast for the traditional life of Jesus of Nazareth.

Let me disqualify it as evidence immediately and thoroughly, because if I do not, a reviewer will, and they will be right.

The ephemerides degrade over two millennia. The traditional dates are theological rather than historical. There was no blinding, no control chart, no pre-registration, no rater protocol, and the sample size is one. The astrologers who described the output as coherent knew what they were looking at. Every methodological safeguard that matters was absent.

Worse, and more interestingly: coherence is exactly what the deflationary hypothesis of §6 predicts. A sufficiently dense symbolic graph will produce internally consistent readings whether or not it corresponds to anything at all. Coherence therefore cannot be evidence, because it is equally consistent with the hypothesis that astrology is empty. Using it as evidence would be a straightforward begging of the question.

What the case is good for is narrower and still worth something: it demonstrated that the architecture behaves as designed. The model refused to fabricate biography from geometry. It declined to infer. It drew parallels only when explicitly invited to compare the computed structure against a known figure, and it marked those parallels as interpretive when it did.

That is a systems result, not an astrological one. It shows the constraint holds under pressure — which is the precondition for running a real test, not a substitute for one.

6.The Deflationary Account, Stated at Full Strength

Intellectual honesty requires putting the opposing case as well as its proponents would, so here it is, and it is strong.

Symbolic compression. Astrology is not a list of meanings; it is an extraordinarily dense symbolic graph. Some two thousand years of human reflection have been attached to a small finite set of nodes — planets, signs, houses, angles, aspects — with weighted, directional, temporally-evolving edges between them. Language models are exceptionally good at expanding compressed symbolic representations. On this account astrology survives not because planets influence psychology, but because it became an unusually efficient symbolic indexing system for human experience. The Chronometer's outputs would then be genuinely impressive artefacts of graph density, and would tell us nothing whatever about the sky.

Constraint improves apparent intelligence. Unconstrained models confabulate. Sufficiently constrained ones become disciplined. Much of what reads as insight in the Chronometer's output may simply be the discipline of a narrowed question. This is a real effect and it flatters the system.

The Forer effect. Bertram Forer gave thirty-nine students an identical generic personality sketch, assembled from horoscopes, presented as an individual analysis. Nearly all rated it as highly accurate (Forer, 1949). Any evaluation depending on subjective recognition of fit is contaminated by this, and astrological validation is almost entirely such evaluation. Note carefully that this contaminates the astrologers' reaction to the Jesus output just as much as it contaminates a consultation.

If I am honest about the current state of the evidence, the deflationary account is the one I would bet on. It explains everything observed so far, requires nothing new, and has the better track record.

The problem is that it also explains everything observed if astrology were true. It is unfalsifiable as currently deployed. That is not a virtue in a scientific position, and it is the reason §8 exists.

7.Geometric Symbolic Reasoning

Strip the astrology out and something survives. Underneath the planets sits a graph: nodes, edges, weights, directionality, temporal evolution — precisely the structures over which contemporary machine reasoning increasingly operates. The formalism below is deliberately lightweight. Its purpose is not to prove a theorem but to make plain that what is being described is graph traversal under constraint, not an emergent property of a language model.

7.1 Definitions

Let the relational geometry at time t be a directed, weighted graph

V

A finite set of entities. In the astrological instance: the bodies and computed angles. |V| is fixed and does not vary with t.

E(t) ⊆ V × V

The set of relations holding at t. Directed, because the relation of v_i to v_j is not in general symmetric. In the instance: aspects.

W(t) : E(t) → ℝ^k

A weight function assigning each edge a vector of scalar attributes. In the instance: orb (angular deviation from exactness), and the signed rate of change dθ/dt, whose sign distinguishes applying from separating.

Critically, G(t) is the output of a deterministic function of ephemeris time:

where 𝒢 is pure arithmetic — Kepler's equation, perturbation series, coordinate transforms. Same t, same G, on any machine, with no model in the loop. This is the first and most important property: 𝒢 is reproducible by anyone, and auditable without reference to the language layer at all.

Symbolic binding. Let Σ be a symbolic lexicon — the accumulated meanings a tradition attaches to entities and relations. Let

where λ_V : V → Σ annotates entities and λ_E : E × W → Σ annotates relations, conditioned on their weights. Writing π for the projection that discards annotations, Φ is structure-preserving by construction:

Φ adds labels. It never adds, removes, or reweights an edge. This is the first invariant.

The language layer. Let L be the constrained language model,

mapping annotated graphs to strings. The entire content of the architecture is the admissibility condition on L. Let ρ(s) denote the set of relational assertions extractable from an output s. Then L is admissible if and only if

Every relation asserted in the output must correspond to an edge that was computed. The model may traverse the graph, describe it, modulate its narration by W, and select among paths. It may not add an edge.

The full pipeline is the composition

read right to left: compute, bind, describe.

7.2 The inversion

The standard generative pipeline runs tokens → embeddings → structure inferred implicitly in latent space → output. Structure is a consequence of language, and it is never explicitly represented.

GSR runs structure computed → structure annotated → language emitted under structural constraint. Language is a consequence of structure.

The practical difference concerns hallucination. In the standard pipeline, a fabricated relation is a failure of a model that was never given a ground truth to violate; there is nothing to check it against. Under condition (2), the violation ρ(s) ⊈ E(t) is a well-defined, mechanically decidable condition — one can enumerate the asserted relations and test membership in a set that was computed independently.

7.3 What the formalism does and does not guarantee

It does not guarantee correctness, and I want to be blunt about this because the notation could easily be read as claiming more than it does.

L remains a stochastic function. Condition (2) is enforced by construction — through prompt scoping and post-hoc validation of ρ(s) against E(t) — not by mathematical necessity. Nothing here proves that L cannot violate it. What the formalism buys is that violations become well-defined and checkable, which is not true of ungrounded generation. That is a difference in kind, not a proof of correctness, and it should not be sold as one.

Second, and more importantly: every contestable assumption in this paper lives in Φ. The choice of lexicon Σ and of the annotation maps λ is a modelling decision carrying the whole weight of the symbolic tradition. The determinism of 𝒢 confers no legitimacy whatsoever on Φ. A perfectly computed graph can be bound to an entirely fictitious lexicon and the pipeline will run exactly as well. Distinguishing these two cases is precisely what §8 is for.

7.4 Relation to existing architectures

GSR is adjacent to, but distinct from, retrieval-augmented generation (Lewis et al., 2020). RAG grounds a model in retrieved documents — unstructured text from which the model must still infer relations. The inferential step remains inside the model's authority, which is why RAG reduces but does not eliminate fabrication. GSR grounds the model in a computed structure, where the relations are given and the only permitted move is traversal. The inferential step has been moved outside the model entirely.

7.5 The symbolic layer is a parameter

Nothing in (1) or (2) requires astrology. Σ and Φ are free parameters; 𝒢 is domain-specific but always deterministic. Substitute the symbolic layer and the engine is unchanged:

Greek gods. Tarot. Chess positions. Constitutional law. Protein interaction networks. Characters in Shakespeare. Economic indicators. Musical motifs. Family systems. Ecological webs. Differential diagnosis.

Astrology's contribution, on this reading, is not truth but maturity of dataset — an unusually well-developed, internally consistent Σ refined over millennia and available at no cost. Whether the sky has anything to do with it is, for the architecture, entirely optional.

8.The Test

This is the section the earlier draft of this paper lacked, and without which the rest is rhetoric.

The separation of 𝒢 from L makes something possible that was previously impractical: generating interpretations at scale, blind, with matched controls, without an astrologer in the loop whose expectations could contaminate the output.

8.1 The hypotheses, formally

Let σ be a permutation on the edge set that preserves the corpus marginals — so that seasonal and diurnal birth-frequency artefacts are held constant — but randomises the relational structure. Write the scrambled graph G^σ. Let B be a subject's biography and m(·,·) a blind matching accuracy.

The symbolic-compression hypothesis of §6 states that all informativeness resides in Σ and Φ, and none in E:

The astrological hypothesis is the strict inequality:

Stated this way, the question stops being "is astrology true" — an unanswerable framing — and becomes something a statistician can act on: does the edge set carry information beyond the lexicon?

8.2 Protocol

1. Corpus. Assemble n ≥ 500 biographical subjects with birth times recorded to the minute from official registration — not self-report. Registration-derived timing is essential, since post-1950 self-reported birth times are a known error source in this literature.
2. Charts. Compute G for each subject, and a matched scrambled control G^σ as defined above.
3. Interpretation. L(Φ(·)) generates an interpretation from each graph, real and scrambled, under identical constraints. The system is never told which is which. No human sees the charts.
4. Blinding. Interpretations are stripped of any content identifying the graph type. Raters — a mix of practising astrologers and naive controls — receive interpretation/biography pairs with no indication of provenance.
5. Task. Forced-choice matching: given one biography and two interpretations, one from G and one from G^σ, select the better fit. Continuous fit ratings collected alongside, for effect-size estimation.
6. Pre-registration. Hypotheses, analysis plan, exclusion criteria and stopping rule registered before data collection. Full corpus and code published regardless of outcome.
7. Analysis. Test the aggregate, not sub-samples. Report effect size and power alongside significance. This is the specific failure Ertel identified in Carlson, and it must not be repeated in either direction.

8.3 Secondary analyses

On the same corpus, isolating the structural claims that have never been tested separately. Each asks whether a component of W carries information:

• Do tight orbs outperform loose orbs? Theory says yes; a null model says orb width is irrelevant.
• Do applying aspects outperform separating ones? Theory predicts asymmetry in the sign of dθ/dt; chance predicts none.
• Does angularity carry weight beyond aspect?

These are the interesting tests, because they are internal. They do not ask whether astrology works — they ask whether the system's own claimed weighting scheme carries information. A positive on the primary with nulls on all three would be a substantially weaker result, and would suggest something other than the claimed mechanism is doing the work.

8.4 The falsification condition

If interpretations generated from G^σ are matched to biographies at the same rate as those from G — that is, if (3) holds — the symbolic-compression hypothesis is confirmed and the astrological hypothesis is disconfirmed for this class of test. I commit in advance to reporting that outcome as a refutation of the position argued in §2 and §3, without reinterpretation.

I expect that outcome. I think it is more likely than not.

I would rather know than be right.

9.What Each Outcome Would Mean

If (3) holds. Astrology's coherence is fully explained by the density of Σ and the Forer effect. The astrological hypothesis takes a serious and properly designed hit — the first one, arguably, aimed at the right target. GSR survives entirely intact, because condition (2) never depended on Σ being true. The architectural claim is unaffected; the empirical claim is dead.

If (4) holds. We have a correlation without a mechanism, in a domain where that is enormously surprising. The correct response is not celebration but replication — independently, by hostile parties, at greater n. Only after replication does the mechanism question become live, and the first places to look are mundane: seasonal, photoperiodic, geomagnetic and gestational variables correlated with birth timing that E may be inadvertently encoding.

If results are equivocal. The most likely outcome, and the least satisfying. It would still be an advance, because it would be equivocation produced by a test aimed at the actual hypothesis rather than at a convenient proxy for it.

Note that only the first outcome is a loss for the position argued here — and it is a loss specified in advance, in falsifiable terms. That is the difference between this draft and the one before it.

10.Conclusion

This paper does not claim astrology has been validated. It has not been.

It claims something narrower and, I think, harder to dismiss: that astrology has not been properly tested; that the absence of a mechanism is a poor reason not to test it; and that the architecture described here makes the real hypothesis testable for the first time, at a cost that no longer requires anyone's permission or funding.

Alongside that sits a claim I hold with much greater confidence and which does not depend on the first at all. Geometric Symbolic Reasoning is a real architectural pattern with obvious applications wherever complex relational systems exist — diagnosis, systems engineering, legal reasoning, ecology, comparative religion, education. The important innovation is not astrology.

The important innovation is refusing to let language invent structure.

Human civilisation has spent millennia constructing symbolic systems. Perhaps their lasting value was never predictive. Perhaps it was always computational — myth, religion, mathematics and story not as competing accounts of reality but as different projections of the same relational landscape.

The Chronometer may simply have stumbled into a better way of letting machines walk across that landscape.

Whether the landscape corresponds to anything in the sky is a separate question, and an open one, and — this is the whole point — an answerable one.

Postscript

It would be wonderfully ironic if humanity spent centuries arguing about whether astrology was true, only to discover that its greatest contribution was not revealing the future but accidentally assembling one of the oldest and most sophisticated symbolic graph databases ever built.

It would be considerably more interesting if we ran the test and found we had been arguing about the wrong question for four hundred years.

Either way, the argument has gone on long enough without anyone bothering to look.

References

this paper was prepared with the aid of Claude Opus 5

Correspondence: toby@ffilm.org