---
title: "Reconstructing Ancient Conic Sundials from 3D Scans"
subtitle: "A Research Programme in Construction Geometry, Measurement, and Literate Coding"
running-title: "Sundial Reconstruction Programme"
book: false
toc: true
author:
  - name: Gerd Graßhoff
    email: gerd.grasshoff@hu-berlin.de
    orcid: 0000-0002-4450-357X
    affiliations:
      - name: Humboldt-Universität zu Berlin
        department: Philosophisches Institut
        address: Unter den Linden 6
        city: Berlin
        country: Germany
        postal-code: 10099
      - name: Deutsches Museum
        department: Forschungsinstitut
        address: Museumsinsel 1
        city: München
        country: Germany
        postal-code: 80538
abstract: |
  The Berlin Sundial Project preserves photogrammetric 3D surface models of
  ancient Greek and Roman conic sundials. The overarching research task is to
  recover, from these high-resolution meshes, the original construction
  proportions in which the instruments were designed: the geographical *klima*,
  the cone opening, and the scale factor. This paper establishes the
  epistemological and computational architecture of that programme. It defines
  the four interdependent layers of the workflow — theoretical geometry,
  automated feature perception, measurement consensus on physical stones, and
  literate coding as an executable specification for AI agents. We provide a
  systematic review of the drafts developed on the research volume through June
  2026, recording what each established, identifying contradictions in earlier
  models, and documenting the governing corrections: that seasonal meridian
  marks determine the cone independently of latitude, that engraved hour lines
  on real stones are circular arcs rather than mathematical generators, and
  that chisel groove notches lie within the scan noise floor.
keywords:
  - conic sundial
  - 3D reconstruction
  - literate coding
  - Berlin Sundial Project
  - construction proportions
  - computational archaeology
  - AI research agents
bibliography: references.bib
format:
  arxivai-typst: default
---

::: {.callout-note title="Supplement S1" #supplement-box}
Supplement S1 to the Ancient Sundials repository (Berlin Sundial Collaboration, ancient-sundials.lettre.ai). This draft documents the programme behind the repository's measurement panel and astronomical picture. It is a preprint draft and has not been peer reviewed. Where a figure in the repository differs from a figure here, the repository, which is rebuilt from the current database, is authoritative.

Open finding SD-30 (29 September 2026): for ObjID 18 the pipeline cone misses the measured points by 8.9 mm, where verified cones fit within 0.39 to 0.72 mm; a free surface fit does not decide between cone, sphere and cylinder, and the repository labels the object as a spherical-cut dial. Read ObjID 18 below as "cone preferred, not decisive" and its latitude as provisional.

Open finding SD-29: the measurement pipeline assumes a cone. The surface fit finds 80 cones, 28 spheres and 5 cylinders among the 113 detected objects, so cone-specific quantities hold only for objects classified as cones.
:::


# The Core Problem: Reconstruction versus Surface Fitting

Ancient Greek and Roman conic sundials represent sophisticated instruments of
mathematical astronomy carved into marble or limestone blocks [@gibbs1976; @schaldach2006].
When archaeologists recover these objects, they are typically weathered fragments:
vertices are chipped, mounting bases are broken, and the original gnomon rod has
disappeared.

High-resolution 3D surface photogrammetry (NXS and GLB point clouds with
sub-millimetre sampling) captures the physical morphology of these fragments
in exquisite detail [@bsdp2014dialface18]. However, raw 3D data do not explain
themselves.

```{=typst}
#arxivai.concept-box(title: [The reconstruction standard])[
  A computational reconstruction of an ancient sundial is not a surface fit to a
  point cloud. It is the deductive recovery of the intended construction
  proportions in the maker's own design vocabulary — the klima ratio $s : g$,
  the opening $rho = tan alpha$, and the scale $d$ — grounded in an explicit error
  model that attributes discrepancies to ancient artisan error rather than
  blurring them into an unweighted average.
]
```

Naive geometric regression — such as fitting an unconstrained cone or sphere to
all surface points — collapses when applied to damaged fragments. A blind
least-squares fit on short arcs produces unstable cone axes, misclassifies dial
types, and leaves the intended geographical latitude undetermined. To recover
the maker's design, computational algorithms must be constrained by the domain
theory of ancient gnomonics.

# The Four-Layer Architecture of the Programme

The reconstruction programme is organized into four hierarchical knowledge
layers. Each layer answers a distinct research question, consumes inputs from
the layer below, and produces verified outputs for the layer above.

```{=typst}
#arxivai.method-box(title: [The four-layer hierarchy])[
  1. Theory: What is the ideal geometric instrument, and what can be recovered?
  2. Perception: How do we extract physical chisel grooves from noisy 3D meshes?
  3. Measurement: How do we derive defended construction proportions from real stones?
  4. Method: How do we orchestrate autonomous AI agents to execute the analysis reproducibly?
]
```

## Layer 1: Theoretical Construction Geometry

The theoretical foundation is established in the keystone geometry paper
[@grasshoff2026conic]. It proves that on a conic dial whose nodus lies on the
polar axis:

* Every seasonal shadow path is an exact circle orthogonal to the cone axis.
* The equinoctial circle invariant $\ell_E = d$ anchors the physical scale.
* The solstitial noon marks obey the workshop rule $x = \tan\alpha\tan\varepsilon$,
  generating the triplet $\ell_W : \ell_E : \ell_S = (1+x)^{-1} : 1 : (1-x)^{-1}$.
* The apex-free marking ratio $Q = (L_S - L_W)/(L_S - L_E)$ recovers $\alpha$,
  $A$, and $d$ in closed form independently of latitude $\varphi$.

## Layer 2: Automated Feature Perception

The perception layer extracts physical features from dense 3D meshes without
human operator bias. It operates across three distinct levels:

1. **The Physical Groove:** A chisel cut of width $2$--$3\,\mathrm{mm}$ and
   depth $1$--$4\,\mathrm{mm}$.
2. **The Feature Point:** The medial centreline coordinate extracted along
   cross-sectional slices of the selection band.
3. **The Geometric Primitive:** 3D circular arcs and lines fitted to extracted
   centrelines.

## Layer 3: Measurement and Consensus Estimation

The measurement layer applies the theoretical framework to individual physical
artifacts, exemplified by Dialface ObjID 18 (British Museum inv. 2547)
[@grasshoff2026measurement; @bsdp2014dialface18] and ObjID 126 (Delos)
[@grasshoff2026objid126].

Rather than seeking a single global fit, this layer evaluates multiple independent
**construction witnesses**:

* **Witness 1 (Front Plane):** The dressed upper front face cut perpendicular
  to the cone axis, yielding $\varphi$ relative to the base.
* **Witness 2 (Cone Axis):** The spatial tilt of the fitted cone axis against
  the seating plane.
* **Witness 3 (Solstitial Hour Division):** The half-day arc $H$ of the
  solstitial day-lines, calibrated by the equinoctial standard $H_E = 90^\circ$.

A hierarchical model separates photogrammetric measurement noise from the
artisan's stonecutting realization spread $\tau$. When a witness deviates
significantly (e.g. $>4\sigma$), a leave-one-out rule identifies it as an
isolated ancient construction error rather than corrupting the consensus.

## Layer 4: Literate Coding and Agentic Orchestration

The methodological layer governs how computational research is conducted. Using
literate programming extended to AI agents, a single Quarto document functions
as the *executable specification* against which AI agents plan, execute, and
refine code in a closed feedback loop.

# Review of Drafts on the Research Volume

The working drafts of the programme reside on the project volume `2025Berlin`
under `ai-sundial3d/arxiv/`. An editorial audit reveals substantial progress
alongside critical theoretical corrections that have shaped the final bundle.

## The Geometry Draft and the Latitude Uncoupling

The initial draft of *Construction Geometry of the Conic Sundial with Horizontal
Gnomon* (12 June 2026) established the forward shadow equations and the $Q$-formula.
However, it contained an internal contradiction:

* Section 6 asserted the equinox invariant $\ell_E / d = 1$.
* Section 7 tabulated $\ell_E / d = \cos\varphi$, mixing the conic dial's axial
  coordinate with the vertical gnomon's horizontal shadow.
* A proposition claimed that the solstice ratio $\ell_S / \ell_W$ served as a
  monotone latitude diagnostic.

The revised geometry paper [@grasshoff2026conic] eliminates this equivocation.
The noon-mark triplet on a conic dial is strictly independent of latitude
$\varphi$. Rotational symmetry about the polar axis ensures that the relative
axial positions of the three noon marks depend solely on $\alpha$ and $\varepsilon$.
Latitude $\varphi$ enters the stone only through the hour division of the
solstitial day-lines and the orientation of the dressed block faces.

## The Measurement Draft for ObjID 18 (BM 2547)

*Geometrical Measurement of an Ancient Conic Sundial from a 3D Surface Model*
(11 June 2026) provides the empirical demonstration on the British Museum marble
fragment [@grasshoff2026measurement]:

* **Consensus Latitude:** $\varphi = 41.45^\circ$ ($68\%$ CI $[41.16^\circ, 41.75^\circ]$),
  corresponding to $s = 53;02$ for $g=60$, matching Ptolemy's Hellespont and
  Rome klimata.
* **Noon Triplet:** $\ell_W : \ell_E : \ell_S = 12 : 15 : 20$ (within $0.6\sigma$),
  confirming the Hellenistic workshop ratio $x = 1/4$.
* **Error Diagnosis:** The summer solstice hour division deviates by $14\sigma$,
  definitively identified as an isolated ancient carving error, corroborated by
  a $4.9\%$ compression of the equinoctial hour division.
* **Hour Lines as Circular Arcs:** The eleven engraved hour lines are circular
  arcs with radii between $113\,\mathrm{mm}$ and $280\,\mathrm{mm}$ (fit RMS
  $0.20$--$0.71\,\mathrm{mm}$), whereas straight generator models leave large
  systematic residuals of $2.2$--$7.6\,\mathrm{mm}$ RMS.

## The Delos Draft for ObjID 126

*The Cone Geometry of ObjID 126 (Delos)* (12 June 2026) analyses an exceptionally
well-preserved dial from the Delos Archaeological Museum [@grasshoff2026objid126]:

* **Direct Cone Surface Fit:** Direct orthogonal-distance regression on 3,345
  points yields half-angle $\alpha = 34.3^\circ$ (RMS $0.39\,\mathrm{mm}$).
* **Equinoctial Calibration:** The engraved equinoctial arc $H_E = 94.74^\circ \pm 0.31^\circ$
  exceeds the theoretical $90^\circ$ by $+5.27\%$, providing the engraver's
  unit scale factor to correct the solstitial arcs.
* **$\varepsilon$-Free Inversion:** Pairing each solstitial circle's placement
  with its own division confirms the Delos klima independently of epoch obliquity.

## Feature Detection: Evolution from Depth to Medial Filtering

*Automated Feature Detection on Ancient Conic Sundial Fragments* (30 March 2026,
amended 11 June 2026) introduced cross-sectional profile analysis [@grasshoff2026detection]:

* **The Depth Estimator Failure:** Early implementations sought the feature
  point as the minimum depth ($\mathrm{argmin}_z$) of each cross-section.
  Empirical noise analysis demonstrated that surface weathering and scan noise
  ($\pm 0.3\,\mathrm{mm}$) exceed the depth of the groove notch, causing depth
  minimizers to jump erratically to the boundary of the selection window.
* **Medial Line Estimation:** The robust feature point is the medial centreline
  of the selection band, which provides an unbiased estimator of the groove axis.

## Model Optimization: Multi-Stage Nonlinear Estimation

*Automated Model Optimization for Sundial Feature Detectors* establishes the
nonlinear parameter estimation framework following Bard [@bard1974]:

* Compares Trust Region Reflective (TRF), Nelder-Mead simplex, and L-BFGS-B
  joint optimization over the 7 fundamental cone parameters.
* Solves the partial-arc bias problem on fragmented day-lines by enforcing
  coaxiality with the robust direct surface cone fit.

## Literate Coding Methodology

The methodology drafts (April and June 2026) formalize the three levels of
operation in AI-assisted scientific discovery [@grasshoff2026literate]:

* Demonstrates how an AI agent operating with IDE tools (sub-agents, skills,
  MCP servers) uses the research paper as a living specification to prevent
  hallucinations and ensure mathematical consistency.

# Literate Coding: Executable Specifications for AI Agents

When computational tasks are delegated to AI coding agents, standard prompt-based
workflows fail because the domain theory is separated from the code. The agent
may generate mathematically plausible Python code that fits a geometric primitive
while violating the physical principles of ancient gnomonics.

```{=typst}
#arxivai.concept-box(title: [Principle of Complexity Reduction])[
  Before delegating an analytical task to an AI agent, construct a single literate
  document that embeds the complete mathematical derivations, data formats,
  quantitative success criteria, and runnable code blocks. The document is the
  agent's complete and bounded world.
]
```

The methodology distinguishes three operational levels:

1. **Level 1 (Reference Background under Permanent Revision):** The Quarto paper
   contains the authoritative domain theory. When higher-level execution uncovers
   an anomaly (e.g. that straight lines fail to fit hour lines), the domain
   theory is revised, updating the agent's reference substrate.
2. **Level 2 (Structured Planning Substrate):** The AI agent reads the literate
   specification to formulate an execution plan. Every computational step must
   cite a specific equation or constraint ID from the document.
3. **Level 3 (Autonomous Agentic Action & Validation):** The agent executes the
   code, checks results against quantitative thresholds, and logs iterations.

```{=typst}
#arxivai.operation-box(
  title: [Autonomous Sundial Feature Inversion],
  question: [Recover cone parameters $(alpha, d, phi)$ from 3D point cloud of ObjID 18],
  steps: (
    (agent: "mesh-ingest", does: [Loads PLY point cloud BSDP0300 and verifies mm coordinate frame]),
    (agent: "groove-extractor", does: [Computes cross-section slices and extracts medial centrelines across 14 feature bands]),
    (agent: "surface-cone-fitter", does: [Fits unlabelled right circular cone by orthogonal distance regression, yielding $alpha = 34.4 degree$]),
    (agent: "apex-free-inverter", does: [Extracts noon marks $P_S, P_E, P_W$, computes marking ratio $Q = 1.395$, and inverts apex $A$ and scale $d$]),
    (agent: "witness-consensus", does: [Combines front plane, cone axis, and winter hour division into consensus $phi = 41.45 degree$, flagging summer division at $14 sigma$]),
  ),
  latency: "1.8 s",
  answer: [$alpha = 34.40 degree plus.minus 0.09 degree$, $d = 101.3 plus.minus 0.4 " mm"$, $phi = 41.45 degree plus.minus 0.30 degree$ ($s = 53;02$), workshop rule $x = 1:4$ ($12:15:20$)],
)
```

# Lessons Learned and Epistemic Standards

The systematic review of the sundial bundle establishes five core methodological
rules that govern all subsequent papers in the series:

1. **Latitude is Uncoupled from Noon Marks:** No pipeline may attempt to invert
   meridian marks alone for latitude $\varphi$.
2. **Hour Lines are Empirically Tested as Arcs:** Feature detectors must evaluate
   both straight line and circular arc model classes.
3. **Medial Filtering Replaces Depth Minima:** Groove extraction on weathered
   stone relies on medial windowing rather than $\mathrm{argmin}_z$.
4. **Discordant Features are Attributed as Historical Errors:** Significant
   outliers are isolated via leave-one-out testing and cataloged as ancient
   craftsmanship errors rather than blurred by global smoothing.
5. **No Unverified Corpus Claims:** Every quantitative claim must cite a
   specific museum accession number (e.g. British Museum inv. 2547) and a
   reproducible pipeline script.

# The Sundial Publication Bundle

@tbl-bundle summarizes the complete publication bundle comprising the
ai-sundial3d reconstruction programme.

| Paper ID | Title | Core Contribution | Status |
|:---|:---|:---|:---:|
| **P1** | *Construction Geometry of the Conic Sundial* [@grasshoff2026conic] | Forward projection, $Q$-formula, noon triplet, marking sufficiency | **Complete** |
| **P2** | *Automated Feature Detection on Sundial Fragments* [@grasshoff2026detection] | 3-level groove model, cross-section slicing, medial line extraction | **Complete** |
| **P3** | *Automated Model Optimization for Feature Detectors* [@grasshoff2026optimization] | Nonlinear parameter estimation, Bard framework, TRF/L-BFGS-B | **Complete** |
| **P4** | *Literate Coding with AI Agents* [@grasshoff2026literate] | 3-level executable specification, agent orchestration loop | **Complete** |
| **P5** | *Geometrical Measurement of ObjID 18* [@grasshoff2026measurement] | Full empirical deduction on BM inv. 2547, witness consensus | **Draft** |
| **P6** | *The Cone Geometry of ObjID 126 (Delos)* [@grasshoff2026objid126] | Direct surface cone fit, $H_E$ calibration, $\varepsilon$-free latitude | **Draft** |

: The comprehensive publication bundle for the ancient conic sundial reconstruction programme. {#tbl-bundle}

# Conclusion

The recovery of ancient conic sundials from 3D scan data demonstrates how
rigorous domain theory, robust feature perception, and literate agentic
workflows combine into a cohesive scientific programme. By replacing naive
mesh regression with the exact proportional geometry of Hellenistic astronomy,
fragmentary stone artifacts can be reconstructed with mathematical certainty,
shedding new light on the metrological practices of the ancient world.

# References {.unnumbered}