antikythera-spectral 0.3.0

Hyperdimensional-computing encoder + Pyodide bridge for the Antikythera mechanism

antikythera-spectral

A hyperdimensional-computing model of the Antikythera mechanism, packaged for Python and the browser.

Encode any date as a state vector that any of the mechanism's dials can be read from. Decode, compare against modern ephemeris truth, simulate the Hellenistic operator's seasonal-recalibration workflow, and run the project's 31-row hypothesis battery — all from one pip install. Pyodide-compatible: the same package runs in a browser via micropip so a web app can drive a digital Antikythera with no Python server.

What is the Antikythera mechanism?

A bronze hand-cranked astronomical calculator built in Hellenistic Greece around 150–60 BCE, recovered in 1901 from a shipwreck off the island of Antikythera. Its surviving 30+ gears predict solar / lunar / planetary positions, eclipses (via the 18-year Saros cycle), and the four-year Olympiad calendar. Freeth et al. 2021 (Sci. Rep. 11:5821) is the current authoritative reconstruction; the device's planetary front face has not survived intact and remains the subject of active research.

The HDC framing this package implements: every gear is a faithful representation of ℤ/nℤ; every mesh is a rational map between cyclic groups; every dial pointer is a hypervector whose components are the phase angles on the various dials. The Greeks built a resonant HDC object before Plate wrote HRR.

Install

pip install antikythera-spectral

The base install pulls only numpy and is fully self-contained as of v0.2.0 — every bridge method works out of the box, including visibility windows, heliacal-rising prediction, eclipse search, and solar-elongation queries (using algebraic synodic-cycle propagation; ~±1 day Antikythera-grade precision). No JPL ephemeris kernel required.

Optional extras for sub-arcsec-precision validation against modern ephemerides:

pip install "antikythera-spectral[ephemeris]"    # adds skyfield + jplephem; pass precise=True
pip install "antikythera-spectral[hypotheses]"   # adds scipy (for chi-square in H-H1)
pip install "antikythera-spectral[plot]"         # adds matplotlib
pip install "antikythera-spectral[all]"          # everything above

In algebraic mode (the default), the four sky-truth bridge methods accept precise=False (the default) and use frozen per-planet anchors + closed-form propagation. With precise=True they switch to skyfield + a JPL DE-kernel for sub-arcsec accuracy. See ADR 0011 for the discipline.

Pyodide / micropip (in-browser)

import micropip
await micropip.install("antikythera-spectral[ephemeris]")

The full Bridge API is available from a Pyodide REPL with no server-side Python.

Quick start (Python)

>>> from antikythera_spectral import bridge, default_encode
>>> result = bridge.get_dial_state(jd_tdb=1684500.0)   # ~205 BCE
>>> result['ok']
True
>>> round(result['dials']['Metonic']['angle_deg'], 2)
355.07
>>> result['dials']['Mars_synodic_period_relation']['modulus']
133
>>> result['state']['dtype']        # v0.3.0: bridge default flipped to bit-packed
'uint64'
>>> result['backend']
'bit'

# Top-level shortcut for "encode any date as a hypervector" (v0.3.0):
# default backend is the bit-packed ALU; pass backend="complex128" for the
# legacy reference encoder. See ADR 0012 for the algebra-first rationale.
>>> hv = default_encode(1684500.0)
>>> hv.dtype
dtype('uint64')
>>> hv.shape
(15,)                              # ceil(940/64) packed words = 120 B (vs 15 KB)

# Mars planetary models (v0.3.0): four longitude models × three named
# reconstruction param sets. The `bronze` model is the algebra/eigenbasis
# projection of the gear-ratio cyclic-group representation through the
# pin-and-slot phase-space transform — see ADR 0012.
>>> from antikythera_spectral import mars_models
>>> round(mars_models.mars_longitude_bronze(
...     1721000.0, mars_models.FREETH_2012_MARS_PARAMS
... ), 2)
124.09

# Reproduce F&J 2012 Fig 39's "nearly 38°" Mars peak error from
# the bare deferent + epicycle on the middle 7 retrogrades vs
# JPL DE422 (lazy; needs the kernel cached).
>>> finding = mars_models.fj_38deg_finding()
>>> round(finding["rms_deg"], 2)
38.85

# Bit-packed binary HDC ALU (v0.3.0): every operation reduces to
# integer bit-ops on packed uint64s. 125x smaller than complex128;
# 2-10x faster bind. See figures/bit_alu_findings.md for the benchmark.
>>> from antikythera_spectral import bit_alu
>>> import numpy as np
>>> rng = np.random.default_rng(0)
>>> a = bit_alu.random_hv(940, rng)
>>> b = bit_alu.random_hv(940, rng)
>>> abs(bit_alu.similarity(a, b, 940)) < 0.1   # uncorrelated random pair
True
>>> bit_alu.hamming_distance(bit_alu.bind(a, b), bit_alu.bind(a, b))
0

Quick start (CLI)

antikythera-spectral encode --jd 1684500.0
antikythera-spectral visibility --planet mars --from-jd 1684500 --to-jd 1685000
antikythera-spectral compare ephemerides --jd 1684500 --body mars \
                                          --kernel-a de421 --kernel-b de441_part1

# v0.3.0: compare Mars models with the new bronze + named param sets
antikythera-spectral compare models --jd 1721000 --body mars \
    --model-a bronze --model-b equant --params freeth_2012 --kernel de422

antikythera-spectral hypotheses --csv-out -

What can you do with it?

• Encode any Julian date as a state vector across all of the mechanism's dials.
• Decode any state vector back to per-dial residues (round-trip is exact for the LCM variant).
• Convert dates between Gregorian, Julian, Athenian archonship + Attic months, and Olympiad year — the four calendar systems an ancient Greek astronomer or modern reader might use.
• Compute visibility windows for each planet (heliacal rising / setting + solar elongation), the astronomical reality that gates the operator's recalibration workflow.
• Search for eclipses in any date band via sky-driven ephemeris enumeration.
• Simulate the operator workflow (§11.6.16 of the research notebook): start at a date, advance, observe at heliacal rising, re-anchor, repeat.
• Compare reconstructions — Freeth 2021 vs Wright vs Price 1974 dial readings simultaneously at any date.
• Compare ephemeris kernels — DE421 vs DE441 vs DE441_part1 deltas at a chosen JD/body in arc-seconds, kilometers, AU.
• Run Hellenistic Mars planetary models (antikythera_spectral.mars_models, v0.3.0): four longitude models (uniform, epicycle-only, equant, bronze — the gear-ratio cyclic-group projection through the pin-and-slot phase-space transform) under three named param sets (Ptolemy / Almagest IX-X, Freeth & Jones 2012, Freeth 2021). mars_models.fj_38deg_finding() reproduces F&J 2012 Fig 39's "nearly 38°" Mars peak error directly from the bare deferent + epicycle on the middle 7 retrogrades of the 1st century BC vs JPL DE422 (lands at 38.85° within 0.85°). See ADR 0012 for the algebra-first scope discipline; full audit in figures/mars_38deg_gap_findings.md.
• Use the bit-packed binary HDC ALU (antikythera_spectral.bit_alu, v0.3.0): a second encoder backend where every operation reduces to integer bit-ops (XOR, popcount, shift) on packed uint64 words — no floating point, no complex multiplies, no matrix work. 125× smaller per hypervector than the complex128 reference (120 B vs 15 KB at D=940); 2–10× faster bind scaling with D. Same algebraic substrate (cyclic-group representation of the dials), pure-bitwise primitives. The cleanest possible incarnation of ADR 0012's algebra-first discipline. Benchmark + cycle-alignment analysis (solar vs sidereal day for permute = sigma_day — solar wins) in figures/bit_alu_findings.md.
• Encode without picking a backend (antikythera_spectral.default_encode, v0.3.0): one-liner that returns a hypervector under the package's default backend (the bit-packed ALU, per ADR 0012). Pass backend="complex128" to recover the legacy reference shape. The bridge's get_dial_state / decode_dial / decode_to_jd likewise default to the bit-packed backend in v0.3.0, with explicit backend="complex128" opt-out for v0.2.x consumers.
• Run the 31-row hypothesis battery that drives the research notebook, get JSON / CSV output.
• Override gear ratios (what-if mode) — re-encode with arbitrary p/q to explore alternative period relations like the canonical Venus 5/8.
• Inventory by fragment (archaeological mode) — list which gears are attested in fragments A/B/C/D vs reconstructed by Freeth.
• Babylonian Goal-Year overlay — given a planet+date, return what an astronomer using the 47-year Mars cycle (or 59-year Saturn, etc.) would have predicted.
• Animation export — emit a time-series of states over a date range for a viewer / animation frontend.

Bridge API

docs/bridge_api.md is the consumer-facing contract. 28 methods grouped by purpose; each returns a Pyodide-JSON-serializable {"ok": True, ...} dict. State-vector payloads serialise differently per backend: backend="bit" (v0.3.0 default) ships state.packed_uint64 as a JSON list of integers (each ≤ 2⁶⁴-1; JS consumers can use BigInt or split into uint32 pairs), while backend="complex128" ships state.interleaved_f32 as [re0, im0, re1, im1, …] length 2·D so JS can wrap it directly in a Float32Array. Per-dial residues / angles are floats / ints either way and fit standard JSON.

v0.3.0 extends bridge.compare_models with the bronze model name and a new optional params argument ("ptolemy" / "freeth_2012" / "freeth_2021"); v0.2.x callers continue to work unchanged. Direct Python access to the new Mars / bit-ALU primitives is via the antikythera_spectral.mars_models and antikythera_spectral.bit_alu facades — these are not part of the bridge contract (no Pyodide JSON serialization needed; numpy arrays / floats / dicts).

v0.3.0 default-backend flip. bridge.get_dial_state now returns a bit-packed state by default — the state sub-dict has dtype: "uint64", shape: [n_words] (= ceil(D/64)), n_bits: D, and packed_uint64: list[int]. bridge.decode_dial / decode_to_jd auto-detect the backend from the input shape (uint64 array → bit decoder; complex128 / interleaved-Float32 → reference decoder). v0.2.x consumers pass backend="complex128" to recover the legacy interleaved-Float32 shape; nothing else changes in the API. The dial residues / angles in the response are backend-independent.

Two HDC backends, one algebra

The package ships two HDC backends — complex128 (FHRR-style; the v0.2.x reference) and bit_alu (BSC-style; the v0.3.0 default). Both implement the same algebraic substrate: a cyclic-group representation of the Antikythera dials. We provide both deliberately because they sit at different points on a structural-fidelity / representation-theoretic-fidelity axis:

• bit_alu shares two structural properties with the bronze mechanism that complex128 does not. The state space is finite ({0,1}^D vs ℂ^D), and the operations (XOR, popcount, bit-rotate) are exact at the representation level rather than floating-point approximations. The bronze realises ℤ/Nℤ literally — a 53-tooth gear has 53 distinguishable rotational positions, full stop — and the binary substrate parallels that discrete substrate one level up the abstraction stack. Hardware-irreducible binary HDC is a real engineering literature (Kanerva 2009 §3 motivates BSC on this basis; Schmuck/Benini/Rahimi 2019 build bit-serial accelerators with no multiplier).

• complex128 is closer to the representation-theoretic eigenbasis. Pontryagin duality on a finite cyclic group gives the complex characters e^{2πi·k/N} as the natural orthonormal basis; FHRR-style binding via complex multiplication is exactly the convolution-on-characters story. So if you ask "what is the textbook eigenbasis of ℤ/Nℤ?" — that's the complex backend's home.

Important caveat we want to flag. Neither backend matches the bronze's actual tooth count. Both are holographic abstractions at dimension D = 940 or 13440, vastly larger than any single gear's tooth count (max ~250). The bit-ALU's discreteness is a representation-level parallel, not a mechanism-level identity. We are not claiming the bit ALU is "the" irreducible bit-level translation of the Antikythera mechanism — the mechanism is irreducibly integer-on-cyclic-groups, and binary HDC is one of several discrete encodings of that, not a privileged one.

Why both ship. The bit ALU is the v0.3.0 default because, in the ADR-0012 algebra-first discipline, "no FPU calls anywhere in the encode/decode path" is a meaningful purity property: floating-point introduces concepts (ULP, gradual underflow, NaN) that have zero counterpart in bronze gearing. The complex128 backend is preserved because the H-battery's algebraic identities (B-H1 round-trip, σ_day unit-operator) are most naturally stated in the FHRR / character-of-cyclic-group language, and because the complex Gaussian basis has cleaner decode behaviour for cross-talk-sensitive H-battery rows.

References for the framing: Kanerva (2009, Cognitive Computation 1:139); Plate (1995, IEEE TNN 6); Schlegel/Neubert/Protzel (2022, AI Review 55:4523, BSC vs FHRR comparison); Schmuck/Benini/Rahimi (2019, ACM JETC 15, hardware-binary HDC). For the bronze: Freeth et al. 2006 Nature 444:587; Freeth & Jones 2012 ISAW Papers 4; Freeth et al. 2021 Sci. Rep. 11:5821.

Hypothesis battery

The package ships the same 31-row H-battery the research scaffold runs. Headlines:

• 22 PASS — encoder round-trips perfectly for all D variants; pin-and-slot encodes T-symmetry breaking; manufacturing tolerance is fine for one Metonic cycle; etc.
• 3 PARTIAL — the proxy-metric Pareto for {7, 17}; the prime spectrum vs null model; H-H1 chi-square against Almagest periods.
• 3 FAIL — including E-H2: uniform Mars encoder peak ≥ 150° (this is expected; the failure mode is the rationale for the §11.6.16 operator-recalibration framing).
• 3 UNDETERMINED — open exploration F-series rows, plus skyfield-gated rows when no ephemeris kernel is on disk.

See the research notebook for the full row-by-row narrative.

Documentation

Topic	Link
Research notebook (the project narrative)	antikythera_spectral_research_notebook.md
Bridge API contract	bridge_api.md
Calendar systems reference	CALENDAR_SYSTEMS.md
Ephemeris kernels (DE421 / DE441 / etc.)	EPHEMERIS_KERNELS.md
ΔT discussion (Earth-rotation drift at -200 BCE)	DELTA_T_MODEL.md
Operator workflow simulation (§11.6.16)	OPERATOR_WORKFLOW.md
Mars 38° gap audit (10-analysis decomposition)	mars_38deg_gap_findings.md
Bit-ALU benchmark + cycle-alignment (v0.3.0)	bit_alu_findings.md
ADRs (architecture decisions)	docs/adr/ — ADR 0011 (algebraic default), ADR 0012 (algebra/eigenbasis vs CAD scope)
Roadmap	ROADMAP.md
Changelog	CHANGELOG.md

Citing

If antikythera-spectral contributes to a paper or write-up, please cite both the package and the research notebook:

@software{antikythera_spectral,
  author = {Kirkland, Steven},
  title = {antikythera-spectral: Hyperdimensional-computing model of the Antikythera mechanism},
  year = {2026},
  url = {https://github.com/lemonforest/mlehaptics/tree/main/docs/antikythera-maths/antikythera-spectral},
  version = {0.3.0}
}

License

GPL-3.0-or-later. Matches the license of the parent monorepo.

See also

Sibling packages in the same monorepo:

• chess-spectral — the same HDC + cyclic-group-algebra framing applied to chess, with native C accelerator.