getter-one 0.2.2

GETTER One: Geometric Event-driven Tensor-based Time-series Extraction & Recognition - Structural event detection and causal network extraction for N-dimensional time series

GETTER One

Geometric Event-driven Tensor-based Time-series Extraction & Recognition

Omnidimensional Network Engine

---

v0.1.2 (2026-04) — 📄 Paper Submission Version

This version corresponds to the manuscript submitted to Information Science.

pip install getter-one==0.1.2

GETTER One is a discrete geometric framework for structural event detection and causal network extraction in N-dimensional time series.

It is not a forecasting model. It detects what is changing, when, and which dimensions are driving the change — using closed-form geometric computations with no MCMC, no gradient descent, and no threshold tuning.

Key Features

• Lambda³ (Λ³) structural analysis — Computes displacement vectors (ΛF), structural tension (ρT), cross-dimension synchrony (σₛ), and cooperative events (ΔΛC) in closed form
• Causal network extraction — Identifies directed, time-lagged causal relationships between dimensions
• Statistical confidence — Permutation tests, bootstrap confidence intervals, and effect sizes (no Bayesian inference required)
• Domain-agnostic — Works on any N-dimensional time series: weather, finance, sensors, biology
• Minimal dependencies — Core requires only numpy, scipy, pandas
• GPU-ready — Optional CUDA acceleration for large-scale datasets via CuPy

Installation

pip install getter-one

Optional extras:

pip install getter-one[gpu]       # CUDA/CuPy acceleration
pip install getter-one[viz]       # matplotlib + plotly
pip install getter-one[full]      # everything

Quick Start

from getter_one.pipeline import run

# One line — full pipeline
result = run("weather.csv", target="precipitation")
print(result.report)

Step by Step

from getter_one.data import load
from getter_one.structures import LambdaStructuresCore
from getter_one.analysis import NetworkAnalyzerCore, assess_confidence

# 1. Load data
dataset = load("weather.csv", target="precipitation", normalize="range")

# 2. Compute Λ³ structures
core = LambdaStructuresCore()
structures = core.compute_lambda_structures(
    dataset.state_vectors, window_steps=24,
    dimension_names=dataset.dimension_names,
)

# 3. Extract causal network
analyzer = NetworkAnalyzerCore(sync_threshold=0.3, causal_threshold=0.25, max_lag=12)
network = analyzer.analyze(dataset.state_vectors, dimension_names=dataset.dimension_names)

# 4. Assess confidence
confidence = assess_confidence(
    state_vectors=dataset.state_vectors,
    lambda_structures=structures,
    network_result=network,
    dimension_names=dataset.dimension_names,
)

CLI

# Run full pipeline
getter-one run weather.csv --target precipitation --report report.md

# Data preparation
getter-one-loader load weather.csv --target precipitation -o prepared.csv
getter-one-loader merge weather.csv air_quality.json --time date -o merged.csv
getter-one-loader info data.csv

# System info
getter-one info
getter-one check-gpu

Pipeline Architecture

Input Data (csv/json/parquet/xlsx/npy)
  │
  ▼
┌─────────────────────────────────┐
│  Lambda³ Structures (Λ³)        │  Closed-form geometric computation
│  ΛF, ΛFF, ρT, σₛ, Q_Λ, ΔΛC   │  No parameters to tune
└──────────────┬──────────────────┘
               │
  ┌────────────┼────────────┐
  ▼            ▼            ▼
┌──────┐  ┌──────┐  ┌──────────┐
│Bound │  │Topo  │  │Anomaly   │  Detection modules
│ary   │  │Break │  │Detection │  (GPU-accelerated)
└──┬───┘  └──┬───┘  └────┬─────┘
   └─────────┼───────────┘
             ▼
┌─────────────────────────────────┐
│  Causal Network Analysis        │  Sync + directed causal links
│  Sync matrix, lag estimation    │  Hub/driver/follower detection
└──────────────┬──────────────────┘
               ▼
┌─────────────────────────────────┐
│  Confidence Assessment          │  Permutation test → p-values
│  No MCMC, no Bayesian inference │  Bootstrap → confidence intervals
└──────────────┬──────────────────┘
               ▼
┌─────────────────────────────────┐
│  Report Generation              │  Markdown report with all results
└─────────────────────────────────┘

What GETTER One Is (and Isn't)

	GETTER One	Traditional Models
Purpose	Structural event detection	Forecasting / prediction
Method	Discrete geometry (closed-form)	Regression / deep learning
Parameters	Window size only	Thresholds, hyperparameters
Output	Events, boundaries, causal network	Predicted values
Relationship	Complements prediction models	Competes with each other

GETTER One detects what changed and why, not what will happen next. It can be used as a feature extractor for prediction models — experiments show that combining raw data with GETTER One features improves Transformer event detection F1 by +5.7% over raw data alone, and outperforms adding 7 additional weather variables (+4.3% vs +6.1%).

Benchmark Results

Standard Benchmark

Causal inference benchmark against 5 established methods on synthetic data with ground truth (6 scenarios, 20 repeats each):

Method	Composite	F1 (dir)	Lag MAE	Sign Acc	Spurious
PCMCI+	0.877	0.740 ± 0.364	0.000 ± 0.000	1.000 ± 0.000	0.000 ± 0.000
VAR Granger	0.853	0.697 ± 0.353	0.010 ± 0.100	0.990 ± 0.100	0.050 ± 0.154
GETTER One	0.821	0.792 ± 0.366	0.000 ± 0.000	1.000 ± 0.000	0.500 ± 0.000
Event XCorr	0.537	0.534 ± 0.403	0.244 ± 0.825	—	0.300 ± 0.340
Transfer Entropy	0.446	0.432 ± 0.368	0.413 ± 1.396	—	0.525 ± 0.112
Graphical Lasso	0.120	—	—	—	1.000 ± 0.000

Per-scenario F1 (directed):

Scenario	GETTER One	PCMCI+	VAR Granger	TE	EventXCorr
S0 null	0.000	0.000	0.000	0.000	0.000
S1 delayed	1.000 ± 0.000	0.825 ± 0.183	0.662 ± 0.141	0.645 ± 0.233	0.450 ± 0.446
S2 asymmetric	0.950 ± 0.122	0.883 ± 0.163	0.658 ± 0.037	0.662 ± 0.177	0.520 ± 0.434
S5 confounder	0.800 ± 0.000	0.990 ± 0.045	0.960 ± 0.082	0.734 ± 0.092	0.623 ± 0.213
S7 event delayed	1.000 ± 0.000	0.825 ± 0.183	0.917 ± 0.148	0.342 ± 0.417	0.842 ± 0.206
S8 event asym	1.000 ± 0.000	0.917 ± 0.148	0.983 ± 0.075	0.212 ± 0.356	0.767 ± 0.262

GETTER One achieves F1 = 1.000 with σ = 0.000 (perfectly deterministic) on all event-driven scenarios, with zero lag error and perfect sign accuracy across all 20 trials. Its closed-form computation produces identical results regardless of noise realization — a direct consequence of the displacement-based (diff) approach inherited from molecular dynamics trajectory analysis.

Hell Mode Robustness Benchmark

Causal detection under non-stationary, non-linear, non-Gaussian conditions (8 scenarios × 5 repeats, ground truth: X0→X1 lag=2, X0→X2 lag=3):

Method	F1	Precision	Recall	Lag MAE	Detected	Spurious
GETTER One	0.394 ± 0.406	0.471	0.362	0.000	1.4	0.118
PCMCI+	0.376 ± 0.271	0.279	0.738	0.306	8.6	0.142
VAR Granger	0.283 ± 0.308	0.217	0.475	0.630	7.5	0.260
Transfer Entropy	0.164 ± 0.240	0.119	0.887	0.770	84.3	0.188

Per-scenario F1 (Hell Mode):

Scenario	GETTER One	PCMCI+	VAR Granger	TE
H1 pulse	0.513	0.494	0.431	0.385
H2 phase jump	0.547	0.183	0.093	0.025
H3 bifurcation	0.000	0.142	0.040	0.025
H4 cascade	0.133	0.649	0.581	0.668
H5 resonance	0.960	0.774	0.000	0.043
H6 decay	0.533	0.428	0.813	0.044
H7 multi-hell	0.467	0.238	0.304	0.029
H8 progressive	0.000	0.100	0.000	0.092

GETTER One achieves the highest F1 and precision under non-ideal conditions, with the fewest detected links (1.4 avg vs ground truth of 2) — demonstrating conservative, high-confidence detection rather than over-reporting. Each method has characteristic strengths: GETTER One excels at phase jumps and resonance, PCMCI+ at cascades, and VAR Granger at structural decay.

Real-World Application: Weather Network

Analysis of 7 cities around Tokyo (hourly data, October 2024) reveals physically correct causal structures:

Precipitation causal network:

• Nagano → Gunma (lag=1h) → Saitama (lag=1h) — inland-to-plains propagation
• Yamanashi → Gunma (lag=2h) — basin-to-plains flow
• Drivers: Nagano, Yamanashi (mountain sources)
• Followers: Saitama, Tochigi (downstream receivers)

Cross-variable causality (42D full analysis):

• Temperature changes → Surface pressure changes (lag=2-3h) — thermodynamic law rediscovered from data alone

Supported Data Formats

Format	Extension	Read	Write
CSV / TSV	.csv, .tsv	✅	✅
JSON	.json	✅	—
Parquet	.parquet	✅	—
Excel	.xlsx, .xls	✅	—
NumPy	.npy, .npz	✅	✅

Project Structure

getter_one/
├── data/           Data loading & merging (multi-format)
├── structures/     Lambda³ structural computation (CPU)
├── core/           GPU infrastructure (CuPy/CUDA)
├── detection/      Boundary, topology, anomaly, phase-space detection
├── analysis/       Network analysis, confidence assessment, report generation
├── pipeline.py     Full pipeline orchestration
└── cli.py          Command-line interface

Related Projects

• BANKAI-MD — GPU-accelerated molecular dynamics analysis built on the same Lambda³ theory. GETTER One is the domain-agnostic generalization; BANKAI-MD is the MD-specialized implementation.

Citation

Paper in preparation. If you use GETTER One in your research, please cite this repository.

License

MIT License — see LICENSE for details.

---

Built with 💕 by Masamichi & Tamaki

CHANGE EAGLE // Aerial Type // Structure Detection