Abstract
The inZORi engine is applied to 13,000 TESS light-curve signals to produce a ranked priority list for follow-up observation — without supervised training, labeled examples, or explicit weighting formulas. Seven signal quality dimensions are evaluated: regularity, stability, SNR, scatter, transit count, duration, and depth. A consensus rank across 4 ranking variants identifies 294 clean transit candidates for priority review; the top candidate achieves rank_sum = 7.
1. Experimental Setup
Signal Quality Dimensions
| Dimension | What it measures | Higher = better? |
|---|---|---|
| Regularity | Period stability across transits | Yes |
| Stability | Baseline flux consistency (out-of-transit) | Yes |
| SNR (normalized) | Signal-to-noise ratio of transit dip | Yes |
| Quality fraction | Fraction of valid (non-gapped) observations | Yes |
| Scatter (normalized) | Noise level — lower is better; filter: ≤0.7 | No (inverted) |
| Transit count (normalized) | Number of detected transit events | Yes |
| Duration / Depth | Physical plausibility of transit shape | Contextual |
Ranking Methodology
Four independent ranking variants are computed over the 7 dimensions, each applying different weighting emphasis (regularity-first, SNR-first, composite, stability-first). A consensus rank_sum aggregates all four ranks — lower rank_sum means stronger consensus across methods. The scatter filter (scatter_norm ≤ 0.7) removes high-noise signals before ranking.
- No neural network, no supervised labels, no predefined transit template
- Criteria weights discovered through inZORi selection, not manually specified
- Output is fully auditable: all 7 dimension scores per candidate are published
- Designed for human-in-the-loop follow-up, not autonomous classification
2. Results
Signal Classification Breakdown
| Class | Count | % of Total | Action |
|---|---|---|---|
| Clean transit candidates | 294 | 2.3% | Priority follow-up |
| Variable / artifact | 513 | 3.9% | Secondary review |
| Eclipsing binary (confirmed) | 1 | <0.01% | Known object |
| Weak signal | 1 | <0.01% | Insufficient evidence |
| Uncertain | 12,191 | 93.8% | Archive |
| Total | 13,000 | 100% | — |
Top Candidate Profile (TIC 55524055)
| Metric | Value | Note |
|---|---|---|
| Rank sum | 7 | Strongest multi-variant consensus |
| Rank max | 3 | Worst rank in any single variant |
| Stability | 0.979 | Excellent baseline stability |
| SNR (norm) | 0.768 | High signal-to-noise |
| Depth (norm) | 0.910 | Deep, well-defined transit dip |
| Scatter (norm) | 0.599 | Within filter threshold (≤0.70) |
| Regularity | 0.456 | Moderate periodic regularity |
3. Key Findings
- 294 clean transit candidates from 13,000 signals — 2.3% yield, consistent with survey expectations.
- Strong class separation: clean transit signals score distinctly higher on stability and SNR vs variable/artifact classes.
- Consensus ranking: rank_sum provides robust prioritization by aggregating 4 independent orderings.
- Fully auditable: all 7 dimension scores for every candidate are in the published JSON report.
- No supervision: the same inZORi engine that navigates 2D environments here evaluates 7D signal quality space.
4. What This Demonstrates
Unsupervised Prioritization at Scale
This scenario shows inZORi operating as a multi-criteria prioritization engine in a scientific domain with no labeled training data. The ranking is consensus-based across multiple evaluation perspectives — a single outlier dimension does not dominate, making output more robust to noise.
Application: Genomics, drug screening, materials discovery, financial signal processing — any domain requiring expert-level triage of large catalogs with multi-dimensional quality criteria.
5. Reproducibility
Framework: inZORi v1.0 | Domain: Astronomy / TESS light-curve analysis
Dataset: 13,000 TESS signals (S0001 snapshot) | Dimensions: 7 | Ranking variants: 4
Output: tests/tess/candidates_report.json (full ranked list)
Note: Evaluation strategy and genome encoding are proprietary. Quality dimensions and normalization are fully disclosed above.