About This Project
Overview
We build ML-guided ranking tools for rare Calabi-Yau-like targets in large search spaces, with reproducible seeds, target-rule checks, and shareable dossiers.
Demo corpora are synthetic Hodge-number draws inspired by published dataset statistics — not a live crawl of the full Kreuzer–Skarke census, and not experimental physics data.
Verified means a candidate passes that dataset’s target rule on those synthetic labels (e.g. |χ| < 100 for Kreuzer–Skarke). It is not experimental verification of a string vacuum.
upg-strings emphasizes reproducibility, transparent methodology, and durable product surfaces (Hall of Fame + shareable dossiers) over inflated precision marketing.
What Makes upg-strings Useful
While tools like CYTools focus on analyzing individual manifolds, upg-strings is a search / ranking layer: which synthetic candidates are worth opening a dossier for?
The problem we sketch
The Kreuzer–Skarke database describes hundreds of millions of reflexive polytopes. Finding geometries with specific topological properties is hard. This site demonstrates ranking + packaging on synthetic draws so the pipeline stays cheap and reproducible.
How we're different (honestly)
Existing tools
CYTools: Analyzes geometry of individual manifolds
Research papers: Classify or generate new manifolds
Traditional approach: Manual selection or random sampling
upg-strings
Ranks & packages: Scores candidates and opens shareable dossiers
Hall of Fame: Persistent board of target-rule hits
Honest metrics: Synthetic retrieval vs random baseline; label features held out of the model
A simple scenario
Goal: Calabi-Yau-like manifolds with small Euler characteristic (|χ| < 100) for phenomenological sketches.
- Full census search of ~474M polytopes → not what this demo runs
- Random sampling of synthetic draws → baseline hit rate in the API response
- ML-guided ranking with |χ| held out of the forest → retrieval lift vs that baseline on synthetic labels
- Open a dossier / pin to Hall of Fame when the target rule passes
Do not read marketing-style “8.7×” or “98% cost reduction” as measured KS physics performance — those were overstated relative to this synthetic setup.
Metrics that matter here
Synthetic retrieval
Precision@k / Recall@k against the dataset target rule on synthetic draws
vs baseline
Compare to random selection rate in the same draw (baseline_random_precision)
Leakage hold-out
Target-defining columns (e.g. absolute χ) are withheld from the RandomForest
Our Approach
- Information retrieval framing: Rank candidates, then verify with an explicit rule
- Reproducible pipeline: Deterministic seeds, checksummed draws, pinned dependencies
- Verification with clear meaning: Target-rule pass on synthetic labels — stated in API
honesty/leakage_note - Durable product: Hall of Fame + shareable candidate dossiers
- Open artifacts: CSV, JSON, reproducibility notes
The Bigger Picture
Think of upg-strings as part of the Calabi-Yau research stack:
- Generation: Genetic algorithms or databases create / enumerate manifolds
- Search: upg-strings ranks promising candidates (← you are here)
- Analysis: CYTools computes detailed geometry
- Metrics: cymetric approximates Ricci-flat metrics
- Classification: Domain models check topological properties
upg-strings answers: "Which candidates should I open a dossier for?"
New here?
Browse the Hall of Fame, try Lookup, open the quintic dossier, or start with the ELI5 walkthrough. Live ranking on the home page is optional.
Draft note (honest ML + certificates + how to cite): Honest landscape dossiers (Markdown).