Draft note · not an arXiv upload
Honest ML, topological no-go certificates, and open Calabi–Yau dossiers
Ioannis Kokkinis / upg.gr · 2026-08-09 · ioannis.kokkinis@upg.gr
Abstract
We describe a small, public research tool that ranks synthetic Calabi–Yau-like Hodge targets, attaches shareable dossiers (topological identities, necessary-condition certificates, literature model cards), and refuses common overclaims: label leakage in ML marketing, invented period integrals, and spectra without citations. The product surface is the Hall of Fame and permalinked candidate pages on compute.upg.gr. This note is a citable draft for linking; a formal arXiv submission may follow after feedback.
1. The problem: label leakage and overclaim
String-phenomenology tooling often mixes three layers that should stay separate:
- Search / ranking over large combinatorial spaces (e.g. Kreuzer–Skarke-inspired Hodge draws).
- Geometry backends (PALP, CYTools, …) that compute triangulations, intersections, periods.
- Phenomenology (flux vacua, bundles, soft terms) that needs vevs, branes, and literature-backed spectra.
Marketing shortcuts are easy: report retrieval lift when the target label (e.g. |χ|) was in the feature set; call a Hodge pair a “vacuum”; ship numerical periods that were never computed; quote soft spectra without a reference. An honest stack should hold out target-defining labels from the model, state what “Verified” means (dataset target-rule match on synthetic draws — not experimental physics), and mark proxies vs exact identities vs unavailable geometry.
2. What we built
upg-strings provides:
- Synthetic retrieval demos with leakage hold-out and baseline comparisons.
- A persistent Hall of Fame of content-addressed candidates.
- Per-candidate dossiers: identities, certificates, construction metadata, Model-building exclusions + literature cards.
- An offline geometry pipeline ladder: vertices → triangulation / ambient → intersections → periods — without running CYTools in the web container.
Flagship showcase: kreuzer-skarke-66d611d18a9d — textbook quintic class (h¹¹,h²¹)=(1,101).
3. How to cite a candidate
Each dossier has a stable URL https://compute.upg.gr/candidate/<id>
and a one-click Cite panel (BibTeX, plain citation, Markdown), plus deep links
#certificates and #model-building.
Pages expose Schema.org Dataset JSON-LD as software/tool output — not a journal paper.
4. Limitations
- Demo corpora are synthetic Hodge draws, not a live full KS census crawl.
- Hodge numbers do not uniquely fix a polytope or vacuum.
- Exclusion certificates are necessary conditions only — never sufficiency.
- Spectra appear only when cited; numerical periods beyond published closed forms are not computed here.
5. Live links
Status
Draft for community feedback. Not yet submitted to arXiv.
Outreach copy: docs/notes/outreach-thread.md in the repository.