# Sources and implementation notes

## Research evidence

Bence Lukacsy, CSE 495 Research Proposal; CSE 496 Progress Report; historical research Reports 1–3; McKinnon–Rosati Laboratory repository Project 1; frozen v2 follow-up protocol and completed THOROUGH_RESULTS.zip. The study is a follow-up informed by earlier results, not a pristine confirmatory benchmark.

The two-page summary and Beamer presentation on this site report the completed experiments. Primary comparisons are in primary_comparisons.csv. Original reference-label agreement is not independent biological validation; optimization seeds are not biological replicates. Native expression conventions remain a limitation in cross-dataset inference.

## Deployed inference

The browser code reproduces the evaluated three-seed rank-based mouse router. It does not substitute the router's generic tuning-selected kNN candidate. Child methods are tuning-selected: pooled-lung large-flat ensemble; masked-flat ensembles for brain and fish. An optional child kNN mode uses its own recorded calibration and gates; the mouse router stays fixed. Other architectures were researched but are not silently substituted in the UI.

Gene names use explicit unambiguous mouse identifier associations. Missing selected genes use the reference's saved native mean before standardization. The minimum selected-gene overlap is 90%. The empirical reference-support score combines saved reference-relative signals and is not P(tissue). Unknown-type experiments demonstrate that these thresholds do not reliably reject every unseen type. Rejection of a malformed or incompatible panel is only an input-contract check.

Data/model artifacts are provided for this research prototype subject to the source and laboratory permissions. No blanket redistribution license for the datasets is asserted. Public weights, reference PCA vectors and labels are downloadable; query expression values are not transmitted to a prediction endpoint.

## Visualization

The default display is an exact first-two-axis PCA projection. When the display-only UMAP build is run, UMAP is fitted on 2,000 saved training-reference PCA points. Query placement is an approximate eight-neighbor inverse-distance barycenter on that map, NOT UMAP.transform. No query-fitted map is used to claim classification success. Each reference family has a separate coordinate system.

Official UMAP transform documentation: https://umap-learn.readthedocs.io/en/latest/transform.html

## Website dependency

Plotly.js 3.3.1, distributed under MIT; its license is in assets/PLOTLY_LICENSE.txt. All JavaScript and plotting assets are local; the application includes no analytics or third-party inference calls.

## Static hosting

Cloudflare Pages Direct Upload, custom domains, platform limits and static-asset pricing:
- https://developers.cloudflare.com/pages/get-started/direct-upload/
- https://developers.cloudflare.com/pages/configuration/custom-domains/
- https://developers.cloudflare.com/pages/platform/limits/
- https://developers.cloudflare.com/pages/functions/pricing/

The application has no Pages Functions or server-side query API. Hosting providers still see ordinary page/model-download requests. Source checked 15 November 2026.
