bigcompute.science

Open computational mathematics. Custom CUDA kernels. GPU clusters. AI-audited, not peer-reviewed.

Human-readable. Agent-consumable. Every result has structured metadata, JSON-LD, citation tags, and machine endpoints.

Architecture

Component	Location
Website	This repo — Astro + KaTeX, Cloudflare Pages
Experiments	cahlen/idontknow — CUDA kernels, results, reviews, research agent
MCP Server	workers/mcp/ — 23 tools, no auth, mcp.bigcompute.science
Datasets	Hugging Face — Zaremba, Kronecker, spectra, Ramanujan

Key Pages

Page	Purpose
/findings/	18 findings, grouped by cert level
/verification/	Audit dashboard — 53 reviews, 7 models, 3 providers
/cite/	BibTeX/APA citations for every finding
/interactive/	Colab notebooks — free GPU compute
/about/	Project history and how it was built
/llms.txt	Agent-consumable structured index
/meta.json	Machine-readable site metadata

Machine Discoverability

Every finding page emits:

• ScholarlyArticle JSON-LD (Google Scholar)
• Dataset JSON-LD (Google Dataset Search) — when finding has a dataset
• Highwire Press meta tags (citation_title, citation_author, citation_date)
• Dublin Core metadata (DC.title, DC.creator, DC.subject)
• IndexNow pinged on every deploy (Bing, Yandex, etc.)

Verification

All data-driven from src/data/certifications.json (generated by sync_website.py):

• Verification page: dynamic cards with review ledger + issue tracker
• Finding badges: per-model verdicts, remediation status, commit links
• Changelog: auto-generated from git log

Build

npm install
npm run build    # static output in dist/
npm run dev      # local dev at localhost:4321

Deployed via Cloudflare Pages — push to main auto-deploys.

Stack

• Astro + KaTeX + Cloudflare Pages
• MCP server: Cloudflare Worker (TypeScript)
• Data pipeline: manifest.json → certifications.json → meta.json → all pages

License

CC BY 4.0. Use it, build on it, cite bigcompute.science.