ML theory library
PublishedA connected library of ML, statistics, probability, optimization, and frontier-model topics with prerequisites, examples, failure modes, exercises, and references.
Browse topicsProjects
TheoremPath is a graph-backed ML research learning system.
The public project is not one page or one app. It is the topic library, prerequisite graph, evidence ledger, learner loop, practice standard, and iOS companion moving through content-depth, learner-loop, evidence, and release gates.
Last reviewed: July 4, 2026
Atlas graph
LiveA prerequisite graph for seeing how topics depend on one another and for tracing paths from foundations to frontier reading.
Open AtlasEvidence dashboard
Live receiptsA public boundary page for source grounding, Lean mappings, diagnostic links, trail coverage, learner-loop receipts, and sparse calibration status.
Open evidencePractice standard
PublishedA standard for turning study into code, derivations, baselines, ablations, plots, paper maps, and short technical reports.
View standardDiagnostic and review loop
BetaSigned-in diagnostics, saved topics, review entry points, and profile state are wired. Effectiveness and item calibration still need more real learner data.
Run diagnosticLearner-loop and iOS continuity
Pre-TestFlightThe companion app has signed-in API continuity receipts and simulator captures. TestFlight upload, privacy metadata, and final screenshots remain separate release gates.
See receiptsFlagship ML pages
The current content work is centered on a small set of high-value ML pages. Each should connect theory, examples, failure modes, and evidence boundaries.
Evidence boundary
The project is strongest when each claim names its evidence type. Working software, source mappings, formal wrappers, and signed-in smoke receipts are different kinds of proof. They should not be blended into a single vague quality claim.
| Layer | Boundary |
|---|---|
| Implemented | Topic graph, governed claims, source locators, Q-matrix rows, diagnostic items, FSRS-style review state, signed-in state, saved topics, review endpoints, and iOS API continuity. |
| Measured today | Route coverage, production smoke receipts, learner-loop API reads/writes, proof-asset captures, and live sparse diagnostic rows. |
| Not claimed yet | Retention lift, calibrated IRT ability estimates, item discrimination, broad PFA effectiveness, causal mastery proof from the Q-matrix, or a public claim that every page is Lean verified. |
What comes next
The launch candidate should keep moving on content depth, evidence integrity, signed-in flows, iOS readiness, and a demo packet built from the real product.
- Collect enough real signed-in attempts for item calibration rather than fitting sparse data.
- Keep flagship ML pages source-backed with examples, failure modes, and diagrams.
- Re-run web and iOS signed-in continuity after auth, learner-state, or release-shell changes.
- Prepare a product demo packet from real screenshots and flows, not staged claims.