Learn Center: v2 roadmap (the gap-to-88 ledger)
This is the honest “not-measured / not-yet-built” ledger for the Learn Center, in the spirit of the Evaluating AI Honestly track. The v1 build was rated by a decorrelated Codex + Gemini panel across four rounds; both raters agree the content is accurate and the prose is strong, and both place the remaining gap to a 90s score in product depth, not correctness. Those depth items are a deliberate v2, scoped here.
Where v1 landed
- 49 lessons across 5 tracks (Foundations, Agents, Vibe Coding, Evaluating AI, Frontier), every lesson with key points, verified links, flashcards, a quiz, and an interactive exercise.
- 47-entry resource library (curated + aggregated lesson links), filterable, with a Start-here path.
- Search, flashcards (with completion + spaced review), quizzes (auto-complete on a perfect score), and four active exercise types. Honest source metadata on frontier lessons.
- Both raters: no factual errors in the prose; dense, level-labeled, non-patronizing.
What would close the gap to a 90s score (v2)
- Richer interactive widgets (highest leverage). The current exercises are MCQ-shaped (predict_output / prompt_repair / spot_the_bug / agent_trace). Add applied formats that need new engine widgets: a confusion-matrix / discrimination calculator, a temperature-sampling demo, a tokenizer playground, an embedding-distance widget, and ordering tasks. This is the single most-requested upgrade from both raters.
- Diagrams and visual learning. The site reads like dense notes. Add inline SVG diagrams for
tokens, attention, the context window, the RAG retrieve-augment-generate loop, the agent loop,
and a confusion matrix. Wire the existing
tracks.jsoniconids to real SVGs. - More breadth. Classical ML (supervised vs unsupervised, train/dev/test, core metrics), LLM-as-judge rubric design, inter-rater reliability, red-teaming, tool permissioning, current benchmark families, and production cost/latency tradeoffs.
- Navigation depth. A persistent lesson sidebar, prev/next controls in the lesson view, level/track filters inside search, and a “resume where I left off” affordance.
- Link hygiene at scale. Add
last_checkedmetadata per link and a CI link-checker so the library ages honestly; widen sources so the same few links are not reused across lessons. - Source discipline everywhere. Extend the
sourcemetadata beyond frontier lessons to the agent-security, eval-methodology, and vibe-coding claims that age quickly.
Known limitations (carried, not hidden)
- Exercises are recognition/diagnosis MCQs, not yet applied calculators or builders.
- No diagrams yet; visual learners are underserved.
- Lesson “complete” is auto-set on a perfect quiz but is otherwise a manual toggle.
- The site is intentionally served from the hub root, so a learn-rooted local server cannot
resolve the absolute
/assets/research-notes.css; this is correct for the GitHub Pages deployment, not a production bug.
v2 shipped (2026-06-18) + where it landed
v2 added the depth both raters asked for: three applied interactive widgets (a live confusion-matrix / discrimination calculator, a temperature->softmax visualizer, ordering tasks), six inline SVG diagrams, and four breadth lessons (classical ML, LLM-as-judge, inter-rater reliability, production cost/latency). 49 -> 53 lessons. The goal-loop ran two more rounds: Codex 78->82->80, Gemini 83->86->84.
The loop then plateaued: a re-rate of a strictly-improved build moved both scores down, which means the rubric scoring is noise-dominated at this quality level. Exited per anti-inflation discipline. Both raters confirm the content is accurate and dense throughout; the gap to a stable 88+ is structural.
v3 (would push past the mid-80s plateau)
- More applied labs: a bootstrap confidence-interval simulator, an ROC/PR threshold slider, a RAG chunking/retrieval sandbox, a prompt-injection permission sandbox, a context-window budget visualizer.
- One capstone per track (build+evaluate a tiny classifier, debug an agent trace, review an AI patch, design an eval, read a frontier paper).
- Breadth: F1/AUC/PR curves, statistical significance, LoRA/PEFT, quantization, residual connections, normalization, positional encodings, BPE tokenization, privacy/copyright/fairness, MLOps/data pipelines.
- Source discipline: clickable per-claim source links with last-verified dates, beyond frontier lessons.
- Navigation: a global index/tag view, per-track mastery, “continue learning,” randomized quiz mode.
- Dedup the few near-duplicate quiz/exercise pairs flagged in pedagogy review.