Learn Center v3: the binary feature checklist (the gate)

GATE RESULT (2026-06-18): ALL GREEN

Every item below CONFIRMED. Decorrelated panel: Gemini (live site) verified all 21 items present + 0 factual errors + self-fetched all 75 links 200; Codex (gpt-5.5, workspace copy; its sandbox DNS blocked the live fetch so it honestly marked live/links CANNOT-VERIFY rather than guessing) confirmed all 21 in source/data and surfaced 4 real accuracy catches. All 4 fixed (commit 3411fa7) + adversarially re-verified by a 3-verifier panel (0 major; 1 minor tightened, commit 470a4a5). Ground truth held throughout: all 76 site URLs curl-200, JSON valid, node-check clean, em-dash-free, Pages deploy succeeded. The binary checklist replaced the noise-dominated 0-100 rubric and the loop converged cleanly (a real defect class found and closed, not a random walk).

v3 replaces the noise-dominated 0-100 rubric (see BUILD_STATE.md plateau finding) with a fixed, binary per-feature gate. Each item is either DONE (present + verified working) or NOT. The build is complete when every box is checked and independently verified by the decorrelated rater panel (Codex gpt-5.5 + Gemini CLI). No fuzzy scores; “present and works” or not.

Doer = Claude (Opus). Verifiers = Codex + Gemini, each checking items against the LIVE site and the source. A box is only checked when the doer has verified it AND at least one rater confirms it (or a ground-truth check – curl, node –check, live preview – confirms it when raters disagree).

A. New interactive labs (engine widgets)

A1. Bootstrap CI simulator – resample a small score set, show the distribution of the mean and a 95% percentile interval that narrows with more resamples. Teaches why a point estimate needs an interval. Wired into an evaluating-ai lesson.
A2. ROC / PR threshold slider – a labeled-example set with scores; a draggable threshold updates TPR/FPR (and precision/recall) live and plots the operating point on an ROC curve. Teaches threshold choice and the precision-recall tradeoff. Wired into an evaluating-ai lesson.
A3. RAG retrieval / chunking sandbox – a short corpus, adjustable chunk size and top-k; a query retrieves the closest chunks (simple term-overlap score) and shows what would land in the prompt. Teaches chunking and retrieval tradeoffs. Wired into an agents lesson.
A4. Context-window budget visualizer – sliders for system / history / retrieved / output reservations against a fixed window; shows what overflows and gets truncated. Teaches context budgeting. Wired into a foundations or agents lesson.

Each lab: renders, is interactive (live recompute on input), math is correct (verified against a hand calc), has a stated learning goal, and clears stale feedback on re-input (the v2 metric_lab bug pattern).

B. Capstones (one per track)

B1. Foundations capstone – build + evaluate a tiny classifier (walk train/dev/test + read a confusion matrix to a decision). Multi-step, synthesizes the track.
B2. Agents capstone – debug a multi-step agent trace end to end (find the fault, choose the fix, justify it).
B3. Vibe-coding capstone – review an AI-generated patch (spot the bug, the missing test, the security issue; decide merge/revise/reject).
B4. Evaluating-AI capstone – design an eval for a stated capability (pick metric, build a rubric, choose a baseline, name the failure mode you would miss).
B5. Frontier capstone – read a frontier paper abstract + claim and judge it (what is measured, what is overclaimed, what evidence is missing).

Each capstone: is a distinct lesson, marked as a capstone, is genuinely multi-step (not a single MCQ), uses a real interactive exercise, and ends the track’s next chain.

C. Breadth lessons

C1. F1 / precision-recall / AUC – the metric family, when each is right, what AUC means.
C2. Statistical significance / p-values / multiple comparisons – what a p-value is and is not, and the multiple-comparisons trap in eval.
C3. LoRA / PEFT / quantization – parameter-efficient finetuning and quantization tradeoffs.
C4. Transformer internals – residual connections, normalization, positional encodings.
C5. BPE tokenization (detail) – how byte-pair merges build a vocab; why token counts surprise.
C6. Privacy / copyright / fairness – training-data provenance, PII, fairness failure modes.
C7. MLOps / data pipelines / deployment monitoring – the lifecycle around a deployed model.

Each breadth lesson: accurate (no factual errors), em-dash-free, level-labeled, has key points + verified links (curl 200) + flashcards + a quiz + an exercise, and is threaded into its track’s nav.

D1. Global index / tag view – a single route listing every lesson, filterable by track and level (and/or tag), each a direct link. Bookmarkable.
D2. Resume / continue learning – the landing surfaces the last-viewed (or next incomplete) lesson as a one-click “continue” affordance.
D3. Per-track mastery view – each track shows progress (lessons complete / total, and quiz pass state) at a glance.
D4. Clickable per-claim source links with last-verified dates – source metadata extended beyond frontier lessons; rendered as clickable links with a visible last-checked date.

E. Content hygiene

E1. Dedup the near-duplicate quiz/exercise pairs flagged in pedagogy review; no two lessons share a near-identical question.
E2. Global invariants hold: all JSON parses; tracks.json order matches each track file’s ids; all prereqs/next resolve; zero em-dashes across learn/; every link curl-200; learn.js node --check clean; live Pages deploy serves the new content.

Verification protocol

For each item: (1) doer builds + self-verifies (preview, node –check, curl, hand-calc the math); (2) commit + push live; (3) Codex + Gemini each get the checklist + the live URL and report, per item, CONFIRMED / NOT / CANNOT-VERIFY with evidence; (4) any NOT gets fixed; (5) repeat until every box is CONFIRMED by the doer AND corroborated (rater or ground-truth). Disagreements resolved by ground truth (curl / node / live preview), never by re-running the score.