README.md

Live pages — plan, look up, test, present

Five self-contained pages published with the docs site, sharing one nav so they read as a single tool. All are plain HTML with no framework, no CDN, and no build step, so they work on GitHub Pages, from a local checkout, and on a locked-down corporate network alike.

Page	What it does	Open it
Study planner	Pick a goal, get an ordered module ladder, tick modules off, watch remaining hours drop	dashboard.html
Reference sheets	Track, concept and module cheatsheets — searchable, printable, every fact cited	cheatsheets.html
Recall cards	Retrieval practice on the curriculum's own Q&A, sorted into three Leitner boxes, every card cited	flashcards.html
Slide deck	Presentable overview of the curriculum and the findings from building it	slides.html
Concept players	Twelve interactive versions of the infographics, one per concept cluster, where a control is what teaches	concepts/index.html

They form a loop: plan what to study, look up what you forgot mid-lab, test whether it stuck, play with the one control that makes a concept click, and present it to find out what you only half-know. Every surface links to the other three: a concept player's footer points at its reference sheet and its recall cards; a reference sheet's cluster card points at its concept player and its recall cards; a recall card's track filter matches the same track colours everywhere.

On GitHub these render as source, not as pages. GitHub does not serve HTML from a repository view. Use the published site (the Deploy docs to GitHub Pages workflow builds it on every push to main), or clone the repo and open the files directly in a browser — both work.

---

Study planner

The planner is the in-repo counterpart to the Cursor canvas dashboard, rebuilt so it works anywhere a browser does.

• Five goal-based paths — write fast kernels, make models fast, design ML systems, interview soon, or everything in dependency order. Choosing one reorders the ladder and recomputes the hours.
• Progress that persists. Ticks are stored in localStorage under ppc.progress.v1, so closing the tab does not lose your place. Nothing is uploaded and there is no account.
• A "next up" recommendation — the first module on your path you have not finished, with its hour estimate, prerequisite depth, and a direct link.
• Per-track bars showing completed versus remaining hours for Tracks A, B, and C.
• Reset clears your ticks after a confirmation.

It degrades honestly: with JavaScript disabled it says so and points at CURRICULUM.md, which needs nothing but a browser.

Recall cards

Every module's Section 9 already holds conceptual Q&A, but it sits inside <details> blocks in a 900-line README — so it is read passively, if at all. This page turns that material into retrieval practice, which is a measurably different activity: Roediger & Karpicke found re-testing produced 61% recall after a week against 40% for re-reading the same material the same number of times, while the re-readers felt more confident (Test-Enhanced Learning, Psychological Science, 2006). Confidence is not retention, which is the same argument the curriculum makes about benchmarks.

• Three Leitner boxes, not a streak. Cards start in Box 1. Solid promotes, Shaky holds, Missed drops to Box 1. Drawing is weighted 3:2:1 toward Box 1, so the deck keeps returning to what you get wrong. The box display is the progress meter — the useful question is "what is still in Box 1", not "how many did I click".
• The answer reveals in place; the question stays. Deliberately not a flip card, which hides the question at exactly the moment you want to compare it against the answer.
• Every card cites a source — a paper, a vendor document, or the module it came from. If a card and its source disagree, the source wins.
• Listen to a card. L (or the button) reads it aloud using the browser's own speech synthesis — the question only, until you reveal, so it never spoils a card you have not attempted. Useful for testing yourself away from the screen.
• Keyboard: space reveals, 1/2/3 grade, → skips, L listens.
• Filter by track; progress persists in localStorage under ppc.cards.v1.

Deliberately a three-box Leitner system rather than a full SM-2 scheduler: interval tuning is not what stands between a learner and this material, and a simpler mechanic you can see is worth more than a better one you can't.

Reference sheets

Recall and reference are different activities, and merging them makes both worse. The cards ask you a question; these answer one you already have, mid-lab, without making you re-read a 900-line module.

• Four tiers, four questions. A track sheet says what that part of the curriculum owns and what you need before starting it. A concept sheet crosses modules, because the concepts do — twelve of them, from memory hierarchy to fabric tiers. A module card is the "what was the one idea in A04 again" answer, one per module. An analogy card gives the everyday version and then the machine's, set in two different typefaces so you can see at a glance which world you are in.
• Every analogy comes with its own limits. The card states where the module says the picture stops being true, because that is the half people skip and the half the engineering lives in. Two modules, A01 and A02, do not state one yet; those cards say so rather than implying the analogy is airtight.
• Every bullet is cited to the module section or glossary entry that derives it, and the numbers are the module's numbers. Nothing here is restated from memory.
• Search and filter. Type to search every sheet at once; / focuses the box and Escape clears it. Track pills narrow to one track.
• The rail is load-bearing. Thirty lanes across the top, one per module, coloured by track — the same "a square is a unit" device the infographics use. Filtering dims the lanes that drop out, so it always answers "how much of the curriculum am I looking at". Click a lane to jump to its card.
• Printable. Ctrl-P gives one concept sheet per landscape page with the citation URLs spelled out, so the printout stands on its own.

The facts live in data/reference.json, not inside the HTML, so they have one home and can be checked mechanically — every bullet must carry a cite. The cost of that choice: browsers refuse to fetch() a sibling file from a page opened with file://, so this one page needs an HTTP server locally (python -m http.server from the repo root). It says so on screen, with the command, rather than rendering blank. The published site is unaffected.

Concept players

One page per idea where a control teaches better than a fixed drawing — concepts/index.html is the hub, one tile per concept cluster, filterable by track and cross-linked to that cluster's reference sheet and recall cards. Twelve clusters, twelve players, all now built:

Concept	Modules	Player
Memory hierarchy & coalescing	A02, A06, B01	coalescing.html
Divergence & occupancy	A03, A04, A05	divergence.html
Fusion & FlashAttention traffic	A07, A08, A09	attention-traffic.html
Roofline & profiling	B02, B03	roofline.html
Precision, numerics & fp8	B04	precision-formats.html
Prefill, decode & serving	B05, B06	prefill-decode.html
Decision thresholds & cost	C01, C02	decision-threshold.html
Optimization ladder	C04, C05, C07	optimization-ladder.html
Drift & monitoring	C03, C06	drift-monitor.html
Scaling limits & cluster shape	D01, D05	scaling-limits.html
Ring all-reduce & collectives	D02, D03, B07	ring-allreduce.html
Interconnect & fabric tiers	D04, D06	fabric-tiers.html

concepts/coalescing.html was the first: change the access stride and the lockstep width, and watch the cache lines touched, the bytes fetched and the bytes wasted move with it. The efficiency turns out to be 1/stride on both a 32-lane warp and a 64-lane wavefront, which is a more convincing way to learn that than being told it. The other eleven follow the same rule and the same visual system (dark chrome, brand tokens from docs/BRAND.md, track-coloured accent bars): every reading is exact arithmetic on a formula the linked module already states and cites — never a new measurement, never a new claim — and every page says so plainly in its own disclosure note, the same convention the module's own analytic labs use.

The bar for adding one is that the geometry has to genuinely change with the input. If the picture is the same shape at every setting, it should stay an SVG in ../assets/, which remains the canonical version that the module embeds and that survives JavaScript being off — every concept player links back to its own canonical SVG for exactly that reason. The two clusters this repository had already flagged as earning a page, the ring all-reduce phases (D03) and the roofline intensity slider (B02), are both now built.

Slide deck

Fourteen slides following the slide template in docs/BRAND.md — 4px cyan accent bar, Inter and JetBrains Mono, dark dividers between parts, one idea per slide.

• Navigate with arrow keys, space, PageUp/PageDown, Home/End, or just scroll.
• Export to PDF with Ctrl-P / Cmd-P. Print CSS gives one slide per landscape page, so the deck doubles as a handout without a separate export step.
• Content: the design intent, the three tracks, the nine-section contract, the dual-vendor approach, the evidence standard, and five concrete findings from building the curriculum (the kernel that was 0.06% of its runtime, the 4.8× transpose, the two incompatible fp8 formats, why pipelining does not save activation memory, and the KV cache that was not the binding constraint).

Use it as-is for a talk, or fork a slide. The source is one file with the content inline as HTML — edit the <section class="slide"> blocks and it is still a deck.

---

Keeping them accurate

The planner and deck hardcode the module list, hour estimates, and depths. The source of truth is CURRICULUM.md; when a module is added or an estimate changes, update:

1. CURRICULUM.md — the status tables and the totals.
2. live/dashboard.html — the M array near the top of the script, and PATHS if the new module belongs on a learning path.
3. live/flashcards.html — add cards to CARDS, reusing an entry from the S source table or adding one. A card without a real source does not belong in the deck.
4. live/data/reference.json — add a modules entry, and a clusters entry if the module brings a genuinely new concept. Every bullet needs a cite. If the new cluster earns a concept player (the geometry genuinely changes with an input), add a concept field pointing at it — cheatsheets.html's cluster card wires the link automatically.
5. live/concepts/index.html — add a tile to the inline CLUSTERS array so the hub lists it. This mirrors reference.json's clusters array on purpose (AGENTS.md trap #7): each live page stands alone with no shared runtime, so the two lists are kept in step by hand.
6. live/slides.html — only if the headline counts (30 modules, ~203 hours) change.
7. The Cursor canvas at canvases/curriculum-dashboard.canvas.tsx, which carries the same data for use inside the editor.

The duplication is deliberate: each surface has to stand alone with no shared runtime. Five places is few enough to keep in step by hand, and a build step to generate them would be more machinery than the problem deserves.

Two things are deliberately not duplicated, and should stay that way. The reference sheets derive their per-track module counts and hour totals from the modules array rather than storing them again, and the recall deck builds its track filter row from the cards that actually exist. Both used to be hand-written lists; both were quietly wrong the moment Track D landed.

Known gap: the recall deck has 49 cards covering Tracks A, B and C, and none for Track D — it was built before that track existed. The reference sheets link to it per track (flashcards.html?track=D), and the deck says plainly that the track has no cards yet rather than showing an empty screen. Writing them is the fix; each needs a real source.