Doing well
✓Activation (4.65) — interaction stays high across the new stretch: L7 plays each of the three nearest takes and walks vector arithmetic (Sax − Jazz + Blues), L8 lets you hear the attention weights and slide a note into the "lost middle", and L10 flips custom cables vs one universal socket while sliders make the n×m → n+m gap jump out — on top of L6's quiz and stakes meter, and now L3/L4's staff, where a playhead sweeps the notes as they sound.
✓Depth (4.59) — the new lessons go past the headline: L7 covers embeddings, meaning-space arithmetic and chunking; L8 reaches multi-head, long-range "lost in the middle", and the quadratic cost of attention; L9 nails the request-vs-run distinction that trips people up on tool calling.
✓Audio (4.57) — still the signature: L7's three nearest takes each have their own motif, L8 rings the weighted blend so you hear the emphasis shift, L9 sounds the metronome / E♭ chord / a real transpose, and L10 snaps a clean major chord as the universal socket clicks in. The withAudio guard means no first-click note drops.
✓Voice (4.79) & Accuracy (4.77) — warm improviser prose across all 37 pages, now through a full version-2 editorial pass that cut repetition, plain-languaged the last of the insider musician terms, and added one sharper real-LLM point per page; every script parses and the maths is verified (softmax, softmax-with-temperature, top-k/top-p, L6's argmax-invariance, and L8's n² comparison counts), with step-by-step worked examples.
To improve
✗Visual richness (4.46 — up from 4.39) — L3 and L4 now draw the phrase itself on a five-line staff, with noteheads, accidentals and a playhead that sweeps across as each note lights and sounds — the notated staff and animated motion the interactives were missing. Elsewhere the diagrams help (L7's meaning-map and vector-arithmetic, L8's "lost in the middle" curve and n² grid, L10's wiring SVG), but most other pages still lean on bars / chips / sliders where the audio carries more than the picture — the staff wants rolling out to the other note-level lessons next.
✗Accessibility of sound (4.58) — legible panel type, a shared scale, and the v2 pass clearing out insider musician terms all help, and several pages lean on diagrams and text as much as audio. But meaning still lives partly in what you hear; the aria/text equivalents must stay strong for anyone who can't play the audio.
✗Coverage — 12 of 15 lessons are built. Retrieval, attention, tools, MCP, agents and mixture-of-experts are done; still ahead are evals, fine-tuning and RLHF — several (evals especially) the music metaphor will have to be redesigned for, not translated.
✗Depth balance — L9 (tool calling), L10 (MCP) and L11 (agents) are strong but single-page, lighter than their multi-page siblings (L7 at 5, L8 at 4). Each could grow a page or two to match — e.g. a build-a-server walk for MCP, or a multi-tool trace for tool calling.