Doing well
✓Voice (4.8) — the distinctive editorial prose carries intact across all twelve lessons; the sport changed but the writing didn't slip.
✓Technical accuracy (4.7) — verified programmatically: every script parses, the match-stats MCP server is valid Python, and the evals figures (precision/recall/F1) are identical to source.
✓Metaphor fit (4.6) — the VAR analogy maps the subject unusually cleanly: GOAL / NO GOAL for the binary decision, real instruments (goal-line tech, semi-automated offside, multi-angle replay) for tool-calling, and actual match footage for retrieval.
✓Learner activation (4.1 — up from 3.3) — every lesson now closes with a predict-then-reveal Quick check, turning passive reading into active recall. Plus all 12 lessons + chrome are built and every link resolves.
To improve
✗Pacing (4.0) — the standout gap now. MCP (L9) runs to 7 pages while L6, L8 and L12 are single pages; evals (L11) still spikes on jargon. Normalise the heavy lessons.
✗Deeper interactivity — the Quick checks lift recall, but there's still no editable or runnable code and no multi-question quiz. That's the next lever beyond a single check per lesson.
✗Visual motion (L10) — the agent-loop diagram is static (a page about a loop that doesn't loop); a couple of other redrawn scenes are functional but plain.
✗Housekeeping — the internal sv_ progress keys keep the cricket course's numbering offset (unique and working, just not renumbered 1–12).