course: module 11 -- AI numbers, the MX block (lessons 31-33)

[gHashTag]

What

Continue the course at t27.ai/#/course with a module on the number formats AI chips use: module 11, "AI numbers: the MX block", lessons 31-33, after the lab module.

#	lesson	widget	what it shows
31	One scale for a block of weights	play-e8m0	the player compiles e8m0.t27 (the OCP MX shared scale byte) on seven backends in the browser and runs its tests; it skips what the browser cannot run and says why
32	The scale byte on a real machine	t27c-e8m0	the same spec through native t27c and Zig on the t27c lab: 18 tests pass, 9 invariants proved at compile time
33	One outlier, many zeros	t27c-ocp-mx (+ t27-vs-nvfp4)	ocp_mx.t27 on a block of 31 small weights and one outlier: how many each element format flushes to zero; one planted bias bug fails exactly one test

Rules it follows

• Words and order live in specs/course/course.t27; the generators write the TypeScript, pages and cards.
• The lessons quote only widget output, arithmetic, and constants of OCP MX v1.0 (block of 32; 4-, 6- and 8-bit elements; E8M0 code 255 = NaN, bias 127).
• The core stays 27 lessons (3 x 3 x 3); the lab and the MX module are counted apart, and the spec's own test says so.

Sources

• OCP Microscaling Formats (MX) Specification v1.0.
• feat(numeric): ocp_mx.t27 -- OCP MX v1.0 and OFP8 elements, E8M0 scale, the block conversion t27#6828 (specs/numeric/ocp_mx.t27, open) and specs/numeric/e8m0.t27.
• arXiv:2606.05017 (GoldenFloat), arXiv:2606.09686 (Golden Ruler).

Part of the widgets epic #1377.