Internals

Internals

How the repository is built and how to reproduce a run.

flowchart LR
  lib[lib: domain, dense, moe, recur, engram, scale, bench, budget, svg, results] --> demos[demos: one lesson each]
  demos --> previews[assets/previews: annotated SVG diagrams]
  demos --> fixtures[fixtures: domain, teacher, recordings, scale, benchmark]
  demos --> results[docs/reference/results.md]
  scripts[scripts/check: the gate] --> demos
  serve[mlpl-serve: recordings for the generic host] --> fixtures

Loading

• Architecture: the model under the lens, the four sparsities, ownership and layering
• Delivery plan: every saga and step, the two scales, the visual and measurement contract
• Saga queue: what is done, active, planned
• Implementation index
• Host handoff: the demo-extensions work order for the recordings
• Cross-repo handoffs
• Teacher fixture schema
• Development rules: the agentrail protocol and project rules
• sw-MLPL: the language and runtime every lesson is written in; the adjacent checkout builds the binary the gate uses
• Live demo: learn/ holds two static playback pages (frame by frame, and the panned landscape driven by fixtures/landscape/scenes-v0.json) published by the Pages workflow from scripts/build-site; just serve-learn serves them locally; scripts/check-landscape checks every scene against its fixtures

Reproducing

just check      # the whole gate: structure, links, style, tests, every demo, recordings
just dense      # one lesson (also: moe, moe2, dispatch, delta, scale, teacher, budget, benchmark, recur, rm, engram, re)

Each lesson has a check mode used by the gate and a write mode that regenerates its diagrams and results row. Lessons with several variants run one training per interpreter process (finding F22) and compose any shared diagram from committed metrics. Timings never appear in diagrams, so regenerated SVGs are byte-stable.

Language-level findings met along the way: Reference.