Experiments

Experiments

One laboratory report per experiment: the question, what changed from the previous experiment, the configuration, results, what was learned, what was not proved, and the raw evidence.

flowchart LR
  DM01[DM01 domain] --> DN01[DN01 dense]
  DN01 --> RT[MX01 routing]
  RT --> MX01[MX01 top-1]
  MX01 --> SD01[SD01 sparse dispatch]
  MX01 --> MX02[MX02 top-2]
  MX01 --> DS01[DS01 data scale]
  MX01 --> LD01[LD01 delta experts]
  DN01 --> TE01[TE01 teacher]
  LD01 --> RB01[RB01 budget]
  RB01 --> GB01[GB01 benchmark]
  DN01 --> RC01[RC01 recurrent block]
  RC01 --> RM01[RM01 recurrent MoE]
  MX01 --> RM01
  DN01 --> EG01[EG01 Engram]
  RM01 --> RE01[RE01 all three sparsities]
  EG01 --> RE01

Loading

Order	Report	Question
1	DM01 synthetic domain	What does the model see, and is the split honest?
2	DN01 dense baseline	How good is a tiny dense model?
3	MX01 routing	What does a router compute before training?
4	MX01 top-1 mixture	Can we store more than we execute at unchanged quality?
5	SD01 sparse dispatch	Does conditional compute avoid the unused work exactly?
6	MX02 top-2 and specialization	What does the second expert buy, and do experts specialize?
7	DS01 data scale	Does more data or more training fix low-data behavior?
8	LD01 low-rank delta experts	How many experts can we afford?
9	TE01 in-repo teacher	Is there a usable source of soft targets?
10	RB01 resource budget	How big is MicroMoE?
11	GB01 generation benchmark	How fast is it?
12	Recordings and host handoff	What does the generic host render, and what does the live demo play back?
13	CD00 campus snapshot A	What does the campus docent see, and is its corpus honest?
14	CD01 docent dense classifier	How well does a flat classifier predict intent and destination, and where does it fail?
15	MB01 matcher baseline	What does deterministic matching already achieve, and by how much must a docent beat it?
16	CD01b docent word vectors	Do word vectors from the campus's own text give the docent meaning its hashed features lack?
17	RC01 recurrent block	Does applying one block several times replace parameters?
18	RM01 recurrent mixture	Does expert selection change over reasoning time, and do experts specialize by recurrence?
19	EG01 Engram from scratch	Does external memory improve loss and tasks, at what table size, and is the hand-written form the builtin's function?
20	RE01 recurrent mixture plus Engram	Do memory sparsity and compute sparsity complement each other?

Index with planned experiments: docs/experiments/README.md. Machine-readable inventory: catalog/lessons.toml.

Conclusions drawn from these: Results.