Forking is typed
Each primitive declares which STT-ANCHOR labels make it fire. The anchor taxonomy becomes the type system.
ACRA Insight field unit / typed recursion issue
MoEA Loop
A typed rewrite system that drives three research skills into a self-extending, auditable research tree.
Static showcase only. The Python CLI runs locally because the loop is stateful, resumable, and long-running. Manual mode is dependency-free; the Claude backend only needs the optional Anthropic package.
System position: the Loop is the orchestration layer of the MoEA stack — MoEA Labeler is the preference-data layer, and the skill arsenal is the distribution layer.
System position: the Loop is the orchestration layer of the MoEA stack — MoEA Labeler is the preference-data layer, and the skill arsenal is the distribution layer.
The loop / no mystery meat in the middle
Brief → anchors → deliverables → typed forks.
The orchestrator does not reimplement the skills. It invokes the Stage 1 XML optimizer and Stage 2 Semantic Triple Transformation as the source of truth, then holds tree state between runs.
- 01
brief
One seed topic plus the tension worth researching.
- 02
Stage 1 XML
The gemini-deep-research-xml skill turns the brief into an anchored research prompt.
- 03
anchored deep research
Gemini, Grok, or ChatGPT Deep Research returns STT-ANCHOR blocks.
- 04
Stage 2 STT
Four deliverables: longform, infographic JSON, social copy, and an artifact prompt.
- 05
typed fork
One deepen primitive plus one diverge primitive are selected from the anchor set.
- 06
recurse
The tree expands until the frontier saturates or the budget runs out.
The hard problem, solved
Repeatable follow-ups without turning every topic into the same topic.
The fixed function is the selection logic. The adaptive part is the fork set it chooses from the anchor labels already present in the research.
Selection is paired
The loop returns one best DEEPEN plus one best DIVERGE, so the two children are complementary by construction.
Novelty is gated
A ledger scores each projected brief against prior theses. Saturated branches terminate instead of restating the tree.
Every fork is a typed rewrite rule with a provenance ledger — the tree is auditable, not vibes.
The engine is skill-agnostic / three terminals, one loop
Same typed forks. Different terminal deliverable.
The orchestrator now drives three skills, not two. Stage 3 swaps the terminal step without touching the fork logic: a research tree becomes a documentation tree. Same selection function, same novelty ledger, same provenance.
gemini-deep-research-xml
Turns the brief into an anchored XML research prompt. Every structural thesis point carries an STT-ANCHOR the downstream stages parse.
Download skill →semantic-triple-transformation
Normalizes anchored research into the editorial deliverables: longform, infographic JSON, social copy, artifact prompt.
Download skill →agentic-dev-docs
Swaps the terminal: anchored research becomes a publication-ready Mintlify MDX page that gates its own output against a Verification Gate.
Download skill →New in v0.3: agentic-dev-docs turns anchored research into Upsun-grade Mintlify MDX that gates its own output. See the full arsenal at /book/skills.
The primitive library / seven ways out
Forks are not vibes. They are typed rewrite rules.
Each primitive is lifted from the README and classified as either convergent pressure or lateral reframing.
| id | role | lens |
|---|---|---|
| ANTITHESIS | deepen | Feynman steelman-the-objection |
| DEBATE | deepen | AI-safety-via-debate / Oxford two-sided |
| DEPTH | deepen | drill one named open question to the studs |
| HOLOGRAPHIC | diverge | Maldacena boundary/bulk reframe |
| ANALOGICAL | diverge | cross-domain isomorphism transplant |
| VERTICAL | diverge | applied instantiation (sector / GTM / product) |
| POWER | diverge | coalition / incentive map |
Local loop / paste, ingest, branch, repeat
Manual mode matches the Gemini-by-hand workflow.
It prints paste-ready skill blocks, ingests the research files you save back, and keeps the tree plus novelty ledger on disk.
python3 moea_loop.py init --brief "Your topic + tension" --domain neuro
python3 moea_loop.py prompt 0
python3 moea_loop.py ingest 0 deep-research-report.md
python3 moea_loop.py fork 0The autonomy ladder / how far the human steps back
In-loop → on-loop → out-of-loop.
The same gate runs at every rung. What changes is who has to say yes. The autonomous runner, moea/moea_auto.py, drives the CLI and only proceeds on a clean Verification Gate.
in-loopYou approve at every gate. The runner prints the diff; nothing proceeds without a human yes.
on-loopAuto-proceed on a clean Verification Gate; a failed gate is queued for review instead of shipped.
out-of-loopPASS deploys, FAIL reflexes back into generation. The human sees exceptions only — the audit log holds the rest.
Start at --autonomy gated; promote to on-loop, then auto --deploy, only once the false-pass rate sits at zero. A failed gate is the corrective signal that re-enters generation.
Related field notes / the same thesis, three lenses