Skip to content

Latest commit

 

History

History
155 lines (115 loc) · 5.87 KB

evolution-lab.md

File metadata and controls

155 lines (115 loc) · 5.87 KB

Evolution Lab integration

z0intelligence is the runtime/control plane: evidence-backed state, typed decisions, deterministic legal-action compilation, receipts, harness adapters, and promoted mechanism execution. Historical OpenJev code remains one backend/training substrate inside this repository.

Evolution Lab is the experiment/search engine. It compares candidate mechanisms against frozen evidence and outcome contracts, then exports credited artifacts back to z0intelligence. Fly/mushroom-body specialists are one candidate family, not a mandatory production layer.

z0intelligence runtime / receipts / replay
                    ↓
             Evolution Lab study
      rules · retrieval · linear · MB/fly
        local SLM · typed backend · ...
                    ↓
       promote / reject / deoptimize
                    ↓
      deploy credited mechanism back into z0int

Who owns what

Repo Role You use it when…
z0intelligence (this repo) Runtime/control plane: state, decisions, authority/receipts, compiler surfaces, backend experiments You need runtime cognition, context, routing evidence, replay, or promoted mechanisms
evolution-lab Evolution: genomes, locked splits, promotion ladder, DAgger, control tables You want to search for a fly recovery specialist or JEV head that beats controls
frontier-kb Claims vault / research notes You need protocol, kill criteria, or receipts
OMP extensions Live wiring (typesafe-jev, flyforge-recovery) You want token-saving decisions inside Cursor / OMP today

Evolution Lab may call the legacy OpenJev training backend for Track B genomes (backend=openjev). z0intelligence does not depend on Evolution Lab at runtime.

Quick setup (recommended)

From the z0intelligence repo root:

bash scripts/setup-flyforge.sh

This will:

  1. Create .venv (if missing) and pip install -e '.[test]' for z0intelligence
  2. Clone kvnloo/evolution-lab as a sibling (or use EVOLUTION_LAB_DIR)
  3. Install evolution-lab editable (numpy-only core; OpenJev already satisfies openjev_phase1)
  4. Lock P0 Hermes recovery splits and Jev Track B splits
  5. Print smoke commands

Environment overrides:

Variable Default Meaning
EVOLUTION_LAB_DIR ../evolution-lab Where to clone/install the lab
EVOLUTION_LAB_REF nightly Git branch to checkout
HF_HOME unset Hugging Face cache (set for Route B 4B models)
SKIP_EVOLUTION_LAB=1 — z0intelligence-only install (no clone)

Manual setup

# 1. z0intelligence (this repo)
python3 -m venv .venv && source .venv/bin/activate
pip install -e '.[test]'

# 2. Evolution Lab (sibling)
git clone https://github.com/kvnloo/evolution-lab.git ../evolution-lab
cd ../evolution-lab && git checkout nightly
pip install -e .
cd -  # back to z0intelligence

# 3. Lock experiment splits (deterministic, in git)
python -m evolution_lab lock-splits
python -m evolution_lab lock-jev-splits

For GPU Route A Jev training, the legacy OpenJev substrate is already installed from this tree. You do not need pip install -e '.[openjev]' on evolution-lab when this z0intelligence checkout is editable.

Two evolution tracks

Track A — Hermes recovery fly (local_plasticity)

Native NumPy in evolution-lab. 640-weight mushroom-body analogue for bounded recovery actions:

{retry, restart_sandbox, escalate, noop, page_human}

cd "$EVOLUTION_LAB_DIR"
python -m evolution_lab seed
python -m evolution_lab run --level 1
python -m evolution_lab table --level 1
python -m evolution_lab dagger-smoke --rounds 2
python -m evolution_lab recovery '{"action":"plan","event":{"kind":"tool_error","message":"timeout"}}'

Deploy path: OMP flyforge-recovery extension → cached local_plasticity student.

Track B — JEV routing heads (jev_tiny, jev_hf_head)

Uses this repo as the training backend (openjev_runner.py in evolution-lab).

cd "$EVOLUTION_LAB_DIR"
python -m evolution_lab seed-jev
python -m evolution_lab jev-smoke --level 0   # needs CUDA + torch from OpenJev

Deploy path: train checkpoint → openjev-predict / OMP typesafe-jev at inference time.

Export contract (lab → runtime)

Artifact Produced by Consumed by
P1 control table evolution_lab table Maintainers / receipts
Recovery advise JSON evolution_lab recovery / advise OMP flyforge-recovery
Jev checkpoint (.pt) openjev-train via lab genome openjev-predict, openjev-eval
Locked splits lock-splits, lock-jev-splits Reproducibility; do not re-sample

Weights are not checked into either repo by default. Reproducibility = genome JSON + locked splits + git SHA.

Token-saving production stack

Use local specialists before subscription LLMs:

1. recovery_advise (fly, 640w)     → tool/harness failures
2. typesafe-jev / openjev-predict  → route, clarify, skill-pick
3. subscription LLM              → only on escalate / low confidence

Smoke checklist

# OpenJev Route B zero-shot (needs GPU + HF cache)
CUDA_VISIBLE_DEVICES=0 openjev-score --mode direct \
  --model Qwen/Qwen3.5-4B \
  --revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a \
  --input examples/decisions.jsonl --output /tmp/results.jsonl

# Evolution Lab unit tests (CPU)
cd "$EVOLUTION_LAB_DIR" && python -m unittest discover -s tests -p 'test_*.py'

# Jev Track B via lab (GPU, OpenJev installed)
python -m evolution_lab jev-smoke --level 0

Sister links