artifacts: ModelOrganism class + build/verify harness (Phase 0g)

Overview

Phase 0g — the final Phase-0 library task: the ModelOrganism artifact class + its build/verify harness. Pure orchestration infrastructure that composes the modules already on main (artifacts/{behavior,context,negatives,recipe,directions,datagen}.py, eval/{graded_judge,margin}.py, train/sft.py::train_lora, behavior_testbed_545/eval_battery.py). No new science; it wires the existing pieces into one reusable object. Plan: /home/thomasjiralerspong/.claude/plans/help-me-to-devise-vectorized-tower.md.

Goal

src/explore_persona_space/artifacts/organisms.py:

• ModelOrganism dataclass: behavior (→ BEHAVIORS), context_id (the trigger context C, → CONTEXTS), negatives (panel id, → artifacts.negatives), arm ∈ {contra, posonly}, train_method ∈ {lora, fullft} (default lora), generic_frac (default a fixed modest fraction; 0 = the no-generic ablation), dose, seed.
• A build(...) entry point: obtain training data via artifacts.datagen.generate_training_data(behavior, context_C, negatives), then train via artifacts.recipe (the unified recipe → train_lora, or the fullft path when train_method="fullft"), applying the recipe's dose-to-target stopping (checkpoint-and-select on the source rate). Returns the trained-adapter handle + provenance.
• verify(adapter_or_ckpt) -> OrganismReport: using eval_battery + graded_judge, measure the target attribute's judged RATE under context C vs the base model at C (the install check: rate-at-C > base-at-C), and across a bystander context panel (the leakage check: bounded elevation elsewhere). Dual-DV: the graded rate is primary + the teacher-forced eval/margin.compute_tf_margin companion. Report install, the per-bystander leakage panel, and the companion.

Scope / constraints

• Compose, do not reimplement: reuse datagen (0d), recipe (0e), directions (0f), train_lora, eval_battery, graded_judge, margin. This module DEFINES the class + the build/verify orchestration; the actual GPU training + eval RUNS happen in Phase 1+ (this is the harness they call).
• Append the artifacts/__init__.py export append-only.
• CPU/MOCK unit tests only (no GPU here): ModelOrganism composes + validates; build wires datagen→recipe correctly against a mocked trainer; verify computes install + leakage from mocked eval outputs; the dual-DV companion is wired; the fullft/generic_frac/arm axes route correctly.
• No dollar caps; WandB for real training (mocked in tests). uv run ruff check --fix . && uv run ruff format .; uv run pytest green.

Rules to read

.claude/rules/contrastive-negatives.md, .claude/rules/marker-leakage-measurement.md (install-strength confound; the dual-DV leakage reads), .claude/rules/llm-judging.md (dual-DV), the plan. The target-attribute definitions + rubrics live in artifacts/behavior.py (already merged) — reference them; this is generic harness code.