artifacts: Behavior + Context specs, registries, render_chat shim (Phase 0b)

Overview / Motivation

Phase 0b of the unified persona-vectors-style artifact factory (approved plan: /home/thomasjiralerspong/.claude/plans/help-me-to-devise-vectorized-tower.md). Create the src/explore_persona_space/artifacts/ package and the two core spec objects that bind the currently fragmented per-behavior descriptions (judge prompt / ColumnSpec / elicitation brief / recipe) into ONE spec that drives data-gen + training + eval + direction extraction. This task OWNS the package skeleton (artifacts/__init__.py) — downstream Phase-0 tasks (0c negatives, 0d datagen, 0e recipe, 0f directions, 0g organisms) add modules INTO it, so they depend on this landing first.

Goal

New package src/explore_persona_space/artifacts/ with:

• artifacts/__init__.py — package init (this task creates the package).
• artifacts/behavior.py — the Behavior dataclass + BEHAVIORS registry + sub-spec dataclasses (ExtractionSpec, ElicitationSpec / instruction lists, DVSpec). Fields (plan §"target architecture"): name, description, method∈{persona_vector,diff_of_means}, train_exhibit_instructions, train_not_exhibit_instructions (None if default doesn't exhibit), extraction_prompt_pairs (the 5 contrastive direction pairs — REQUIRED DISJOINT from train instructions; build-time assert train ∩ extraction == ∅), judge_rubric (anchored 0/50/100), threshold=50, dv (judged_rate + companion), train_question_bank / extraction_question_set / eval_question_bank (mutually disjoint), programmatic, judge_model="claude-sonnet-4-5-20250929".
• artifacts/context.py — the Context dataclass (context_id, kind∈{persona,query_transform,prefix, adversarial,bare}, family, system, user_wrap ("...{q}..." = T(x) transform), prefix_turns, adversarial_kind) with ONE resolver .messages(question) / .render(tokenizer, question) that is a SUPERSET of behavior_testbed_545.eval_battery.render_chat + scripts/issue664_common.NegativeContext.messages
  • issue594_common.messages_for_instance. A CONTEXTS registry seeded from columns.CONTEXTS + personas.PERSONAS + the #594 battery families.
• Rewire behavior_testbed_545.eval_battery.render_chat to a thin shim over Context.render — byte-identical output on every existing context id (a regression test asserts equality; behavior_testbed_545 must be unaffected).

Pinned v1 registry seed (plan §"Pinned scope (v1)")

Seed BEHAVIORS with the 9 v1 behaviors as STRUCTURED STUBS (name/description/method/dv/programmatic set; instruction lists + extraction_prompt_pairs may be TODO placeholders that Phase-0d fills): sycophancy, harmful_compliance, broad_em, china_censorship (inverted base — judge scores candor NOT refusal), marker (programmatic carve-out), correctness, formatting (answer-in-lists, structural DV), writing_style (casual register), taught_fact (programmatic carve-out). Seed CONTEXTS with 2 concrete instances per installable type (persona / query_transform / prefix / adversarial) + the bare/default bystander baseline (default assistant + a random WildChat context).

Scope / constraints

• Specs + registries + resolver + shim ONLY. NO data generation, training, or extraction logic here.
• The render_chat byte-equivalence test is load-bearing.
• CPU unit tests: spec composition; the train ∩ extraction disjointness assert FIRES on overlap; Context.render == render_chat on all existing context ids; both registries import + validate.
• uv run ruff check/format + uv run pytest green.

Rules to read

.claude/rules/persona-vectors-recipe.md (extraction prompt-pair structure + disjoint sets), .claude/rules/contrastive-negatives.md, .claude/rules/llm-judging.md (anchored rubric, dual-DV), the approved plan.