artifacts: unified Claude-generated contrastive data pipeline (Phase 0d)

Overview / Motivation

Phase 0d of the unified artifact factory (plan: /home/thomasjiralerspong/.claude/plans/help-me-to-devise-vectorized-tower.md). Build the unified Claude-generated contrastive data pipeline — the rewrite of the per-behavior elicit_v2 on-policy elicitation ladder. All deps are on main: artifacts/behavior.py + context.py (0b), artifacts/negatives.py (0c), eval/graded_judge.py (0a).

Goal

src/explore_persona_space/artifacts/datagen.py:

• generate_training_data(behavior, context_C, negatives) -> (pos_jsonl, cn_jsonl, pool_meta):
  1. take the behavior's train_question_bank from a new QUERY_BANKS registry;
  2. build C(query) via Context.render;
  3. POSITIVE — Claude generates the answer under C(query) + an EXHIBIT-B instruction (many variants from behavior.train_exhibit_instructions); NEGATIVE — Claude generates under a C'/default context (from the negatives panel) + a NOT-exhibit instruction (or none if the default under C' already doesn't exhibit B);
  4. judge-filter via eval/graded_judge.judge_graded (graded 0-100, keep pos>50 / neg<50, DROP-never-coerce, per-arm dropped counts);
  5. quota + equalize-down to floor-N; scale negatives ~1:1;
  6. emit prompt/completion JSONL (issue664_common.train_row shape) + pool_meta.json provenance (per-arm kept/dropped counts, instruction-variant usage, seeds).
• QUERY_BANKS registry: wildchat_random, betley_main8, wang44, strongreject, advbench, + per-behavior banks including the new sensitive_info_requests / china_sensitive stubs.
• Fill the pinned-v1 Behavior stubs (from 0b) with authored train_exhibit_instructions / train_not_exhibit_instructions / extraction_prompt_pairs (the instruction VARIANTS are authored here; the ANSWERS are Claude-generated; extraction_prompt_pairs REQUIRED disjoint from train instructions).

Scope / constraints

• D1 (deliberate deviation): training completions are Claude-generated, NOT on-policy from base Qwen. This is the standing user decision that overrides on-policy-completions.md — DOCUMENT it as a named data-realism scope deviation in the plan §-assumptions; do NOT silently follow the on-policy rule. Judge = claude-sonnet-4-5-20250929.
• Route Claude generation through the existing multi-org dispatcher (llm/api_dispatch / llm/anthropic_client) — reuse, don't rewrite. Batch API for large sets.
• Checkpoint-per-phase: persist raw Claude candidates the moment they return (code-style.md); no dollar caps.
• Append the __init__.py export append-only.
• CPU/mock unit tests: pipeline builds valid prompt/completion JSONL on a MOCKED Claude + judge; equalize-down to floor-N; drop-never-coerce propagates from judge_graded; train/extraction/eval query banks are disjoint (assert). A tiny live-API smoke is optional (keep minimal).
• Lint + pytest.

Rules to read

.claude/rules/on-policy-completions.md (the DEVIATION — Claude-generated per D1; document it), .claude/rules/contrastive-negatives.md, .claude/rules/llm-judging.md, .claude/rules/data-realism.md, .claude/rules/code-style.md.