test(parity): eighth-pass §L.5 closeout — Track B B=1000, NSW+PSID-1, agent bench

- coverage_monte_carlo: add explicit-rate run_b1000.py harness; ship
  coverage_b1000.json with measured 95% CI coverage at B=1000 for
  OLS (0.952), 2x2 DiD (0.955), strong-Z IV (0.962); all inside the
  99% Wilson band [0.935, 0.967] around nominal 0.95. Full slow
  pytest sweep at B=1000 also passes 8/8 (753.51 s).
- orig_parity: add module 04b on causalsens::lalonde.psid (true
  Dehejia-Wahba NSW+PSID-1, 2675 obs); sp.regress + sp.psm match R
  bit-equal and recover the published -15,205 to rel 1.5e-05.
- agent_bench: add CausalAgentBench harness (50 prompts L1/L2/L3 x
  6 cells x 3 reps = 900 trials) with deterministic mock LLM,
  grader emitting H1-H5 directional table, and frozen OSF pre-
  registration protocol. Mock dry-run completes in <1 s; production
  run is one --mock -> --api both flag away once OSF + API budget
  clear.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: brycewang-stanford

tests/agent_bench/README.md

@@ -0,0 +1,87 @@
+# `tests/agent_bench/` — CausalAgentBench harness
+
+This directory holds the infrastructure for the JSS plan §5.5 /
+manuscript §7 LLM-agent behavioural benchmark. The harness is
+**fully built** but the production 900-trial run (6 cells × 50
+prompts × 3 reps with Anthropic and OpenAI APIs) is gated on:
+
+1. an OSF pre-registration deposit (the protocol is in
+   `prompts/_protocol.md` and ready to upload),
+2. an API budget commitment (~$300–600), and
+3. a green light from the project lead.
+
+Until those land, the harness runs end-to-end against a
+deterministic mock LLM (`runners/mock_llm.py`) so the scoring,
+result-aggregation, and statistical-test pipeline can be smoke-
+tested. The mock-LLM run produces fake transcripts and a real
+score table; flipping the harness to the real API is a one-line
+change in `runners/runner.py`.
+
+## Layout
+
+```
+tests/agent_bench/
+├── README.md
+├── prompts/
+│   ├── _protocol.md        # OSF pre-registration text (frozen for upload)
+│   └── prompts.json        # 50 prompts (L1 × 20, L2 × 20, L3 × 10)
+├── golds/
+│   └── golds.json          # gold answers + 5-dimensional rubrics
+├── runners/
+│   ├── mock_llm.py         # deterministic stub for harness smoke
+│   ├── runner.py           # main loop: run all 900 trials
+│   └── grader.py           # apply rubric + emit per-trial scores
+├── sandbox/                # per-trial scratch directories
+└── results/
+    ├── trials.jsonl        # one line per trial: {prompt_id, cell, rep, output}
+    ├── scores.csv          # per-trial 5-dimensional scores
+    └── headline.md         # H1-H5 hypothesis test outcomes
+```
+
+## Pre-registered hypotheses
+
+The OSF document freezes:
+- **H1**: StatsPAI cells (C1, C2) achieve task success ≥ 90% on L1,
+  ≥ 70% on L2, ≥ 50% on L3.
+- **H2**: StatsPAI cells exceed Pythonic-stack cells (C3, C4) on
+  L2–L3 task success by ≥ 15 percentage points.
+- **H3**: StatsPAI hallucination rate < 5%; Pythonic-stack ≥ 15%.
+- **H4**: StatsPAI uses ≤ 60% of the tokens of Pythonic-stack at
+  fixed prompt+model.
+- **H5**: R-via-MCP cells (C5, C6) match StatsPAI on task success
+  but require ≥ 1.5× the tokens.
+
+Statistical test: cluster bootstrap with prompt as the cluster,
+α = 0.05, Bonferroni correction across H1–H5.
+
+## Six experimental cells
+
+| Cell | Toolset | Agent |
+| --- | --- | --- |
+| C1 | StatsPAI + MCP | Anthropic Claude (latest) |
+| C2 | StatsPAI + MCP | OpenAI GPT (latest) |
+| C3 | Pythonic stack (statsmodels, linearmodels, DoubleML, grf-python) | Claude |
+| C4 | Pythonic stack | GPT |
+| C5 | R via MCP (radian + Jupyter R) | Claude |
+| C6 | R via MCP | GPT |
+
+900 total trials = 50 prompts × 6 cells × 3 reps.
+
+## How to run
+
+Mock-LLM smoke (no API cost):
+
+```bash
+cd tests/agent_bench
+python3 runners/runner.py --mock --cells C1,C3,C5 --prompts L1
+python3 runners/grader.py
+```
+
+Full production run (after OSF + budget approval):
+
+```bash
+export ANTHROPIC_API_KEY=...
+export OPENAI_API_KEY=...
+python3 runners/runner.py --cells C1,C2,C3,C4,C5,C6 --prompts all
+python3 runners/grader.py
+```