Treatment definitions (frozen for the pilot; re-frozen before the confirmatory stage)
All arms share: the same TASK.md and bundle; the same client (Claude Code 2.1.314, --bare), model (claude-opus-5-5, CLI default effort), tool inventory (Bash, Read, Write, Edit, Glob, Grep; no web, no sub-agents), permission policy (--permission-prompts none, listed tools pre-approved), offline executor node with the same pinned Python environment, the same resource envelope (4.00 USD, 1,500 s wall, 80 turns per session), the same user prompt (protocol/arms/task_prompt.md) and the same common preamble (protocol/arms/common_preamble.md: environment facts, output contract, integrity rules). The only difference is the text appended after the common preamble in --append-system-prompt-file:
| Arm | Appended text | Words (preamble + arm) | Controller behaviour after the session |
|---|---|---|---|
| A — Open instruction | arm_A.md (one sentence) |
285 | collect, post-hoc contract check (not fed back), evaluate |
| B — Structured protocol | arm_B.md (phases P0–P5, stopping rules, recovery rules, phase ledger, claims ledger) |
944 | same as A |
| C — Protocol + executable gates | arm_C.md = B + mandatory gate program (~/wp29/gates/pgar_gate.py) |
1,106 | run the pristine gate; if it fails, resume the session with the gate output (≤ 3 cycles, within the envelope); evaluate |
| D — Protocol + gates + separate review | arm_D.md = C + announcement of a separate review |
1,200 | as C; then a separate reviewer session (fresh session id, reviewer.md system prompt, sees only TASK.md + out/) writes out/review/review.json; if revise, the executor is resumed once with the review; gate again; evaluate |
Measured initial context (first assistant turn, smoke run, tokens): A 2,238; B 3,717; C 4,111; D 4,319 (sessions.first_turn_input_tokens). These include the client's own system prompt and the task prompt; the difference between arms is the protocol text only.
Content-equivalence audit: the arm texts add no scientific information (no data facts, no method hints beyond "compute the simplest valid baseline", which is already a required output in every TASK.md that defines a baseline key). The common preamble is byte-identical across arms (protocol_hash per run = SHA-256 of preamble+arm text; effective_context_hash adds TASK.md and the user prompt). A length-matched unstructured control (arm_A_long.md, same word count as B, generic research advice without phases, ledgers or gates) is a Phase 08 ablation if the budget allows.
What the gate cannot do: it contains no reference values and cannot judge scientific correctness; it checks the output contract, internal consistency and reproducibility. What the reviewer cannot do: it has no reference answers and no access to the executor's transcript (it is told not to look for it; the transcript lives outside the workspace under ~/.claude/projects on the node — a residual, documented leak path that was not exercised in the smoke run: the reviewer's tool uses referenced only workspace paths).