Implement ## Why this experiment adversarial interrogation gate at task creation, PM dispatch, and /issue Step 0

Goal

Force the user to articulate four concrete answers (Decision / Branches / Cut / Application) before any experiment can advance from proposed → planning. The user is prone to running experiments without a concrete reason; this gate is friction designed to interrupt that impulse.

Critical design constraint: the user MUST personally articulate the four answers in chat. The agent's role is to interrogate and transcribe, NEVER to draft, synthesize from context, or auto-extract from prior materials. If the user delegates ("you decide", "use the ideation doc"), the agent refuses and re-asks.

This task itself is bootstrap — its ## Why this experiment section is intentionally absent because the gate doesn't exist yet. User explicitly waived the dogfood test.

Application: infra — serves Audit + Predict (forces stating-the-decision before compute commits).

The four questions (asked one at a time)

1. Decision this changes: What concrete choice in your queue or proposal hinges on this outcome?
2. Expected outcome + branches: What do you expect, and what alternative outcome would route you to a different next experiment?
3. What gets cut: Which experiment in your queue doesn't run this week as a result?
4. Application: detect | predict | defend | audit | infra — one sentence on which deliverable this advances.

Four-layer gate dynamic (per question)

1. Ask the question.
2. Refuse non-answers (pattern check; see reject list below).
3. Challenge the substance (research-taste critique) — ONE round per question, max.
4. Defer + transcribe when the user defends, reframes, or overrides.

User always has the last word via override ("I'm right, move on" / "defer" / "ship it"). Override is unconditional and logged.

Refuse list (non-answer patterns)

Reject answers containing only these phrases, or answers under ~15 words that don't name a specific decision/branch/cut:

• "just curious" / "want to see what happens" / "to explore" / "to check" / "to understand X better"
• "use the ideation doc" / "you decide" / "look at the parent" / "summarize from context"
• "because it would be interesting" / "to add to the literature"

Implementation: regex pattern list + LLM substance check ("does this answer name a concrete X?").

Challenge lenses (one per question, only when applicable)

If the agent doesn't see a substantive issue, it skips the challenge entirely. No gratuitous challenges.

Termination rules

Three accepted user replies to a challenge:

• Defense: original answer + brief defense → transcribe both, move on.
• Reframe: user changes their answer → transcribe new answer, no further challenge.
• Override: "I'm right" / "defer" / "ship it" → transcribe original + user_overrode_challenge: true marker, move on.

Three enforcement points

1. PM session — dispatch gate (primary)

When user says "work on #N" or "spawn #N" in the PM session:

• Read body.md; if ## Why this experiment is missing or stubby (any of 4 lines empty / under threshold), block dispatch.
• Run the four-question interrogation in chat.
• After all four are filled (with any overrides logged), call task.py set-body <N> --file <draft.md> and post the epm:gate-filled events.jsonl marker.
• THEN spawn the per-issue session via spawn_session.py spawn-issue --issue N.

Update .claude/agents/research-pm.md Mode 5 (DISPATCH) to include this pre-check.

2. /issue <N> Step 0 — safety net

If PM is bypassed (user opens a per-issue session directly), Step 0's existing clarifier check extends:

• If ## Why this experiment missing or stubby → run the four-question interrogation in-session.
• Legitimate AskUserQuestion use because it IS a gate.

Update .claude/skills/issue/SKILL.md Step 0 (body completeness check).

3. task.py new --kind {experiment|survey|infra} — creation gate

If --body-file doesn't contain ## Why this experiment, refuse with:

ERROR: experiment/survey/infra tasks require `## Why this experiment` (4 lines).
This section must be filled by an interrogating agent, not written from scratch.
Use the PM session, `/ideation`, or `/experiment-proposer` to draft it.

Note: ideation/proposer outputs already produce four-line candidates; promoted candidates pass automatically. The CLI guard catches "I'll just type it manually" bypass attempts.

verify_task_body.py changes (new check #12)

• Frontmatter has application: <enum> field (one of: detect, predict, defend, audit, infra).
• Body has ## Why this experiment H2 section with four labeled lines: **Application:**, **Decision this changes:**, **Expected outcome + branches:**, **What gets cut if we run this:**.
• Each line non-empty, ≥40 chars after the label.
• Application enum value matches frontmatter application:.
• FAIL if missing, stubby, or mismatched.
• Skipped when frontmatter contains legacy_why_unset: true.

Logging

Every gate fill writes one events.jsonl row:

{
  "kind": "epm:gate-filled",
  "version": 1,
  "gate": "why-experiment",
  "filled_by": "<agent_session_id>",
  "challenges_fired": ["decision", "branches"],
  "user_overrides": []
}

After ~30 fills the user samples whether challenges are firing on substantive issues or going soft. If soft, sharpen the lens prompts.

Migration

One-shot migration script scripts/migrate_add_legacy_why_sentinel.py:

• 130 completed + 60 archived + 33 awaiting_promotion → add legacy_why_unset: true to frontmatter.
• 55 proposed → leave WITHOUT the sentinel. Next time the user wants to run any of them, the PM gate fires.
• 2 running + 1 blocked → add the sentinel (already past the gate point); user can fill retroactively at their option.
• Migration commit is atomic; rollback by removing the field.

Files touched

• scripts/task.py — new subcommand rejects experiment/survey/infra without ## Why this experiment.
• scripts/verify_task_body.py — new check #12 (application enum + section + line non-empty + frontmatter match).
• scripts/migrate_add_legacy_why_sentinel.py — new one-shot migration.
• .claude/agents/research-pm.md — Mode 5 (DISPATCH) extended with pre-spawn gate.
• .claude/skills/issue/SKILL.md — Step 0 extended with gate check.
• .claude/skills/why-experiment-gate/SKILL.md — NEW skill, the interrogation protocol with tight system prompt (interrogator only, no drafting, refuses non-answers, one challenge max per question, user override terminates).
• .claude/workflow.yaml — add why-experiment to the canonical gate list under § gates.
• Body template under tasks/ docs — update with the four-line section.

Test plan

1. Unit: verify_task_body.py correctly FAILs on missing section, stubby lines, missing frontmatter field, mismatched application enum; PASSes on well-formed body; SKIPs on legacy sentinel.
2. CLI: task.py new --kind experiment --body-file empty.md rejects with clear error; task.py new --kind experiment --body-file with_why.md succeeds.
3. Migration: dry-run prints diff; apply commits one atomic change; sentinel correctly added to 226 tasks, absent on 55 proposed.
4. Manual end-to-end: Run the gate on a chosen proposed task (suggested test case: #114 "Use activation oracles to see persona" — small enough that the gate cost is felt). Confirm:
   • Interrogator asks Q1, refuses "just curious", accepts substantive answer.
   • Challenge fires when the named decision is weak.
   • User override ("ship it") terminates the loop, marker logged.
   • Section lands in body.md, frontmatter application matches, verifier PASSes.
5. Per-issue safety net: spawn per-issue session for a freshly-gated task; confirm /issue <N> Step 0 sees a complete section and proceeds without re-asking.

Acceptance criteria

• Gate can be entered from PM session OR /issue Step 0 OR task.py new (refusal path).
• Agent never drafts content from context; user words transcribed verbatim.
• One challenge per question, never more.
• User override is unconditional and logged.
• Verifier check #12 enforces the floor; conversation enforces the substance.
• Migration adds sentinel to non-proposed tasks; 55 proposed tasks remain ungated until first-use.

Explicit cuts (not bundled into this task)

• Auto-extract from /ideation outputs (anti-pattern: defeats the friction).
• Multi-application secondary field (anti-pattern: escape hatch — if you need two, split the task).
• Budget-cap removal (separate workflow change, queued next).
• Auto-continuation hardening (separate workflow change, queued after budget removal).
• Per-experiment dashboard (separate workflow change, scope still TBD).