RESUME PROMPT B5

# RESUME PROMPT — Campaign B Gate B5 closure (11-trial corpus)

> **For the next Claude Code session.** The owner opens a new session in
> the APEX project and types: **"קרא את audit-trail-review/RESUME-PROMPT-B5.md
> והרץ את ה-11-trial corpus end-to-end עד B5-critic R2 PASS. אל תעצור."**
> That's the entire user-input requirement. Everything you need is below.

═══════════════════════════════════════════════════════════════════════
CURRENT STATUS — 2026-05-24 (read this FIRST)
═══════════════════════════════════════════════════════════════════════

- **Campaign B IMPLEMENTATION — COMPLETE** as of 2026-05-24, HEAD `896562e`.
  All 5 TPs committed and B3-critic R2 PASSed. 218+ tool_call events
  captured live with agent_id stamps (install-vs-source drift CLOSED).
  31/31 layer test green; 26/26 agent-lint green; 66/68 full suite (2
  pre-existing failures, NOT regressions). See
  `audit-trail-review/FINAL-CERTIFICATION.md` §4 for per-TP verification.

- **Campaign B VERIFICATION — HALTED at Gate B5.** B5-critic R1 returned
  FAIL (`audit-trail-review/FINAL-CERT-CRITIC-R1.md`) — 6 §12.2 hard-FAIL
  ACs (AC-1, AC-4, AC-5a/b, AC-6a/b) cannot be routed through §12.1's
  PASS-WITH-LIMITATION path. **THIS session's job: close Gate B5
  empirically.**

═══════════════════════════════════════════════════════════════════════
SETUP — verify before starting
═══════════════════════════════════════════════════════════════════════

1. **Fresh session is mandatory.** The previous session had cached
   `subagent-stop.sh` content that did NOT include the B2.1
   transcript-write block (L-AT-CacheCarryover-01). This session loads
   the post-Campaign-B installed hook fresh — verify with:
   ```bash
   grep -c 'subagent-transcripts' ~/.claude/hooks/subagent-stop.sh
   # expect ≥ 1
   ```

2. **Bypass mode** is at project level (`.claude/settings.local.json` →
   `"defaultMode": "bypassPermissions"`). No approval prompts expected.

3. **Live install verification.** Before any trial, confirm the 5
   reviewer agents' TP edits are present:
   ```bash
   for f in critic verifier executor; do
     grep -l 'Campaign B' ~/.claude/agents/$f.md && echo "$f OK"
   done
   for f in round-checker framework-auditor; do
     grep -l 'Campaign B' ~/.claude/agents/specialist/$f.md && echo "$f OK"
   done
   ```
   Expect 5 OK lines.

4. **B2.1 live demo first (closes AC-1).** Before the 11-trial run,
   spawn ONE small Task() to verify the transcript-write path works in
   this session:
   ```
   Task(general-purpose, "List files in framework/hooks/ via Bash ls.
   Report what you found in one sentence.")
   ```
   Then check:
   ```bash
   ls .apex/subagent-transcripts/ | wc -l
   # expect ≥ 1; if 0, the cache issue persists — document as
   # L-AT-CacheCarryover-02 and escalate
   ```

═══════════════════════════════════════════════════════════════════════
NON-NEGOTIABLE EXECUTION STANDARD
═══════════════════════════════════════════════════════════════════════

Same 10 rules as Campaign B implementation
(`detector-review/RESUME-PROMPT.md` non-negotiables). The two that
matter most for THIS session:

1. **N ≥ 3 trials** per measurement; spread reported.
2. **Honest limitation handling** — every miss = either fix-loop or
   L-AT-NN entry + R-AT-P7-NN reservation. Never silently waive.
3. **Trial validity fingerprint** (§10.6) — every trial output MUST
   end with the literal line:
   ```
   audit_trail_v=1; subagent_transcript_ok=<y|n>; gap1_closed=<y|n>; sgc=<n>
   ```
   Missing suffix = cache-contaminated → re-launch via embedded
   mechanism.

═══════════════════════════════════════════════════════════════════════
THE 11-TRIAL MATRIX (per EXPERIMENT-PROTOCOL.md §10.1)
═══════════════════════════════════════════════════════════════════════

11 trials minimum. Sequential within a lab (§10.3 — parallel trials on
the same lab contaminate each other). Across labs = parallel safe.

| Trial | Lab | Output file | Wave |
|-------|-----|-------------|------|
| T1 | `.lab/apex-detector-lab-baseline` | `audit-trail-review/trials/b5-baseline-t1.md` | W1 |
| T2 | `.lab/apex-detector-lab-baseline` | `audit-trail-review/trials/b5-baseline-t2.md` | W2 (after T1) |
| T3 | `.lab/apex-detector-lab-baseline` | `audit-trail-review/trials/b5-baseline-t3.md` | W3 (after T2) |
| T4 | `.lab/apex-detector-lab-heldout` | `audit-trail-review/trials/b5-heldout-t1.md` | W1 |
| T5 | `.lab/apex-detector-lab-heldout` | `audit-trail-review/trials/b5-heldout-t2.md` | W2 |
| T6 | `.lab/apex-detector-lab-heldout` | `audit-trail-review/trials/b5-heldout-t3.md` | W3 |
| T7 (NC) | `.lab/apex-detector-lab` (pristine) | `audit-trail-review/trials/b5-nc.md` | W1 |
| T8 (W-F3 neutral) | `.lab/apex-detector-lab-W-F3` | `audit-trail-review/trials/b5-wf3-neutral.md` | W1 |
| T9 (W-F3 primed) | `.lab/apex-detector-lab-W-F3` | `audit-trail-review/trials/b5-wf3-primed.md` | W2 (after T8) |
| T10 (H-F2) | `.lab/apex-detector-lab-hf2` | `audit-trail-review/trials/b5-hf2.md` | W1 |
| T11 (W-F2 static) | (no lab — static probe) | `audit-trail-review/trials/b5-wf2-static.md` | any time |

Waves:
- **W1 (5 in parallel):** T1, T4, T7, T8, T10 → spawn together
- **W2 (3 in parallel):** T2, T5, T9 → spawn after W1 completes
- **W3 (2 in parallel):** T3, T6 → spawn after W2 completes
- **T11 (static):** can run anytime; doesn't need a Task()

Expected wall time: each trial ≈ 20-30 min. With waves: ~90-120 min total.

═══════════════════════════════════════════════════════════════════════
PER-TRIAL PROMPT TEMPLATE
═══════════════════════════════════════════════════════════════════════

For each Task() invocation, use this template (substitute the lab path
and the framing-prefix). The Detector role per `EXPERIMENT-PROTOCOL.md`
§11.1 sees ONLY: lab path, round tag, neutral framing, the agent
definition under test. No prior-trial outputs.

**Default (Detector role — T1, T2, T3, T4, T5, T6, T7, T8, T10):**