The plan file is ready. Let me provide the plan summary here since the artifacts directory has restricted permissions.

Replace the "optimistically compliant" quality system with ruthless, multi-layer validation across 5 layers.

Replace *) item_passed=true with item_passed=false + optional AI verification. Unknown items fail by default. Also improve existing test/coverage/debug-statement checks to be more thorough.

• Read quality_threshold from pipeline config (default 80, replacing hardcoded 60)
• Read critical_blocks config (default true)
• Critical findings ALWAYS block pipeline regardless of score
• Post detailed score breakdown table to GitHub issue

New multi_perspective_review() function running 3 distinct auditor personas:

1. Security auditor: injection, auth bypass, secrets, CSRF, SSRF
2. Correctness auditor: off-by-one, null access, race conditions, edge cases
3. Architecture auditor: naming violations, coupling, circular deps, pattern breaks

Each writes to adversarial-{persona}.md, combined summary to JSON.

Each compound quality cycle goes deeper, not wider:

• Cycle 1 (fast): E2E, security scan, quality checks
• Cycle 2 (medium): + adversarial, negative prompting, DoD
• Cycle 3 (deep): + simulation, architecture, multi-perspective review
• Cycle 4+ (targeted): Only re-run audits that previously found issues

Controlled by progressive_deepening config flag. Legacy behavior preserved when false.

• Adaptive budget: Complexity <= 30 → 10 iter, <= 60 → 20, > 60 → 30 (only when not explicitly set)
• Diminishing returns: Warn when velocity < 5 lines over last 3 iterations
• Enhanced guard_completion: Reject if uncommitted changes or TODO/FIXME markers in diff

Change adversarial_enabled, simulation_enabled, architecture_enabled to true in daemon-config.

1. DoD unknown item fails (not auto-passes)
2. Quality score blocks below threshold
3. Quality score passes above threshold
4. Critical findings block regardless of score
5. Progressive deepening cycle 1 = fast (adversarial skipped)
6. Progressive deepening cycle 3 = deep (all audits run)
7. Multi-perspective review produces 3 persona files
8. Convergence plateau detection

• Task 1: Replace DoD auto-pass with zero-auto-pass default in run_dod_audit()
• Task 2: Improve DoD verification for test, coverage, and debug checks
• Task 3: Make quality score threshold configurable (replace hardcoded 60)
• Task 4: Add critical-findings-always-block logic
• Task 5: Implement multi_perspective_review() in sw-adversarial.sh
• Task 6: Wire multi-perspective review into stage_compound_quality()
• Task 7: Implement progressive deepening (fast/medium/deep/targeted)
• Task 8: Add adaptive iteration budget to sw-loop.sh
• Task 9: Add diminishing returns detection to build loop
• Task 10: Enhance guard_completion() with uncommitted/TODO checks
• Task 11: Update all 8 pipeline templates
• Task 12: Enable intelligence defaults in daemon-config
• Task 13: Write tests for all quality gates
• Task 14: Run full test suite and fix regressions

• *) item_passed=true replaced with item_passed=false + AI verification
• Quality score threshold configurable per template (not hardcoded 60)
• Critical findings block pipeline regardless of score
• Multi-perspective review generates 3 auditor reports
• Progressive deepening runs appropriate checks per cycle depth
• Adaptive iteration budget based on complexity scoring
• Diminishing returns detection active
• All 8 templates updated with new quality config
• Intelligence features enabled by default
• All new code tested with PASS/FAIL assertions
• Full test suite passes (npm test exits 0)
• Bash 3.2 compatible