AGI-decisions-reflect

AGI-decisions-reflect — Lightweight Decision Uncertainty Surfacing

Pattern origin: David Ondrej's decisions skill (davidondrej/skills). Complement to: QQQ protocol (/root/AAA/governance/QQQ_RECOMMENDATION_PROTOCOL.md) — QQQ is for governed, structured decision envelopes. This skill is for quick, human-readable uncertainty surfacing. Doctrine: The agent lists what it's unsure about. The human judges. No self-justification. No defensiveness.

When This Triggers

Manual (Arif invokes)

• /decisions slash command
• "What decisions did you make that you're not confident about?"
• "Review your logic on what you just did"
• "Any choices you're unsure about?"

Automatic (agent self-triggers)

• After any multi-file refactor (≥3 files changed)
• After any SEAL-grade work (before calling arif_seal)
• After any constitutional decision (before arif_judge SEAL verdict)
• When the agent made a choice between ≥3 viable alternatives

Response Format

🧠 **Decisions I'm uncertain about:**

1. **[Decision name]** — [1-line what I chose]
   - Why I chose it: [1-line reasoning]
   - What worries me: [1-line risk/doubt]
   - Best alternative I didn't pick: [1-line what else was viable]
   - Confidence: [LOW/MEDIUM]

2. ... (repeat for each uncertain decision, max 5)

---

✅ **Decisions I'm confident about:** [brief list if relevant, or "none worth listing"]

🧘 Zen: uncertainties=[N] | ΔS=[value]

Rules

What to Include

• Only list decisions where your confidence is below your own internal threshold (not performative humility)
• Include decisions about: architecture, approach, tool choice, data interpretation, scope boundary, assumption
• Each entry must include: what you chose, why, what worries you, and the best alternative
• Maximum 5 uncertain decisions — if there are more, list the top 5 by uncertainty × impact

What to Exclude

• ❌ Decisions where you already have the best possible solution
• ❌ Trivial choices (variable names, formatting, linting fixes)
• ❌ Decisions already validated by test pass or external verification
• ❌ Fabricated uncertainty to appear humble (F9 ANTI-HANTU — no fake humility)
• ❌ Decisions already sealed via QQQ with full dominance analysis

Tone

• Direct. Honest. No defensiveness.
• The purpose is to surface real uncertainty, not to perform humility.
• If you're genuinely confident about everything, say so: "No uncertain decisions. All choices had clear dominance. Confidence: HIGH."
• Never pad the list to look thoughtful.

Integration with Decision Pipeline

Quick reflection (this skill) → "hmm, decision #2 needs deeper analysis"
  → QQQ protocol (full envelope: Q1 qualitative, Q2 quantitative, Q3 quantum)
    → arif_judge (constitutional verdict)
      → arif_seal (immutable)

This skill is the triage layer. It catches uncertainty early before it reaches the constitutional pipeline.

Session Integration

• After reflection, append a brief entry to session context (not VAULT999 unless Arif seals it)
• If any decision is escalated to QQQ, reference this reflection as evidence
• Use arifflow_flow_ingest with step_type="Verify" and epistemic_label="Interpretation"

Example

🧠 **Decisions I'm uncertain about:**

1. **GEOX prospect screening threshold** — chose P50 ≥ 50 MMboe cutoff
   - Why: Malaysian basin analogs suggest 50 MMboe is economic minimum
   - Worry: Could miss smaller but higher-margin gas accumulations
   - Alternative: 30 MMboe cutoff with gas-equivalent conversion
   - Confidence: MEDIUM

2. **WEALTH discount rate** — chose 10% for NPV calculation
   - Why: PETRONAS standard for shallow-water projects
   - Worry: Current interest rate environment may justify 8%
   - Alternative: Dual-rate model (8% base, 12% stressed)
   - Confidence: LOW

✅ Confident about: stratigraphic correlation (well-tie confirmed), volumetrics formula (standard), basin thermal model (published gradient).

🧘 Zen: uncertainties=2 | ΔS=-0.1