Iterative decision simulation. Spend compute to think better. Five persona analysts as independent subagents, anonymized peer review, mechanical convergence — until the answer is robust.
-
NEVER simulate in a vacuum. Phase 1 (GROUND) is mandatory. If WebSearch yields nothing, mark the run UNGROUNDED in the brief header — do not proceed silently.
-
Phase 1.5 (ELICIT) runs after GROUND, before the loop, unless --skip-elicit. The single biggest quality lever — never default-off.
-
Each persona runs as a SEPARATE Agent subagent. Genuine context-window independence. Sequential authoring in one context destroys diversity. Non-negotiable.
-
The main conversation IS the orchestrator. Walk the phases yourself. Spawn agents for parallelizable tasks (5 personas, critique + adversary). NEVER spawn one agent to "run the loop" — that agent can't spawn grandchildren and the council collapses. See engine-protocol.md "Orchestration Model."
-
The 5 canonical personas are the default. Optimist, Pessimist, Competitor, Regulator, Customer. ELICIT may modify (rename, specify a competitor), add (e.g., "Investor" for fundraising), or remove (e.g., Regulator when irrelevant) personas. Custom personas follow the same structure: optimization objective, blind spot, contrarian question, no-hedging rule. Names defined in references/persona-council.md "Canonical Persona Names."
-
Every effect carries a stable effect_id, a probability, a probability_range, and explicit assumption keys. The Judge compares by ID across iterations — descriptions drift, IDs don't. No implicit assumptions. Assumption keys are as stable as effect_ids: iteration 2+ personas receive the full all_assumptions map from iter-1's effects-chains.json and MUST reuse keys verbatim for conceptually-identical assumptions. Renaming market_has_demand to market_demand_exists between iterations fakes instability and breaks the Judge's assumption_stability metric. See phases/simulate.md "Assumption Key Stability."
-
Persona disagreement IS the uncertainty signal. The probability range is the data — never average it away.
-
Generate 2nd-order effects for ALL 1st-order effects. No probability gate. Tail risks matter most.
-
Anonymize during peer review. Personas review "Analysis A", "Analysis B" — never by name. Mapping randomized per iteration.
-
The Decision Brief is for humans. Never emit snake_case, never backtick raw effect_ids in prose. Use the description field. See references/phases/decide.md.
-
Phase 8.5 (VALIDATE-BRIEF) is mandatory for full/medium/revise/quick. It means literally invoking scripts/validate-brief.py against the schema. Writing a custom inline Python validation script (checking for your own invented section headers, declaring "13/13 passed" against a list you authored) IS NOT Phase 8.5. It is self-certification. Self-certification against invented headers is a HARD protocol violation — it silently lets a structurally broken brief ship. If the named script cannot run (e.g., python3 missing), fall back to the Step 5.5 self-check in phases/decide.md and emit the structural-self-check footer. Do NOT roll your own validator. On HARD_FAIL, re-prompt DECIDE once; if still failing, prepend VALIDATION_FAILED and continue. See references/phases/validate-brief.md.
Writer must not invent section headers. The brief's H2 structure is defined by references/brief-schema.json and is MANDATORY. Inventing headers like ## Context, ## Decision tilt, ## The possibility map, ## Methodology, ## Analysis Approach (any header not in the schema) is a HARD_FAIL, even if it reads well. Improving readability is the schema's job, not the writer's. Do Step 4a's pre-write checklist before composing a single line.
-
Stop when the Judge says so. Max iterations configurable (default 2, up to 5). The iteration folders ARE the memory — read previous iteration's convergence-summary.md (≤500 tokens) before starting the next, never the full JSON.
-
Subagent nesting. If /autodecision runs inside another agent, the Agent tool may be unavailable. STOP and ask the user (full protocol in engine-protocol.md). Never silently degrade.
-
Read references/engine-protocol.md BEFORE starting any phase. This is not "on demand" — it is the first file you read after SKILL.md. If you have not read engine-protocol.md in this conversation, stop and read it now. The protocol defines the loop structure, file writes, and orchestration model. Without it, you will write a one-shot memo instead of running the actual loop.
-
Intermediate files are mandatory — no brief without a loop. Before Phase 8 (DECIDE) can start, the run directory MUST contain: config.json (Phase 0), ground-data.md (Phase 1), at least one iteration-{N}/effects-chains.json (Phase 3), and convergence-log.json (Phase 7). If these files do not exist, the loop did not run. Do NOT write a brief from memory or from the context file alone. Go back and run the missing phases. A Decision Brief without upstream data files is fabricated, not analyzed.
-
The Decision Brief has exactly 16 H2 sections in full mode. They are, in order:
## Executive Summary, ## Data Foundation, ## Hypotheses Explored, ## Effects Map, ## Council Dynamics, ## Minority-View Winners (optional), ## Stable Insights, ## Fragile Insights, ## Adversarial Scenarios, ## Key Assumptions, ## Convergence Log, ## Recommendation, ## Appendix A: Decision Timeline, ## Appendix B: Complete Effects Map, ## Appendix C: Quick Mode vs Full Loop Comparison (optional), ## Sources.
Writing a brief with different section names — "Evidence Summary", "Options Considered", "Adversary Findings", "Sensitivity Analysis", "Persona Council Results", "Critique Findings" — is a HARD_FAIL. These are NOT schema headers. Use the exact headers above.
Read on-demand as each phase begins. Each phase file opens with a self-describing metadata block (phase number, when it runs, what it reads, what it writes, what gates it has) — always read that block first.
