decision-space-expansion

decision-space-expansion is a decision methodology distilled from the decision-workbench project. Before an important decision, it uses a heterogeneous multi-model committee to widen the option space, structures the divergence, folds real constraints back in — and lets a human make the final call.

Expand the option space before you decide.

The methodology ships in the Agent Skills format (a single Markdown file with YAML frontmatter). It works with any tool that can consume instruction files — no code, no dependencies.

Links

• Repository: sscodeai/decision-space-expansion
• Files: README.md (English) / README.ja.md (Japanese) / SKILL.md (full skill definition)
• License: MIT

Why

• Decision power = size of the option space, not the ability to pick. Your cognitive surface is your decision ceiling.
• A single model (or one model in many roles) only re-samples its own training distribution. True divergence requires different model families (Anthropic / OpenAI / DeepSeek / Google / local OSS), isolated from each other to prevent anchoring.
• Model disagreement is often post-hoc rationalization plus ecosystem bias — but ecosystem blind spots are exactly the highest-value divergence.

The loop (3 stages)

Mandatory question patterns

• “Why didn’t you choose X?” — forces out excluded options and reasons (highest information density)
• “Give one non-mainstream option with a real justification.” — kicks the model off its training distribution
• “Where is this plan most vulnerable?” — cross-examination a single model never performs on itself

Core beliefs

“AI shouldn’t make important decisions for you. It should make the decision space harder for you to misunderstand.”

The committee expands your known unknowns; it cannot reach unknown unknowns. And being forced to understand the divergence while converging is the cognitive growth. The best posture is to ask AI when you don’t know what you don’t know, not when you already know what you want.

Pitfalls

• Don’t treat the committee as an “auto-optimal-answer” machine — open tools (MoA / LLM-Blender) fuse answers by default and delete the human arbitration step, which is the opposite direction.
• Pseudo-divergence wastes time: models will fabricate complete rationales for any assigned stance.
• When requirements don’t constrain the choice, models output their own familiar default — not an optimum.
• Models silently assume hidden constraints (concurrency, budget, team); humans must backfill the real ones.

Install into AI tools

SKILL.md is tool-agnostic. Swap “different model families” for whatever pool your environment can reach (an aggregator, multi-CLI, local Ollama/vLLM) and the method stays the same.

Reference implementation

The reference implementation lives in the decision-workbench repository (FastAPI + React, mock mode with zero API keys). Its README reports the hidden-gold eval as single-model 26% → naive multi-agent 78% → decision-workbench 100% (12 of 12).

Next

• See the implementation in decision-workbench
• Read SKILL.md in the repository
• Browse other SSCodeAI projects in Projects