autodecision
A decision operating system for high-stakes choices — business, strategy, career.
Simulates disagreement, stress-tests assumptions, and converges on what actually holds up.
Quick Start · Why Decisions Fail · How It's Different · Examples · Roadmap
Applies Karpathy's autoresearch + LLM Council patterns to decisions.
Why decisions fail
Most bad decisions don't look bad upfront. They fail later — in second-order effects, edge cases, and under stress.
People routinely:
- Think in first-order effects only ("cut prices, sales go up")
- Ignore second-order effects ("…and then competitors match, margins compress, existing customers renegotiate, NRR drops")
- Underestimate worst-case scenarios ("what if we're wrong by 2x?")
- Miss black swan events ("what if visa risk and market risk are correlated?")
By the time these failure modes appear, the decision is already in motion and hard to reverse.
Autodecision exists because these failure modes are predictable — if you force yourself to look for them.
What autodecision does differently
Most AI tells you what's likely to happen.
Autodecision shows what happens next, what breaks, and what flips the outcome.
| Dimension | What it means |
|---|
| Second-order effects | Not just the immediate consequence — the cascade that follows, with probabilities and timeframes |
| Worst-case scenarios | Treated as decision drivers, not footnotes. Every run includes an adversary red-team phase |
| Black swan stress tests | Models correlated risks, irrational actors, and rare events explicitly |
| Assumption sensitivity | Shows exactly which assumption, if wrong, flips the recommendation — with the threshold |
| Grounded in real data | Every analysis starts with a web search for market comps, benchmarks, precedents — no vacuum reasoning |
| Explicit disagreement | 5 independent personas argue. The disagreement range IS the uncertainty signal |
| Mechanical convergence | A Judge measures stability across iterations. Stops when insights stabilize, not when the model runs out of things to say |
What it does
You give it a high-stakes decision. It:
-
Decomposes the decision into sub-questions
-
Grounds every analysis in real-world data — market comps, benchmarks, case studies, precedents
-
Reviews assumptions, personas, and data with you before simulating
-
Spawns 5 independent personas as parallel subagents — each with its own context window, genuinely unable to see the others:
| Persona | Sees | Blind Spot |
|---|
| Growth Optimist | Upside, creative alternatives | Execution risk |
| Risk Pessimist | Downside, failure modes | Opportunity cost of inaction |
| Competitor Strategist | Market dynamics, game theory | Overestimates rationality |
| Regulator | Legal, compliance, constraints | Overweights unlikely regulation |
| Customer Advocate | User value, adoption, retention | Ignores unit economics |
-
Simulates first-order AND second-order effects with probabilities and explicit assumption tracking
-
Critiques via anonymized peer review — personas rank each other without knowing who wrote what
-
Red-teams with worst-case scenarios, irrational actors, and black swans
-
Analyzes sensitivity — which assumption, if wrong, flips the conclusion?
-
Iterates until a Convergence Judge mechanically measures that insights have stabilized
-
Produces a Decision Brief — a structured strategy memo
Best for
Autodecision is built for a specific class of decisions:
- Strategic business decisions — pricing, expansion, M&A, capital allocation, build-vs-buy, hiring senior roles
- Career decisions with asymmetric outcomes — role moves, relocation, founder vs operator, fundraising terms
- Situations with uncertainty AND second-order effects — where the obvious answer isn't the robust answer
- Irreversible or hard-to-reverse decisions — where you only get one shot and want to stress-test it first
It's less useful for:
- Simple factual questions (use a single LLM call)
- Low-stakes everyday decisions (the overhead isn't worth it)
- Decisions where you already have high conviction (this will just slow you down)
Output: The Decision Brief
Structure is defined by brief-schema.json (v1.1). The writer emits all 16 H2 headers in order; the Phase 8.5 validator HARD_FAILs on deviation.
