Open · Adversarial · via MCP

Irrational

A field guide to the mistakes your mind makes — and an engine that refracts your decision into the biases bending it.

↓ audit a decision below — free, in your own AI


The engine

Audit a decision you're facing.

Assembles the adversarial prompt — it runs in your AI; we keep no model and store nothing.

…or watch it work on a pre-built scenario

The call

Don't greenlight the big-bang rewrite. Carve off the worst module first (strangler-fig), behind the current interface, with a hard 3-month checkpoint.

confidence ~25% · top bias: sunk cost
The audit

This is a System 1 call wearing a suit — fluent, certain, and missing every number that matters. Watch what is actually steering it.

Sunk cost — the “two months on a POC” are gone either way — but they are doing most of the talking. You are reasoning from what you have spent, not from what the next month returns.
Planning fallacy — “about six months” is a best-case, single-point guess with no reference class — and the POC tested the fun 20%, not the six years of edge-cases that are the other 80%.
Survivorship — “Stripe and others did it and won” quotes the visible survivors and steps over the graveyard of rewrites that quietly overran or were abandoned.
WYSIATI — “the old thing is beyond saving” is a confident verdict built on what is annoying today — it ignores the undocumented business logic the old system silently encodes.
Outside view. Big-bang rewrites of mature, in-production systems routinely run 2–3× the estimate, and a sobering share never fully replace the original. Against that base rate, “six months” is the floor, not the plan — anchor your timeline on your team’s own estimate-vs-actual history instead.
Premortem. Picture it a year out, failed: hidden rules surfaced mid-migration, you ended up running both systems at once, velocity dropped instead of rising, “six months” became eighteen, and leadership pulled the plug half-done — leaving two unfinished systems.
Sit with this
  • If you'd never built the POC, would a full rewrite still be the obvious move today?
  • What's your team's real estimate-vs-actual ratio on big projects — and why are you the exception this time?
  • What logic lives only in the old code that nobody fully understands, and what happens to it on cutover day?
  • Could an incremental strangler-fig capture ~80% of the velocity gain at ~20% of the risk?
The recalibration

Not “never rewrite.” The version that survives scrutiny: an incremental strangler-fig — carve off the worst module first, behind the existing interface, with a hard 3-month checkpoint and a timeline built from your own track record. Treat the POC as a learning artifact, not a down payment.

original full rewrite ~25% · incremental path at recoverable risk: moderate–high

Verdict-first, every bias tied to the words you wrote. Browse all 22 biases →

Why not just ask your AI?

Because your AI is trained to agree — to help you execute the framing you walked in with. Irrational assumes you're biased, refuses to validate, and argues the other side — and every bias it names is cited to its primary source, not improvised.