A one-way door decision is hard or impossible to reverse. A two-way door can be walked back cheaply. This explorer asks Jev, the decision model from TypeSafe, to classify a decision, rate how much work it would take to undo, and say how much human oversight it warrants. Reversibility is the reason this matters for AI. The cheaper a decision is to undo, the safer it is to let an agent make it alone.
This text is the state Jev reasons over. Only the decision text is sent to the model, together with the three questions below. The reference reading is not part of the prompt.
The metaphor. The one-way / two-way door framing comes from Amazon's 2015 letter to shareholders. Two-way-door (reversible) decisions can be made fast and delegated widely, because if you're wrong you just walk back through the door. One-way-door decisions are consequential and near-irreversible, so they deserve slow, careful deliberation.
It's a spectrum, not a binary. "Reversible" hides a range of cost. Flipping a feature flag off is instant and free. Reversing a credit-card charge takes a refund or a dispute but is routine. A Bitcoin transaction settles on-chain with no one able to claw it back, and an 83(b) election with the U.S. Internal Revenue Service (IRS) is generally irrevocable. The "Work to undo it" scale is where this artifact tries to place a decision.
Why it matters for AI. If an AI agent picks the wrong two-way door, the cost is a quick reversal. If it picks the wrong one-way door, the cost is permanent. So reversibility is a natural gate for autonomy: let agents act freely on cheap-to-undo decisions, ask for human confirmation on costly ones, and keep irreversible ones with a person. That's exactly what the "oversight" question probes.
Reversibility can change out from under a model. A decision's door type is not fixed for all time. An outside catalyst, often a change in law or policy, can turn a two-way door into a one-way door. Converting a traditional individual retirement account (IRA) to a Roth IRA used to be reversible. You could unwind it with a recharacterization. The 2017 Tax Cuts and Jobs Act removed that option starting in 2018, so every conversion is now final. A model trained mostly on older text can be confidently wrong here, reading the decision the way it worked before the change. The catalyst tag on an example marks a case where the answer moved after some cutoff. Treat it as a prompt to check whether the model's world is still current.
Is Jev any good at it? That's the open question. Every example ships with a Claude analysis: Claude's own call on how reversible the decision is, and why. Hold Jev's numbers against it. Both sides are models guessing, and either one can be wrong.