Free Tool

What is the cheapest thingthat can make this decision?

Not which model is smartest — which level is sufficient. Eight questions about one task, and it lands on deterministic code, a decision model, a fast general model, a frontier reasoning model, or a person. Two gates can push it toward a human and never away from one, because a decision you cannot grade is one you cannot automate.

Example

Loaded with a support-ticket routing task so you can see a finished result. Change any answer, or clear it and start from your own.

01Could someone write the rule down?

A formula, a lookup table, a threshold. If the answer follows from stated rules with no judgment call, no model is needed at any price.

02Is the set of valid answers known before you see the input?

A fixed list of departments, a 1-5 severity, a yes/no. A bounded answer space is what makes a task a decision rather than a generation.

03What has to come out the other end?

A decision layer returns a value software acts on. If a person needs to read prose, something has to write the prose.

04Could two competent reviewers label 200 examples and agree?

This is the gate. If reasonable people disagree about the right answer, there is nothing to measure against — and a decision you cannot grade is one you cannot automate at any level.

05How much has to be connected to decide?

One look at the input, or a chain of inference across several facts. Depth is what separates a decision from a piece of reasoning.

06How often does this run?

Volume does not change what a task is. It changes how much the wrong level costs you, and how fast a small error rate becomes a large number of errors.

07If it gets one wrong, what happens?

Not how likely it is to be wrong — what it costs when it is. Reversibility matters more here than severity.

08Does a person see it before it takes effect?

A review step is not a formality. It is the difference between a wrong decision and a wrong outcome.

The cheapest sufficient level

Level 2Decision model

Classification, scoring, routing, gating, ranking — a typed answer software acts on.

Fix first

Nothing. This task is already at the cheapest level that can do it, it is gradeable, and its failure mode is covered. There is nothing here to change.

Both gates

  • clearGround-truth gate
  • clearIrreversibility gate

The ladder

  • L1Deterministic codeKnown rules, arithmetic, validation. No model, no per-call cost, no drift.
  • L2Decision modelClassification, scoring, routing, gating, ranking — a typed answer software acts on.
  • L3Fast general modelModerate reasoning and text generation where the answer space is open.
  • L4Frontier reasoning modelComplex planning, analysis and synthesis across many connected facts.
  • L5Human judgmentHigh-stakes calls, exceptions, and anything nobody can grade.

Nothing you answer leaves your browser. There is no upload and no request — the eight answers are scored on your own machine, which is also why the same answers always give the same level.

Why the ladder, and not a model comparison

Most AI buying decisions get framed as a contest between models, which is the wrong contest. A task that a formula settles should not call a model at any price; a task that has to write a document cannot be a typed decision however cheap typed decisions become. The level is a property of the task, not of the vendor — so the useful move is to place the task first and shop second. Where a task lands two levels above where it could sit, the gap is usually the answer space: bounding it is the largest single cost and latency drop available.

Common questions

What does the Decision Fit Check do?

It places one task on a five-level ladder — deterministic code, a decision model, a fast general model, a frontier reasoning model, or a person — and, where something is short, names the single change that would move it down a level. It answers the useful question, which is not which model is smartest but what is the cheapest, fastest, sufficiently capable intelligence for this specific decision.

Why would I not just use the best model for everything?

Because the best model is the slowest and most expensive one, and most tasks inside software are not reasoning problems. Routing a ticket, scoring a lead, or flagging a transaction is a decision with a known set of valid answers. Sending that to a frontier model costs more, takes longer, and returns prose your code then has to parse back into the value it wanted.

What is a decision model?

A model built to return a typed value rather than text — a choice from a fixed list, a score on a known scale, or a probability that something is true. Software acts on the value directly with no parsing step. TypeSafe AI's Jev is the current example, which is what made this distinction worth a tool.

Why does it refuse to automate some tasks at all?

Because of the ground-truth gate. If two competent reviewers could not label 200 examples and agree on the answer, there is no measurable right answer — so there is nothing to check a model against, at any level. A decision you cannot grade is one you cannot measure, and one you cannot measure is one you cannot safely automate. That verdict is the honest one even though it is the expensive one.

What is the second gate?

Irreversibility. If a wrong answer cannot be undone and nobody sees it before it takes effect, the task routes to human judgment whatever a model would score. Cheap decisions become dangerous precisely when they become numerous — a small error rate over a large volume of irreversible actions is a lot of unrecoverable outcomes.

Can a gate ever make the answer cheaper?

No, and that is deliberate. Both gates are worsen-only: they can move a task toward human judgment and never away from it. The tool can refuse to automate something; it can never bless automating something it has not been told is safe. That direction is the whole safety property.

Does this send my task description anywhere?

No. There is no free-text box and no upload — you pick one of three answers to each of eight questions, and the scoring runs in your own browser. Nothing is transmitted, stored, or logged. That is also why the result is reproducible: the same eight answers give the same level every time.

It says level 2 but gives me no fix. Is that a bug?

No. When a task is already at the cheapest level that can do it, is gradeable, and has its failure mode covered, there is nothing to change and the tool says so. Naming a fix on a clean result would be telling somebody with a working setup to go and repair it.

Where to go next

If this lands at level 2 and you are about to install a third-party server or skill to get there, the MCP Server & Skill Trust Gate grades the artifact before you install it. If the question is what the decision layer will cost once it is everywhere, the Token Economics Workbook prices it per outcome rather than per token.