Hunch

How we reviewed this: we installed the gem from the repository in Docker on Ruby 3.3, ran its test suite, and made live calls against the real API — a probability, the predicates, a pick, a rating, and a batched decide block.

What it does

Three primitives, named the way a Ruby developer would name them:

Hunch.chance("this order is fraudulent", given: order)        # => 0.74
Hunch.likely?("this order is fraudulent", given: order)       # => true
Hunch.pick(:billing, :fraud, :support, given: order,
           question: "which team should handle this order?")  # => :fraud
Hunch.rate(:low, :medium, :high, given: order,
           question: "how risky is this order?")              # => #<Rating high>

Those are our results from a live call on a made-up order — a US billing address, a Russian shipping address, a twelve-item basket on a zero-day-old account. The predicates are thresholds over chance: possibly? at 0.25, likely? at 0.5, probably? at 0.75, almost_certainly? at 0.93, definitely? at 0.99. Reading if Hunch.probably?("fraudulent", given: order) in a Rails model is the whole pitch, and it lands.

pick takes bare symbols or described ones (pick(hold: "stop the order", ship: "send it", given: order)), and rate returns a level plus a position on the scale, so you can branch on the level or compare the number.

One request, several questions

The part that makes it practical is decide:

r = Hunch.decide(given: mail) do |q|
  q.probably? :urgent, "does this convey urgency?"
  q.pick      :team, billing: "payments and refunds", technical: "bugs", sales: "pricing"
  q.rate      :mood, :calm, :frustrated, :livid
end

On a refund complaint we wrote, live:

urgent:  0.95 (true)
team:    :billing   {technical: 0.0, billing: 1.0, sales: 0.0}
mood:    :frustrated

Three judgments, one round trip, and the probabilities are there when you want to argue with the answer. That is the correct shape for this model, and several projects in this directory get it wrong by asking one question at a time.

Tests without the network

The stub backend is the other reason to like this gem:

Hunch.backend = Hunch::Backends::Stub.new(fraud: 0.95, team: :billing, mood: :calm)

Answers are keyed by the question’s key, missing answers raise unless you pass default:, and every call is recorded in calls for assertions. Raising on a missing answer is the right default: a test that quietly invents a probability is worse than one that fails.

The repository’s own suite is 46 runs and 118 assertions, all passing, and the example/ directory is a small Rails app with the README’s validations, error triage and inbound-mail examples plus their tests.

What to know before you use it

given: is the state. Whatever you pass — a string, a hash, an ActiveRecord attributes hash, a raw email — is what goes to TypeSafe or OpenRouter. In a Rails model that is easy to be careless with: given: order.attributes includes every column. Send the fields the question needs.

A failed call raises. The README’s Rails examples all carry rescue Hunch::APIError and return nil, which is the honest pattern — a validation that cannot reach the API should not block a signup — but it means the rescue is your job at every call site, and a missed one turns an outage into a 500.

It is early. Created 18 September, nine commits, one author, MIT, published as hunch 0.3.0 on RubyGems, CI green. Nothing here is half-built, but the version number is accurate.

Verdict

The nicest-feeling library in this batch. It does not ask you to think about requests, questions or schemas: you write the sentence, you get a Ruby value, and when you want several answers about the same object you get them in one call. The stub backend means the tests you already have keep working.

Pass the narrow slice of your record rather than the whole thing, wrap the call sites the way the README’s own examples do, and treat 0.3.0 as the number it is.

For the Rails-specific version of the same idea see Jevalyn; for TypeScript, Advocaat.

See how it compares with other tools in Best Jev tools, tested hands-on.

Review updated Sep 21, 2026. Numbers quoted from the project are its author's own; we don't publish our own measurements of Jev.