Showcases

Decisions people keep rewriting by hand.

Each one is a small graph: unstructured text in, a typed branch out, thresholds in your code. Open any of them in the playground.

01 · Support

Triage without a summary

A help desk does not need a paragraph about the customer. It needs a queue, a flag if the note is on fire, and a sense of how angry the writer is. Three questions, then an if.

Run triage

02 · Agents

A gate in front of side effects

The agent already chose refund_charge. Jev does not send the refund. It returns allow, confirm, review, or deny, plus a risk score, and your policy map picks the threshold. Money moves only in your code.

Run the guard

03 · Agents

Don’t spend a reasoner on a lookup

Some tasks are a database read. Some need a slow model. Some should be split before anyone writes a plan. proceed_fast / deep_review / split_task / block is that switch.

Run the router

04 · Research

The quote has to carry the sentence

Pass the claim and the passage you already retrieved. The verdict is supported, mixed, unsupported, or insufficient. There is no web search hiding inside the call.

Check a citation

05 · Ops

Code the memo, not the PDF layout

Vendor, amount, and a messy memo. The choice is an account, a split, or “a human should see this.” Line math stays in the ledger.

Code an invoice

06 · Context

Compaction that does not paraphrase

Long agent traces die by summary. Ask, per snippet, whether to keep it verbatim, truncate it, or drop it. Anything kept is still the original text.

Sort three snippets

How the code finishes the job

Each showcase stops at a typed answer on purpose. The rest is a normal program.

Triage

If team.confidence is under your floor, the ticket goes to a person with the probabilities attached, not to a guessed queue. Above the floor, billing looks at urgent to decide whether the note skips the backlog. Frustration picks a priority flag. The reply, if you send one, is written by something that can write, using the team and the original note as input.

Guard

Deny and review both stop the tool. Confirm asks the user with the arguments shown, not a paraphrase. Allow still checks the risk score: a read of a record can auto-run, a refund over your policy amount cannot, even when the label says allow. Log the distribution. Later arguments about “why did it refund” are easier with the probabilities than with a boolean.

Router

proceed_fast stays on the cheap model or on code. deep_review is the only path that should start a long chain of thought. split_task should become several narrower Jev calls, not a bigger prompt. block returns to the user with what was missing. Low confidence on the route itself is treated as deep review, so a muddy task does not sneak down the fast path.

Citation

Publish the sentence only when the verdict is supported and the probability on that option is high. Mixed and unsupported go back to the writer with the passage quoted. Insufficient means retrieve again. Do not average this with a second model’s opinion unless you have a reason: two uncalibrated yeses are not a calibration.

Invoice

A split means your books create two lines. The model told you the memo contains two kinds of cost. It did not tell you the dollar split. Parse the amounts in code, or send the memo to review when the figures are not explicit. Review is the right bucket for a memo that is only a person’s name.

Context

Drop removes the snippet from the next prompt. Truncate keeps a slice you cut in code, ideally the first and last lines, not a summary. Keep copies the original string. Run this before you pay a large model to read a trace, not after. The decision is cheap relative to the tokens you would have resent.