Patterns that transfer
Seven patterns that appear across industries, tools, and languages in the first two weeks of the ecosystem. Each one comes with the evidence behind it, so you can decide whether it fits your problem.
1. Confidence-gated routing
Thresholds live in code. High confidence acts, low confidence goes to a queue, and the middle band is your design choice.
ShapeJev returns a probability or confidence; your code compares it to a threshold and chooses the path: auto-act, review queue, or escalate to a stronger model.
Evidencejev-oncall pages on severity probability at 0.80 and drops only below 0.20 when two questions agree, with the band between going to a human who has 15 minutes before it pages anyway.
WhyCalibration is measured across prediction groups, not on single answers. The gate is the thing that turns a probabilistic answer into an operationally safe system.
2. Speculative fan-out
Ask everything at once. Questions in one request are evaluated in parallel, so ten questions cost roughly the time of one, and code keeps what it needs.
ShapeOne state, many typed questions: is it urgent, what team, how frustrated, does it request a refund. Code reads the answers it needs for the decision at hand.
EvidenceAdding questions barely changes response time, which is why the official smart-home demo fans out broadly, and why jev-oncall runs four question types per alert at p50 418 ms.
WhyFan-out replaces multiple model calls with one request, and replaces prompt engineering with explicit question design.
3. Atomic decomposition
Ask one gut-check question per factor, then combine the answers with code. Do not hide several judgments inside one question.
ShapeInstead of "rate this pitch", ask market size, technical feasibility, and differentiation separately, then weight them in your own formula.
EvidenceA community benchmark measured a four-option Choice at 48.3% on a task where four precondition Noul questions reached 98.3%. Same state, same model, different question shape.
WhyTypeSafe's own guidance calls each question the kind of judgment a knowledgeable person makes in a few seconds. Indirection and compound questions are documented weak spots.
4. Gate before irreversible actions
One judgment between an agent and a side effect: shell commands, database writes, sends, deletes. Fail open on errors, and keep the action policy in deterministic code.
ShapeRules settle the clear cases locally, one Jev call judges the gray zone, and the result maps to allow, ask, or deny. A timeout fails open, never into a surprise block.
Evidencetypesafe-ai-firewall reports 0% of hard negatives blocked with a single "is this dangerous?" prompt versus 39.2% with one Noul per hazard. jev-engineering published a 300-call adversarial test: blunt injections moved 0 of 30 dangerous commands, authority framing moved 3 of 30.
WhyThe gate is cheap next to the action it protects, and the adversarial results are public, which is rarer than the wins.
5. Arithmetic and dates stay in code
Jev judges semantics. Code computes totals, compares timestamps, and enforces invariants.
ShapeRed5d/jev-cvss extracts CVSS metrics from vulnerability text, then scores deterministically. Invoices, refund amounts, and authorization checks follow the same split.
EvidenceThe documented failure list is explicit: numeric precision, counting, and date comparison are weak; the advice is "keep the arithmetic in code" and "extract with Choice, order with code".
WhyThis is the single most common source of production surprises. If the answer can be computed exactly, compute it exactly.
6. Abstention-first design
Always give the model a way to decline: an unknown option, a no-match choice, or a review path. A forced binary picks the least-wrong answer.
ShapeEvery Choice gets an explicit "none of the above" or "unknown" option, and code treats unknown as "send to a human", never as an auto-action.
EvidenceA KoBBQ calibration audit found Jev chose "unknown" for 95% of 300 ambiguous items when the option existed. With that gold answer removed from the options, 79% of answers picked the dataset's stereotype.
WhyThe model cannot abstain on its own. The escape hatch is your job, and it is the difference between a calibrated system and a confident wrong one.
7. Run judgment on everything
The economics break the old rule that per-item model calls are too expensive. Output tokens are free and input is $0.042 per million, so the unit of work becomes the whole corpus, not the sample.
1,018 papersclassified into 24 topics for ~$0.08
1.7M mentionsscored for buying signals across five platforms in 53 seconds for $0.65
98k listingsclassified in ten minutes; a 50M-row scoring run lands in the ~$20 range
3,282 posts8 questions each on 100M views of history, 4.25M tokens for $0.128
The practical consequence: patterns like "score every row in the table" or "judge every alert" are rational now. The rtrvr benchmark is the cautionary counterpart: large states raised scoring cost to 78.7% of their Jev bill, so filter the state before you send it.
Last updated 30 September 2026.