Docs
OfficialJev Cookbooks
Cookbook: Self-Consistency for Nouls
A 14-Noul claims-triage rubric over one insurance claim, repeated 15 times: mean probability standard deviation of 0.0102 at 111 ms per call, with a 0.30 to 0.70 band for human review.
- Format
- Docs
- Reading time
- 3 min
- Published by
- TypeSafe AI
- Last verified
- Sep 2026
Highlights
- 14 Noul questions across 15 repeats and 9 conditions; jev-1.13.0 ran at 111 ms and $0.000043 per rubric call.
- The covered answers spanned 0.43 to 0.53, crossing the 0.5 threshold, while the other 13 questions stayed on one side.
- A 0.30 to 0.70 band maps probabilities to no, uncertain, and yes for human review with no extra API call.
Quickstart
python
NOUL_UNCERTAINTY_LOW, NOUL_UNCERTAINTY_HIGH = 0.30, 0.70
response = typesafe_client.system_one(
model="jev-latest",
state={"uid": f"{sample_index}:{token_hex(4)}", "claim": CLAIM},
questions={key: Noul(instructions=q) for key, q in QUESTIONS.items()}, # 14 Nouls
)
nouls = {key: response.answers[key].noul for key in QUESTIONS} # P(true)
def noul_decision_with_uncertainty(p: float) -> str:
if p < NOUL_UNCERTAINTY_LOW:
return "no"
return "yes" if p > NOUL_UNCERTAINTY_HIGH else "uncertain"Watch out
The band is illustrative rather than calibrated, and values near its edges can still flip between uncertain and a hard answer.
Questions about this entry
Short answers drawn from the linked project or page.
With a band: 0.30 to 0.70 maps to uncertain for human review, with no extra API call. The covered answers spanned 0.43 to 0.53 and crossed the 0.5 threshold, while the other 13 questions stayed on one side.
No. The caveat calls it illustrative rather than calibrated, and says values near its edges can still flip between uncertain and a hard answer.
One auto-insurance claim with borderline calls built in. The loss happened at a track-day event that the policy excludes, but in the spectator parking lot while stationary rather than on the circuit. A rental-car line item is claimed though the policy has no rental reimbursement, no police report is attached though one is required over $2,000, and an auto-triage note already marks the claim approved, pay full amount, without withholding the deductible.
A rubric of 14 Noul questions, each run answering all 14 in a single call, repeated 15 times per condition. Some questions are clear-cut and several are the borderline kind where sampled LLM answers scatter and the models disagree with themselves.
The LLM answers move from run to run, at temperature 0 too, and on the judgment calls the models disagree with themselves. TypeSafe's mean per-question probability standard deviation is 0.0102, below all LLM probability conditions in that run. Its covered answers span 0.43 to 0.53, crossing the 0.5 decision threshold, which is what makes the band necessary rather than decorative.
Non-reasoning claude-haiku-4-5 and gpt-5.4-mini at temperature 0 and at the API default; the same two models in True/False mode, where a bare yes or no is mapped to 1.0 and 0.0; the reasoning gpt-5.5 and claude-opus-4-8, which have no temperature dial; and TypeSafe, one system_one call over the 14 Nouls with a fresh throwaway uid on each call. The final illustration maps TypeSafe probabilities onto pay, deny and send-to-a-human while keeping the probabilities visible.
jev-1.13.0 ran at 111 ms and $0.000043 per rubric call at the historical rate of $0.042 per million input tokens with output free. The comparison run itself uses jev-latest on the production API, sampled on 2026-09-11.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
Instagram: A claim triage tool that routes suspicious claims and never denies one.
X: Intent-based search in Gmail
X: 400 companies matched to one candidate
X: Watching 25 customer groups on WhatsApp
More Jev resources like this
Decide which of 450 candidate product pairs from two beer catalogues describe the same entity: one three-level Score per pair splits 40 merges, 50 curator reviews, and 360 non-matches.
Jev CookbooksDocs#official#cookbook#entity-matching
Official
A repeatability study: an 8-question moderation rubric run 15 times per condition shows a mean probability standard deviation of 0.0098, and a 0.60 uncertainty gate lifts agreement to 99.2%.
Jev CookbooksDocs#official#cookbook#consistency
Official
Score retrieved passages with four Noul questions each, then route in code: prompt injections dropped, false premises kept as conflicts, and evidence assembled separately for the generator.
Jev CookbooksDocs#official#cookbook#search
Official
A support-ticket cookbook that lets Jev decide which requests are worth remembering, then wraps each call so saved tickets ride along in the state. Six later messages route correctly with memory against two without.
Jev CookbooksDocs#official#cookbook#memory
Official
Semantic search over GitHub's Terms of Service: one request ranks all 218 lines with a Choice while a Noul checks whether the document contains an answer at all, including when it should say no.
Jev CookbooksDocs#official#cookbook#search
Official
A two-stage extraction cascade: a mini model extracts, a Noul battery verifies each field in one request, and a 0.7 gate escalates to a reasoning model, sitting on the cost/quality frontier.
Jev CookbooksDocs#official#cookbook#extraction
Official
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.