Patterns worth looking for

These are composable starting points. Explore the current use-case map and cookbooks before designing. “Can an LLM do this?” is often yes; the opportunity is whether cheaper, faster repeated decisions change the product.

Pattern	Question design and code composition	Where the benefit could come from
Route work to a handler, tool, skill or model	Choice over concrete handlers; separate complexity Score and evidence-sufficiency check. Code applies policy, fallback and permissions.	Avoid a generative router on every request; reserve expensive models for cases that need them.
Retrieve many useful items	Noul per candidate for inclusion, or Score per candidate for graded utility; batch over shared query/state. Code enforces diversity and context budget.	Inspect a wider pool or keep retrieval off the latency-critical path.
Check work continuously	One Noul per defined defect, or distinct quality Scores. A generator revises only flagged work; rerun checks on the revision.	Feedback while an artifact is being written, rather than a slow end-only review.
Extract by selection	Parser/generator proposes possible values with source IDs; Choice picks the intended ID or missing/ambiguous. Code copies and normalizes.	Remove expensive free-form extraction or reduce its retries without inventing values. Candidate recall remains essential.
Verify extraction or citations	Given source + candidate claim/field, Noul for support or Choice supported/contradicted/not established.	Cheap verification of many fields, with a stronger model only on failures. Source support is not universal truth.
Recover document structure	Classify line/block roles and whether adjacent spans belong together; code reconstructs layout from original content.	Produce useful transformed documents without regenerating every token. Preserve source order/content unless intended otherwise.
Semantic find	Enumerate source spans by ID; one Choice over the IDs ranks them, a Noul checks that an answer exists at all, and per-span Nouls confirm when several spans must be kept. Code returns exact source excerpts.	Natural-language search inside documents, logs or code, with stable IDs.
Entity matching	Score each proposed pair using concrete merge/review/non-match criteria; code applies uniqueness/business constraints.	Judge fuzzy pairs cheaply after deterministic blocking/retrieval. Avoid all-pairs explosion.
Reusable quality features	Separate Scores for independent properties; code changes weights or a trained small model consumes the features.	Re-rank or personalize without redoing unchanged judgments. Hard requirements stay separate.
Interactive state interpretation	Present named current facts and candidate actions; Choice for a bounded next step, with success/block checks.	Faster small decisions in games, home workflows, browser/agent loops. Fresh state and external execution still matter.
Gate on confidence	Read the answer for what to do and its confidence for whether to act. Below a threshold chosen from labeled data, fall back to a broader label, a reasoning model or a person. An explicit uncertain option on a Choice, or a review band around the middle of a Noul, keeps borderline cases visible instead of forcing them.	Automate the clear majority while the uncertain tail gets proper attention, instead of one quality level for everything.
Guard LLM input and output	One request screens a message with Nouls for named hazards (jailbreak attempt, hidden instruction in a retrieved passage, leaked secret) and Scores for severity. Code blocks, strips or escalates.	Screening on every turn and every retrieved passage becomes affordable on the request path. Jev does not treat state as hostile, so test with injected text.
Extract dates and quantities	Ask for the parts named in the text (day, month, relative unit, direction) as Choices; code resolves, validates and compares.	Robust date handling without asking the model to order or subtract dates, which it does poorly.
Sensory evidence hybrid	An audio/vision/frontend detector supplies semantic observations and candidates. Jev adjudicates those alongside text context.	Combine sensory evidence with task intent; native image/audio understanding is not implied.

Official worked references:

• Function calling, speculative fan-out, skill selection.
• Reranking, semantic find, hierarchical classification.
• Candidate extraction, citation checks, extraction cascade.
• Structure recovery, entity alignment, feature discovery, composite scoring.
• Confidence-gated routing, intent routing, classification using confidence, self-consistency for Nouls and Choices.
• LLM guardrails, classifying RAG passages (contradiction and prompt-injection screening), date extraction, parallel questions (batching cost and latency).

For every adaptation, produce real questions and state requirements. A list of these pattern names is not a completed Jevify assessment. Keep generation for novel content and multi-step reasoning when it is needed. Compute exact numbers, counts, dates, times, identifiers and constraints in code. Serialized raw spectra or embeddings do not acquire semantic meaning merely by fitting JSON; test a meaningful frontend representation first.