Jev Cookbook

100 notebook recipes for typed decisions with Jev, covering text and images, starting from a first Choice question.

Quick start · How a recipe works · Decision types · Levels · Categories · Recipes · Sources · Contributing

---

Jev is TypeSafe AI's System One model. It does not write text. You send it a state and a set of typed questions, and it returns structured answers your code can branch on, sort by, and route with. This cookbook is a curriculum of Jupyter notebooks that teach that way of building, one bounded decision at a time.

Note

The generated badge and recipes table show what has shipped. The catalog lists the planned use cases, categories, and difficulty levels, and it can be revised as recipes are built. A recipe counts as published once its notebook.ipynb is in the repository, and its row then links to the notebook. Every other row is marked coming soon. It links to the issue tracking the recipe when the catalog records one; a row for an entry with no issue yet is marked coming soon with no link. Each published recipe states in its notebook and README which mode it ran in (synthetic fixtures or recorded responses), and for recorded responses the model version the API returned and the date. Metrics from synthetic runs check that the pipeline behaves as designed. They are not Jev results. Any number reported as a Jev result states the number of examples it was measured on and the model version, comes from recorded inference on a held-out set, and makes no general claim about Jev's quality, latency, or cost. Where a recipe has no such run, its README says "not measured live".

Quick start

You need Python 3.10 or newer and git. No API key is needed to run anything offline.

git clone https://github.com/Jev-Engineering/cookbook.git
cd cookbook
python -m venv .venv
. .venv/bin/activate   # Git Bash: source .venv/Scripts/activate; PowerShell: .venv\Scripts\Activate.ps1
pip install -e ".[dev]"
python tools/execute_notebook.py recipes/_template   # run the template notebook offline

The template in recipes/_template/ is a small, complete recipe, and every recipe works the same way; recipe 01 is the start of the curriculum once its row below links to a notebook. Three pages take it from here:

• Getting started: install, open a notebook, run it offline, run a recipe's tests, switch to live.
• Offline and live: what each mode is, what provenance says, why synthetic metrics are pipeline checks, and what "not measured live" means.
• Glossary: state, question, Choice, Noul, Score, criteria, probabilities, confidence, threshold, replay key, fixture, provenance, and run mode, each linked to its source.

How a recipe works

Each notebook keeps the same division of labour. Jev supplies a narrow semantic judgment. Python does everything exact: arithmetic, retrieval, policy, permissions, and execution.

%%{init: {'theme': 'base', 'themeVariables': {'lineColor': '#E551BA'}}}%%
flowchart LR
    F["Prepared fixtures<br/>text or JSON"]:::panel --> Q["Typed questions<br/>Choice · Noul · Score"]:::pink
    Q --> J["Jev<br/>one focused judgment each"]:::ink
    J --> A["Typed answers<br/>with probabilities"]:::pink
    A --> P["Your Python<br/>policy, arithmetic, execution"]:::panel
    P --> E["Evaluation<br/>labels, baselines, held-out sets"]:::ink
    classDef ink fill:#1E1E1E,stroke:#F386A1,color:#FEFEFE
    classDef pink fill:#F386A1,stroke:#1E1E1E,color:#1E1E1E
    classDef panel fill:#DEDEDE,stroke:#1E1E1E,color:#1E1E1E

Loading

A question is a few lines of Python. This is the shape recipe 01 starts from, using the TypeSafe Python SDK:

from typesafe_sdk import Choice, TypeSafeClient

with TypeSafeClient() as client:  # reads TYPESAFE_API_KEY
    response = client.system_one(
        state={"review": "Arrived fast, but the lid cracked within a week."},
        questions={
            "sentiment": Choice(
                instructions="What is the overall sentiment of this review?",
                criteria={"positive": None, "neutral": None, "negative": None, "mixed": None},
            ),
        },
    )

print(response.choices["sentiment"].choice)

What every recipe assumes:

You	Know basic Python. No machine learning background is needed for levels 1 to 3.
Data	Prepared text or JSON fixtures that ship with the recipe. The inputs are written for the recipe, never real data. Stored answers are synthetic unless the recipe says they were recorded from real calls, with the model and the date.
Actions	Simulated. No recipe sends a message, moves a file, or touches a real system.
Model calls	Offline by default, replaying fixtures. Live Jev calls are opt-in with your own key.
Claims	Runs on synthetic or scripted answers show that the pipeline works. Statements about quality, latency, or cost need recorded inference on a held-out set, with the model version, the capture date and the number of examples (N), and a recipe says which one you are looking at.

Decision types

The catalog uses Jev's own vocabulary. Every recipe is tagged with the primitives it asks for.

Type	Role	How to read it
Choice	Select one option	Select one item from supplied alternatives, including none, other, or uncertain outcomes when the task needs them. Returns the choice, a probability per option, and a confidence.
Noul	Judge a proposition	Return the probability that a stated proposition holds. Tasks that allow several labels ask one independent question per label. There is no separate confidence field.
Score	Evaluate a rubric	Return a position and a distribution over ordered rubric levels you define. Exact measurements and arithmetic stay in code.
Box	Locate one target	Return one bounding rectangle for one target, with four edges as fractions of the image's width and height and a confidence. It is never a mask and never a set of detections, and it always returns a rectangle, even when the target is absent, so check presence separately, for example with a paired Noul. Requires the Jev 1.14 preview and the experimental images header.
Combination	Compose judgments	Combine independent questions in Python, and make a later request when one judgment depends on an earlier answer.

A single label can stand for several independent questions of that type. Thresholds are a property of the task, so each recipe that uses one chooses it from examples and evaluates it.

Difficulty levels

In the original track (ranks 1 to 60) the levels rise with rank: rank 1 is the easiest recipe and rank 60 the most demanding, so follow those ranks for a full learning sequence. The multimodal track (ranks 61 to 100) is ordered by category block, not by difficulty, so its levels rise and fall as the rank increases; use each recipe's level, or pick a category, to choose a path through it.

Level	Name	What a notebook at this level involves	Recipes
1	Beginner	One bounded judgment over a short prepared input with direct inspection of typed results.	11 01 to 10, 61
2	Easy	Context-sensitive matching, scoring, or verification with candidate sets and explicit uncertain or no-match outcomes.	24 11 to 23, 66, 69 to 76, 85 to 86
3	Intermediate	Multiple judgments or stages, structured evidence, deterministic composition, and task-level evaluation.	34 24 to 39, 62 to 65, 67 to 68, 77 to 79, 81 to 82, 87 to 93
4	Advanced	Stateful workflows, provenance, adversarial cases, model calibration, or several integrated decision stages.	22 40 to 52, 80, 83 to 84, 94 to 97, 99 to 100
5	Expert	Controlled optimization studies or closed-loop simulations requiring budgets, baselines, and held-out or trajectory evaluation.	9 53 to 60, 98

Ranks are editorial estimates of implementation and evaluation effort. Neighbouring recipes are ordered for learning, not by a measured difference in difficulty.

Browse by category

Category	Recipes	Levels	Recipe numbers
Data quality & knowledge graphs	9	2 to 5	14 22 33 34 37 39 40 47 57
Agent orchestration	8	2 to 4	11 27 28 29 43 44 49 51
Bounding boxes & spatial reasoning	8	3 to 4	77 78 79 80 81 82 83 84
Image understanding & visual judgments	8	1 to 3	61 62 63 64 65 66 67 68
Language & content	8	1 to 2, 4	01 03 05 06 07 12 13 48
Multiframe & temporal understanding	8	2 to 3	85 86 87 88 89 90 91 92
Multimodal & long context	8	3 to 5	93 94 95 96 97 98 99 100
Visual criteria & reference matching	8	2	69 70 71 72 73 74 75 76
Evaluation & optimization	7	2 to 5	23 35 38 50 53 54 55
Search & retrieval	6	1 to 4	08 10 17 20 30 52
Trust & security	6	2 to 5	15 16 26 41 42 56
Workflow & service operations	6	1 to 3	02 04 09 18 24 25
Games & simulation	5	2 to 3, 5	21 36 58 59 60
Software engineering	3	2 to 3	19 31 32
Memory & context	2	4	45 46

The recipes

Original track (01–60)

Text and structured-state decisions using Choice, Noul, and Score, from single judgments to evaluated workflows and simulations.

Level 1 · Beginner

One bounded judgment over a short prepared input with direct inspection of typed results.

#	Recipe	Category	Decision	Status
01	Sentiment classification Classify a short customer review as positive, neutral, negative, or mixed using a fixed set of labels.	Language & content	Choice	Open notebook
02	Refund intent detection Judge whether a customer message explicitly requests a refund so Python can flag it for the appropriate workflow.	Workflow & service operations	Noul	Open notebook
03	Response clarity scoring Score a support response against a clearly defined clarity rubric to identify examples that need editing.	Language & content	Score	Open notebook
04	Support ticket routing Assign a support ticket to a fixed service category or an unclear-request outcome using its subject and description.	Workflow & service operations	Choice	Open notebook
05	Document classification Classify a short document as an invoice, meeting note, policy, technical guide, or other document type from its text.	Language & content	Choice	Open notebook
06	Multiple topic labels Tag customer feedback with every applicable topic by asking an independent yes-or-no question for each predefined label.	Language & content	Noul	Open notebook
07	Word sense selection Select the intended meaning of an ambiguous word from a fixed sense inventory using its surrounding sentence.	Language & content	Choice	Open notebook
08	FAQ selection Select the best matching FAQ from a short candidate list, including a no-match outcome when none addresses the question.	Search & retrieval	Choice	Open notebook
09	File organization Recommend a destination folder from a fixed catalog using a file's name and text excerpt so Python can preview the proposed organization.	Workflow & service operations	Choice	Open notebook
10	Answer relevance check Judge whether a candidate response addresses the user's question so a notebook can separate relevant answers from off-topic replies.	Search & retrieval	Noul	Open notebook

Level 2 · Easy

Context-sensitive matching, scoring, or verification with candidate sets and explicit uncertain or no-match outcomes.

#	Recipe	Category	Decision	Status
11	Clarification selection Select a useful follow-up question from a predefined catalog when a task description omits information needed for the next step.	Agent orchestration	Choice	Open notebook
12	Thesaurus word selection Choose a context-appropriate synonym from a supplied thesaurus list while retaining the original word when no alternative preserves its meaning.	Language & content	Choice	Open notebook
13	Candidate rewrite selection Choose a supplied sentence rewrite that preserves the original meaning and requested tone, with a no-suitable-rewrite outcome when necessary.	Language & content	Choice	Open notebook
14	Source span selection Select the supplier name from text spans already extracted by Python, with a not-stated outcome when the document lacks that field.	Data quality & knowledge graphs	Choice	Open notebook
15	Sensitive text triage Flag synthetic documents that may contain personal information so a simulated review queue can prioritize redaction checks.	Trust & security	Noul	Open notebook
16	Discord moderation triage Classify sample Discord messages as allowed, review-needed, or potentially violating a supplied community rule to populate a simulated moderator queue.	Trust & security	Choice	Open notebook
17	Passage reranking Score retrieved passages against a question so Python can rank candidate evidence and retain each passage's source reference.	Search & retrieval	Score	Open notebook
18	Duplicate incident matching Select a matching incident from a retrieved ticket shortlist, or no match, using descriptions of the symptoms and affected service.	Workflow & service operations	Choice	Open notebook
19	CI failure classification Classify recorded continuous-integration failures as test regressions, dependency problems, infrastructure failures, or unknown to select the next diagnostic workflow.	Software engineering	Choice	Open notebook
20	Claim support classification Classify whether a provided passage supports, contradicts, or leaves unresolved a specific claim while preserving the passage identifier.	Search & retrieval	Choice	Open notebook
21	Quiz answer adjudication Judge whether a typed quiz response matches an accepted answer by meaning, with partial-match and review outcomes for ambiguous responses.	Games & simulation	Choice	Open notebook
22	CMDB asset matching Link differently worded software asset descriptions to a canonical configuration-management record from a bounded candidate list or a no-match outcome.	Data quality & knowledge graphs	Choice	Open notebook
23	Pairwise answer evaluation Compare two candidate answers against one explicit rubric criterion while retaining tie and insufficient-evidence outcomes for evaluation against human labels.	Evaluation & optimization	Choice	Open notebook

Level 3 · Intermediate

Multiple judgments or stages, structured evidence, deterministic composition, and task-level evaluation.

#	Recipe	Category	Decision	Status
24	Incident priority composition Score reported business impact and urgency against separate rubrics so Python can map the scores to defined categories and apply a fixed service-priority matrix.	Workflow & service operations	Score	Open notebook
25	Change evidence review Check whether a proposed change includes sufficient implementation, validation, and rollback evidence so a deterministic checklist can identify review gaps.	Workflow & service operations	Noul	Open notebook
26	Prompt-injection screening Flag instruction-redirection attempts in retrieved text using adversarial and benign quotation fixtures to measure false positives and missed attacks.	Trust & security	Noul	Open notebook
27	Function selection Select a function from a preapproved catalog for a request while Python validates supplied arguments before simulated execution.	Agent orchestration	Choice	Open notebook
28	Skill selection Score retrieved skill descriptions for relevance to a task so Python can select a suitable skill or request clarification.	Agent orchestration	Score	Open notebook
29	Tool outcome verification Classify recorded tool outcomes as complete, partial, failed, or unverifiable using the original request and observed evidence.	Agent orchestration	Choice	Open notebook
30	Citation verification Check each answer claim against its cited passage and flag unsupported or contradicted claims while Python verifies citation identifiers.	Search & retrieval	Choice	Open notebook
31	Patch risk triage Assess a small patch for separate authentication, input-validation, and data-loss concerns to prioritize human review using explicit rules.	Software engineering	Noul	Open notebook
32	Regression test selection Score candidate tests against the behavior changed by a patch so Python can propose a focused regression suite.	Software engineering	Score	Open notebook
33	Schema field mapping Map source field descriptions to a bounded target schema with ambiguous and unmapped outcomes before Python validates the proposed mappings.	Data quality & knowledge graphs	Choice	Open notebook
34	Evidence relationship labels Label candidate entity pairs with an allowed relationship type or no-supported-relation using provided evidence passages.	Data quality & knowledge graphs	Choice	Open notebook
35	Semantic feature extraction Convert text into predefined semantic features with independent judgments and compare a classifier trained on those features with a fixed baseline.	Evaluation & optimization	Noul	Open notebook
36	Card-game action selection Select a legal action in a turn-based card-game simulator using structured state and fixed action descriptions while Python enforces game rules.	Games & simulation	Choice	Open notebook
37	Research domain tagging Assign independently judged research-domain labels to evidence passages while Python preserves provenance and enforces access controls.	Data quality & knowledge graphs	Noul	Open notebook
38	Abstention and review queues Route uncertain ticket classifications to review using thresholds selected from validation examples and measure accuracy versus coverage on a held-out set.	Evaluation & optimization	Choice	Open notebook
39	Document extraction cascade Select candidate values for document fields, verify each selection against its source span, and route uncertain or inconsistent records to review.	Data quality & knowledge graphs	Choice + Noul	Open notebook

Level 4 · Advanced

Stateful workflows, provenance, adversarial cases, model calibration, or several integrated decision stages.

#	Recipe	Category	Decision	Status
40	Hierarchical classification Route documents through a taxonomy in stages while Python carries competing branches forward when early judgments are uncertain.	Data quality & knowledge graphs	Choice	Open notebook
41	Approval scope review Flag semantic mismatches between an approved task and a proposed tool action while Python enforces exact permissions and approval bindings.	Trust & security	Noul	Open notebook
42	Outbound disclosure screening Screen proposed outbound messages for disclosure concerns against supplied sharing rules while deterministic controls hold flagged or uncertain samples for review.	Trust & security	Noul	Open notebook
43	Failure recovery selection Select retry, alternative method, or escalation from a bounded recovery catalog using execution history while Python enforces retry budgets and stop conditions.	Agent orchestration	Choice	Open notebook
44	Plan constraint assessment Evaluate each candidate plan against individual user constraints so deterministic policy can reject ineligible plans and route unresolved judgments to review.	Agent orchestration	Noul	Open notebook
45	Memory reconciliation Classify a new memory statement as corroborating, superseding, contradicting, or unrelated to existing evidence so a versioned store can preserve provenance and unresolved conflicts.	Memory & context	Choice	Open notebook
46	Semantic context paging Judge stored context chunks for relevance to the current task so Python can retain, page out, or rehydrate them while preserving original content and source links.	Memory & context	Score	Open notebook
47	Evidence conflict diagnosis Classify apparent claim conflicts by factual disagreement, time, entity, or scope using source excerpts and temporal relationships already computed by Python.	Data quality & knowledge graphs	Choice	Open notebook
48	Document revision consistency Select among externally supplied sentence revisions and check their consistency with paragraph context as Python assembles a versioned document.	Language & content	Choice + Noul	Open notebook
49	Multi-agent task assignment Score each agent's suitability for a subtask so a deterministic scheduler can assign work under capability, dependency, and budget constraints.	Agent orchestration	Score	Open notebook
50	Model comparison and calibration Compare Jev and alternative decision backends on identical labeled tasks, calibrating each model separately and measuring held-out quality, latency, and cost.	Evaluation & optimization	Choice + Noul + Score	Open notebook
51	Stalled-loop detection Classify progress across recorded agent turns as advancing, repeating, regressing, or complete so Python can apply a continue, change-strategy, or stop policy.	Agent orchestration	Choice	Coming soon · #51
52	Retrieval pipeline evaluation Combine passage filtering, reranking, and claim verification in a retrieval pipeline and measure each stage's effect on answer support using a fixed labeled corpus.	Search & retrieval	Choice + Score	Coming soon · #52

Level 5 · Expert

Controlled optimization studies or closed-loop simulations requiring budgets, baselines, and held-out or trajectory evaluation.

#	Recipe	Category	Decision	Status
53	Semantic feature discovery Judge text against externally proposed semantic features, select a feature set through a budgeted validation search, and evaluate the frozen downstream classifier on an untouched test set.	Evaluation & optimization	Noul + Score	Coming soon · #53
54	Question and policy optimization Optimize Jev question wording, criteria, and decision thresholds with an external search loop, then compare frozen variants on an untouched test set.	Evaluation & optimization	Choice + Noul + Score	Coming soon · #54
55	Adaptive model routing Use Jev judgments to drive a model-routing policy and compare quality, latency, and cost across traffic mixes while Python enforces budgets and fallback rules.	Evaluation & optimization	Choice + Score	Coming soon · #55
56	Adversarial harness evaluation Evaluate Jev screening at input, retrieved-context, proposed-action, and output boundaries using adversarial replay while measuring attack success, false positives, and task completion.	Trust & security	Noul	Coming soon · #56
57	Evidence graph updates Combine entity, relationship, and evidence judgments into a proposed graph update that Python validates and applies transactionally in a simulated store with replayable provenance.	Data quality & knowledge graphs	Choice + Noul	Coming soon · #57
58	Simulated process supervision Select supervisory responses to simulated PLC fault descriptions and telemetry while Python enforces interlocks and a separate controller handles timing and actuation.	Games & simulation	Choice + Noul	Coming soon · #58
59	Observation and action selection Choose the next permitted observation and subsequent legal action in a partially observed game, then compare complete trajectories with a rule-based baseline.	Games & simulation	Choice	Coming soon · #59
60	Factory production control Select bounded production goals and recovery actions from structured factory state while a Factorio-style simulator executes predefined action templates and measures throughput, stalls, and resource use.	Games & simulation	Choice + Noul + Score	Coming soon · #60

Multimodal track (61–100)

Experimental image judgments, visual reference matching, Box localization, ordered still-frame analysis, and multimodal long-context evaluation.

Multimodal level 1 · Beginner

One bounded judgment over a short prepared input with direct inspection of typed results.

#	Recipe	Category	Decision	Status
61	Screenshot error-state recognition Classify an application screenshot as a crash, loading state, authentication failure, normal operation, or unknown using fixed visual categories.	Image understanding & visual judgments	Choice	Coming soon

Multimodal level 2 · Easy

Context-sensitive matching, scoring, or verification with candidate sets and explicit uncertain or no-match outcomes.

#	Recipe	Category	Decision	Status
66	Technical schematic recognition Identify a predefined circuit, wiring, or infrastructure symbol from a cropped schematic image, retaining an unknown outcome for ambiguous or unlisted symbols.	Image understanding & visual judgments	Choice	Coming soon
69	Visual SKU recognition Select a product variant from a fixed SKU catalog whose choice criteria contain reference photographs, with no-match and uncertain outcomes.	Visual criteria & reference matching	Choice	Coming soon
70	Brand identity verification Compare a displayed logo or marketing asset with an approved reference image using bounded visual conformity checks and route uncertain results to review.	Visual criteria & reference matching	Noul	Coming soon
71	Golden-sample assembly checks Score an assembly photograph against rubric levels illustrated by good and defective reference examples so Python can identify candidates for inspection.	Visual criteria & reference matching	Score	Coming soon
72	Board-game piece recognition Identify a token, tile, figurine, or card from fixed categories defined by reference images, including an unknown outcome for unseen or ambiguous pieces.	Visual criteria & reference matching	Choice	Coming soon
73	Visual UI state identification Classify a cropped UI component as selected, disabled, active, hovered, error, or unknown using visual examples in each choice criterion.	Visual criteria & reference matching	Choice	Coming soon
74	Material swatch matching Match a textile, paint, flooring, or material photograph to a controlled collection of visual swatches, with no-match and uncertain outcomes for appearance ambiguity.	Visual criteria & reference matching	Choice	Coming soon
75	Generated-art style consistency Score a generated illustration against separate art-direction criteria illustrated by a supplied moodboard, using explicit rubric levels and retaining each criterion score.	Visual criteria & reference matching	Score	Coming soon
76	Plant appearance matching Compare a plant photograph with labeled visual examples to suggest a candidate appearance group, including unknown for unsupported or ambiguous matches.	Visual criteria & reference matching	Choice	Coming soon
85	Door state transition detection Classify a door as opened, closed, unchanged, or uncertain across chronologically ordered PNG or JPEG frames sampled by Python.	Multiframe & temporal understanding	Choice	Coming soon
86	Object arrival/departure detection Determine whether a specified package, vehicle, or other object appeared, disappeared, remained, or is uncertain across ordered still frames of a scene.	Multiframe & temporal understanding	Choice	Coming soon

Multimodal level 3 · Intermediate

Multiple judgments or stages, structured evidence, deterministic composition, and task-level evaluation.

#	Recipe	Category	Decision	Status
62	Visual regression detection Compare before-and-after screenshots for missing elements, misalignment, clipping, and other predefined visual changes, then score each change against an explicit severity rubric.	Image understanding & visual judgments	Noul + Score	Coming soon
63	Manufacturing defect inspection Check synthetic product photographs for scratches, dents, missing components, and incorrect assembly with independent judgments and a defined defect-severity rubric.	Image understanding & visual judgments	Noul + Score	Coming soon
64	Field maintenance photo triage Classify predefined equipment conditions in maintenance photographs and independently flag visible corrosion, leaks, or disconnected cabling for a simulated review queue.	Image understanding & visual judgments	Choice + Noul	Coming soon
65	Retail shelf compliance Compare shelf photographs with an approved planogram image using independent checks for misplaced or missing product groups while Python combines the findings.	Image understanding & visual judgments	Noul	Coming soon
67	Product finish verification Compare a product photograph with a visual reference using separate color, material, and finish checks and explicit similarity rubrics under controlled imaging conditions.	Image understanding & visual judgments	Noul + Score	Coming soon
68	Visual accessibility screening Flag apparent text clipping, low contrast, crowding, and overlapping UI components with independent screenshot judgments to identify candidates for further accessibility testing.	Image understanding & visual judgments	Noul	Coming soon
77	Visual browser element locator Locate a named button, form input, or menu item in a screenshot and check target presence before Python draws the rectangle and previews a simulated browser action.	Bounding boxes & spatial reasoning	Box + Noul	Coming soon
78	Signature and stamp localization Check whether a synthetic scanned document contains a requested signature or approval stamp and locate one matching region for Python to crop and evaluate.	Bounding boxes & spatial reasoning	Box + Noul	Coming soon
79	Engineering diagram component locator Check for and locate a specified valve, connector, or symbol on an engineering diagram while Python converts normalized box edges into a visible annotation.	Bounding boxes & spatial reasoning	Box + Noul	Coming soon
81	Defect localization Check for one specified visible scratch, dent, or broken component, then draw and evaluate its bounding rectangle while suppressing boxes for absent targets.	Bounding boxes & spatial reasoning	Box + Noul	Coming soon
82	Reference-guided object search Locate a particular product on a cluttered shelf or workbench using its reference photo in the instructions, with an independent presence check before accepting the rectangle.	Bounding boxes & spatial reasoning	Box + Noul	Coming soon
87	Game phase transition detection Recognize transitions between combat, menus, map selection, and rewards from ordered game screenshots while Python records the phase history and handles uncertain transitions.	Multiframe & temporal understanding	Choice	Coming soon
88	Visual action verification Compare screenshots before and after a simulated browser action to judge whether the intended UI change is visible while Python preserves the action and frame records.	Multiframe & temporal understanding	Noul	Coming soon
89	Conveyor jam detection Evaluate ordered synthetic factory frames for visible blockage or halted conveyor movement and score the evidence using a fixed rubric while Python controls sampling and timing.	Multiframe & temporal understanding	Noul + Score	Coming soon
90	Assembly progression recognition Identify the predefined assembly stage reached in a chronological photograph sequence, retaining unknown or inconsistent outcomes for review.	Multiframe & temporal understanding	Choice	Coming soon
91	Sports action phase recognition Classify a short ordered still-image sequence as preparation, execution, recovery, or an unrecognized movement using supplied phase definitions.	Multiframe & temporal understanding	Choice	Coming soon
92	Video highlight candidate detection Score frame sequences for separately defined highlight signals such as visible reactions or unusual scene changes while Python samples video and maps candidate windows to timestamps.	Multiframe & temporal understanding	Score	Coming soon
93	Image-grounded bug report verification Check whether a screenshot supports each claimed bug symptom using the description, error logs, and expected behavior, preserving evidence links for a simulated review queue.	Multimodal & long context	Noul	Coming soon

Multimodal level 4 · Advanced

Stateful workflows, provenance, adversarial cases, model calibration, or several integrated decision stages.

#	Recipe	Category	Decision	Status
80	Game HUD perception Identify and locate a requested energy meter, card, action button, or enemy indicator in a synthetic Slay the Spire 2-style screenshot, with absent and ambiguous outcomes before any simulated action.	Bounding boxes & spatial reasoning	Box + Choice	Coming soon
83	Hierarchical visual search Select among Python-generated image tiles, retaining no-match and uncertain outcomes, then progressively localize a small target while Python maps the final rectangle back to the full image.	Bounding boxes & spatial reasoning	Choice + Box	Coming soon
84	Bounding-box self-verification Check target presence, localize it, and ask independent questions about whether it continues past each crop edge so Python can adjust the rectangle within fixed retry and size limits.	Bounding boxes & spatial reasoning	Box + Noul	Coming soon
94	Long-form visual evidence audit Compare inspection photographs with individual statements in an extended report or maintenance history using opt-in long context while Python preserves evidence provenance and review outcomes.	Multimodal & long context	Noul	Coming soon
95	AI livestream overlay verification Compare gameplay frames and recorded agent state with displayed strategy, phase, and action overlays to flag unsupported status claims in a simulated livestream review.	Multimodal & long context	Noul	Coming soon
96	Multilingual visual instruction verification Classify whether photographs and signage support instructions across languages, comparing pinned Jev 1.13 and 1.14 versions on identical labeled tasks with an uncertain outcome.	Multimodal & long context	Choice	Coming soon
97	Photo-based equipment troubleshooting Combine a photographed symptom, equipment model information, and a long troubleshooting manual to select candidate diagnostic checks from a supplied catalog for simulated review.	Multimodal & long context	Choice	Coming soon
99	Historical screenshot sequence analysis Check candidate state transitions against a long sequence of screenshots and textual test logs so Python can identify the earliest supported visible transition associated with a defect.	Multimodal & long context	Noul	Coming soon
100	Multimodal agent handoff verification Classify a reported agent result and independently check its claims against screenshots and tool records before deterministic policy permits a simulated handoff or requests review.	Multimodal & long context	Choice + Noul	Coming soon

Multimodal level 5 · Expert

Controlled optimization studies or closed-loop simulations requiring budgets, baselines, and held-out or trajectory evaluation.

#	Recipe	Category	Decision	Status
98	Long-context versus focused-context evaluation Compare judgments using contexts up to the opt-in 262k-token limit with retrieval and staged decisions across all four decision types on identical labeled tasks, with frozen policies, held-out evaluation, and measured resource use.	Multimodal & long context	Choice + Noul + Score + Box	Coming soon

Repository layout

The layout the recipes are being built into. catalog/recipes.json is the source of truth for this README, and a recipe flips from coming soon to a notebook link when its notebook.ipynb lands.

cookbook/
├── catalog/recipes.json      the 100 planned recipes, machine readable
├── recipes/NN-slug/          one folder per recipe: notebook.ipynb, README.md, fixtures/
├── src/jev_cookbook/         shared helpers: backends, fixtures, evaluation, plotting
├── tools/render_catalog.py   regenerates the tables in this README
├── orchestration/PROMPT.md   the multi-agent build plan for this repository
└── CONTRIBUTING.md           the recipe contract every notebook has to meet

Sources

The use cases are proposed applications. Their ranks are curriculum judgments, not ratings published by the sources. The references below were reviewed on 6 October 2026 (S01 to S08) and 15 October 2026 (S09 to S12). Each reference's review date is recorded in catalog/recipes.json, and the references that apply to a recipe are listed in that recipe's catalog entry. S07 describes limitations of the documented Jev 1.13 model specifically.

ID	Reference	What it supports
S01	TypeSafe AI: Introduction	Typed questions over supplied state and atomic judgments composed in code.
S02	TypeSafe AI: Primitives (Questions)	Choice, Noul, Score, independent questions, and dependent follow-up calls.
S03	TypeSafe AI: Confidence	Distribution-derived confidence and task-specific thresholds informed by observed outcomes.
S04	TypeSafe AI: How to build with TypeSafe	Narrow semantic judgments integrated with workflows, deterministic logic, and side effects controlled by code.
S05	TypeSafe AI: Example use cases	Application families covering classification, routing, scoring, retrieval, and verification.
S06	TypeSafe AI: Cookbooks	Worked patterns for extraction, citation checking, skill selection, hierarchical classification, and feature discovery.
S07	TypeSafe AI: Jev 1.13 jaggedness	Version-specific limitations affecting arithmetic, context selection, adversarial content, and generation.
S08	TypeSafe AI: System One Adapter	A typed interface for comparing Jev with decision adapters backed by generative models.
S09	TypeSafe AI: Images	Experimental PNG/JPEG assets in state, instructions, and criteria; image comparisons and ordered still frames with the images header.
S10	TypeSafe AI: Box	Jev 1.14 single-target bounding rectangles, normalized edges and confidence, absent-target checks, tiled search, and crop verification.
S11	TypeSafe AI: Models and long context	Versioned model IDs, moving aliases, and opt-in 262k-token per-request and per-question context for Jev 1.13 and 1.14.
S12	TypeSafe AI: Jev 1.14 release	The October 14, 2026 release of Jev 1.14 and the experimental status of image, Box, and long-context capabilities.

Contributing

Recipes are built against one contract, described in CONTRIBUTING.md: offline by default, synthetic fixtures, Python in control of every side effect, and no quality claim without a measured run. Progress is tracked in issues and milestones, with each recipe's assigned issue recorded in its catalog entry (issue is null until one is assigned). A recipe's rank is not its issue number; see CONTRIBUTING.md.

_{An independent project by Jev-Engineering · Steward: Complete Tech LLC · MIT licensed
Not an official TypeSafe AI publication. The pink, ink and pixel-window styling is a nod to typesafe.ai; Jev, System One and TypeSafe are their respective owner's names.}