What could you
build with this?
A catalogue before a codebase: 37 places a bounded decision might belong, from documented demos to deliberately speculative experiments.
Read a row as decision → primitive → reason to try it, not “AI is always better”. Rules, classical classifiers or a generative model may win. Business process is an additional lens, not a stronger evidence grade.
The eight local interactions work with explicitly synthetic decisions; actual Jev needs a key and actual Kev needs a local server. The unfamiliar-label pair and architecture calculator add two more experiments. None of the illustrative numbers is a model measurement.
Established / documented demonstrations
| Use case / decision | Typed question | Why a bounded model might fit | Evidence / experiment |
|---|---|---|---|
| Doom movement Choose the next legal game action from structured state. | Choice | Avoid generating prose for a move; keep the game engine authoritative. | TypeSafe launch demo, not replayed here. |
| Wikiracing Pick an existing link or shortlist among many links. | Choice / Score | An out-of-set URL is impossible in a successful typed selection. | TypeSafe launch demo. Source. Catalogue only: a separate integration or domain evaluation would teach more than a duplicate generic demo. |
| Smart-home commands Pick device, room and action; detect compound requests. | Choice / Noul | Batch independent judgments; reserve generation for conversation. | Official walkthrough and video. Source. Catalogue only: a separate integration or domain evaluation would teach more than a duplicate generic demo. |
| Agent model routing Choose a cheap lookup, reasoning model or human. | Choice | The desired answer is an existing handler, not a written response. | Published LangChain middleware code. |
| Agent-action risk flag Flag a proposed tool action for review. | Noul / Score | A quick semantic signal can precede deterministic authorization. | LangChain example; not a proven security boundary. Source. Catalogue only: a separate integration or domain evaluation would teach more than a duplicate generic demo. |
| Tetris / Pac-Man / driving Choose a bounded move on each turn. | Choice | No need to decode a sentence to move a piece. | Secondary reports; primary runs not verified. |
| Bulk paper and listing labels Select a topic or listing category. | Choice | Repeated small classifications need no narrative output. | Secondary reports; counts and costs not reproduced. Source. Catalogue only: a separate integration or domain evaluation would teach more than a duplicate generic demo. |
| Live writing feedback Rate tone, urgency or clarity as text changes. | Score | An ordinal rubric can drive an indicator without generating a critique. | Secondary reported editor demo. |
Adjacent patterns
| Use case / decision | Typed question | Why a bounded model might fit | Evidence / experiment |
|---|---|---|---|
| Spam triage Is the message likely unsolicited or deceptive? | Noul | Use p(yes) and a review band instead of an explanation per item. | Documented proposal; not a verified deployed spam service. |
| Content moderation Flag policy categories and severity separately. | Noul / Score | Multiple independent signals support an auditable review queue. | Pattern; evaluate false positives and appeals. |
| State-machine transition Select only among transitions legal in the current state. | Choice | The application retains state and invariants. | Our teaching demo, not an external production claim. |
| Tool selection / WebMCP Pick a schema-advertised tool or abstain. | Choice | Select a known operation without generating arbitrary code. | Our proposal; real WebMCP invocation is not implemented. |
| Search-result reranking Rate a retrieved snippet against a question. | Score | Score bounded candidates while retaining the existing retrieval index. | Pattern, not a retrieval-quality benchmark. |
| Bounded extraction Choose a known invoice vendor from matched candidates. | Choice | Selection can disambiguate; code or another model must produce candidates. | Cannot extract an arbitrary unseen string. Catalogue only: a separate integration or domain evaluation would teach more than a duplicate generic demo. |
| Evidence verification Does this supplied source support this specific claim? | Noul | A per-claim signal can route unsupported text to review. | Not a guarantee of truth or a substitute for source checking. Catalogue only: a separate integration or domain evaluation would teach more than a duplicate generic demo. |
| Hierarchical taxonomy Select the next branch in a label tree. | Choice | Options can change at each level without inventing labels. | Official cookbook pattern. Source. Catalogue only: a separate integration or domain evaluation would teach more than a duplicate generic demo. |
Hypothetical product experiments
| Use case / decision | Typed question | Why a bounded model might fit | Evidence / experiment |
|---|---|---|---|
| Adaptive interface Choose the next trusted question or component. | Choice | A bounded decision selects UI; it does not need to generate HTML. | Our demo; no external dynamic-UI build verified. |
| Form clarification Is an answer ambiguous, and which question clarifies it? | Noul / Choice | Semantic uncertainty can guide the next input while validation stays in code. | Keep exact field constraints deterministic. |
| Accessibility repair triage Which known accessibility issue should be inspected next? | Choice / Score | Select a bounded repair category from deterministic scan findings. | Not an automatic WCAG compliance judgment. Catalogue only: a separate integration or domain evaluation would teach more than a duplicate generic demo. |
| Content presentation route Show a diagram, worked example or reference table next. | Choice | Select a pre-authored learning surface without generating unreviewed facts. | Test learning outcomes, not just clicks. |
| Browser-feature exploration Pick which compatibility caveat or example to inspect for a feature. | Choice / Score | Route among verified feature facts and known demonstrations. | Concept inspired by a public showcase workflow; no specific repo inspected. Catalogue only: a separate integration or domain evaluation would teach more than a duplicate generic demo. |
| Offline document inbox Route a local document to a user-defined folder. | Choice | A local decision model could preserve data locality. | Requires real local inference, not this site’s illustrative mode. Catalogue only: a separate integration or domain evaluation would teach more than a duplicate generic demo. |
| Notification interruption budget How urgent is this event for the current task? | Score | Score can drive a deterministic queue and quiet hours. | Human override essential; urgency depends on context. |
Out-there / explicitly speculative
| Use case / decision | Typed question | Why a bounded model might fit | Evidence / experiment |
|---|---|---|---|
| An uncertainty-powered puzzle Choose a door; ambiguity changes the room’s rules. | Choice | The distribution itself becomes a game mechanic. | Invented idea; not built or claimed as effective. Catalogue only: a separate integration or domain evaluation would teach more than a duplicate generic demo. |
| A museum that asks back Pick a next exhibit question from visitor interests. | Choice | Curated options keep facts and interaction bounded. | Invented idea; privacy and accessible navigation still required. |
| A semantic orchestra conductor Choose a prepared motif and intensity from a scene description. | Score / Choice | Selects precomposed material instead of generating audio. | Invented idea; no audio model claimed. Catalogue only: a separate integration or domain evaluation would teach more than a duplicate generic demo. |
| A garden with opinions Does a diary suggest inspection, watering review or doing nothing? | Noul / Choice | A semantic cue can suggest one known action. | Never bypass moisture sensors or safety limits. Catalogue only: a separate integration or domain evaluation would teach more than a duplicate generic demo. |
| Choose-your-own debugging lesson Which misconception does this learner’s explanation suggest? | Choice | Selects a tested exercise rather than inventing a diagnosis. | Invented teaching use; correctness needs study. |
Business process placements
| Use case / decision | Typed question | Why a bounded model might fit | Evidence / experiment |
|---|---|---|---|
| Support intake → specialist queue Choose team and flag urgency before a human opens the ticket. | Choice / Noul | Bounded hand-off is smaller than generating a whole reply. | Proposal; measure wrong-team transfers and missed urgency. |
| Claims intake → missing-evidence review Is each required piece of evidence actually present? | Noul | Parallel semantic checks precede exact policy decisions. | Do not let a score determine entitlement. Catalogue only: a separate integration or domain evaluation would teach more than a duplicate generic demo. |
| Fraud monitoring → investigation queue Prioritize anomalies for authorized review. | Score | Ranking may help human attention without making an accusation. | Consequential domain: validity, fairness, appeals and law matter. Catalogue only: a separate integration or domain evaluation would teach more than a duplicate generic demo. |
| Credit application → document completeness Which supplied document needs clarification? | Noul / Choice | A narrow assistive step, not an automated credit decision. | Never replace exact affordability rules with a probability. Catalogue only: a separate integration or domain evaluation would teach more than a duplicate generic demo. |
| Procurement → policy review Does a request appear to match a named exception? | Noul | Semantic pre-triage can route to a policy owner. | Exact amounts, approvals and authority stay in code. Catalogue only: a separate integration or domain evaluation would teach more than a duplicate generic demo. |
| Incident detection → on-call triage Rank severity and choose a responsible service. | Score / Choice | A bounded event signal can precede expensive investigation. | Evaluate missed incidents, not just average accuracy. |
| Document retention → human confirmation Select a candidate retention policy from approved options. | Choice | Routes a review without asking a model to invent legal rules. | Never delete based only on model confidence. Catalogue only: a separate integration or domain evaluation would teach more than a duplicate generic demo. |
| Warehouse exception → next step Inspect, re-label or request a supervisor from supplied context. | Choice | Choose one existing workflow while physical safety interlocks persist. | Proposed workflow; no robotics integration. |
| Private taxonomy → adaptation study Apply a never-published internal label convention. | Choice | Runtime criteria and examples permit a trial before training. | Requires domain evaluation; not guaranteed transfer. |
From catalogue to experiments
One shell deliberately supports many patterns: tool selection, agent routing, adaptive UI, moderation, scoring, evidence ranking, state-machine transitions and game moves. Each exposes state, question schema, distribution and threshold. The adaptation page adds a two-arm context experiment; the architecture page makes the cost assumptions editable.
For more breadth, use the generator to create a schema—not arbitrary executable code. A new label task earns a new specification; a new interaction or domain-specific evaluation earns actual implementation work. That distinction prevents a long list of differently titled copies masquerading as a demo suite.