Patterns: gates, simulation, personas and other shapes (P38+)

TL;DR Seven shapes from the 2026-09-23 sweep that fit no earlier pattern. Worth copying: wake gates (P38), persona panels as a relative pre-screen (P40), engine-graded tutoring (P41), label-tuned questions (P42). P43 and P44 are poor: read them to recognise the shape, not to build it.

How to read this

Same format and fit scale as Patterns: agent internals, context and coding agents: strong = bounded options, semantic judgment, code owns the rest (still start with an experiment); workable = real, caveat named; poor = a weakness from Jev 1.13 jaggedness: known failure modes or System One Models sits in the middle. Posted numbers are author-reported and unverified. Full index: Decision patterns from the community (with fit verdicts).

Gates and triggers

P38 Wake and interjection gate

• Decision does this event deserve a full LLM turn, or should the assistant speak at all right now
• State what the agent said it was waiting for, plus the event (email, webhook, fresh observation) or the recent usage/transcript
• Ask Choice over wake | not yet | unrelated; for assistants a Noul "is the user stuck / addressing me?"
• Code fail open: user messages, timeouts, errors, missing probabilities and a max-skips counter all wake without asking; sleep intervals, dwell times and numeric targets ("below $300") computed in code
• Fit strong for "event vs stated goal" (one bounded judgment in front of an expensive call); workable for "should I speak", since hesitation is timing data (math, weakness #2) and events are attacker-writable (#6)
• Map Confidence-gated routing, Patterns: agent internals, context and coding agents P36 (a sibling: that one triages findings, this one gates the wakeup)
• Seen shitianfang/wakegate (repo): skips only below 0.2 on "wake"; author reports 21/21 on 21 self-written scenarios, p50 253 ms; the first yes/no wording scored 16/21, erring on numeric "not yet" cases, so the three-way Choice replaced it; savings unmeasured. Calls go via Vercel AI Gateway typesafe-ai/jev (gateway route, not in TypeSafe's docs). @sotak: an in-product Clippy that wakes only when the user seems stuck. @npaka123 (via @StingCloudLab): a voice assistant with no wake word that answers only when context calls for it; StingCloudLab flags always-on listening as an unanswered privacy question

Simulation and synthetic crowds

P39 Branched rollouts: fork the state, Jev steps each branch

• Decision the next move or state transition inside one simulated branch
• State the branch's snapshot as compact JSON plus the legal moves or the moves an LLM proposed
• Ask Choice over move ids per step, per branch; many branches in parallel
• Code snapshot, fork and discard branches; score survivors with the simulator's own outcome (alive / dead), never with Jev's opinion of the future
• Fit workable when you own a real simulator and Jev only picks moves (Patterns: judging, search, documents, real-time and markets P24 plus search); poor when Jev is the simulator of the world, e.g. predicting how a patient responds: that is an outcome the text does not contain (Decision patterns from the community (with fit verdicts) anti-patterns). Not clinical advice
• Map Speculative fan-out, Patterns: agent internals, context and coding agents P05
• Seen @theappcypher ("Mario Never Dies": Jev picks every move; on a death the microsandbox VM forks into 4 timelines and the survivor becomes canon). @chris_not_busy: Grok 4.7 proposes treatment moves, Jev makes state decisions in under 0.3 s across 20+ futures of a simulated cancer patient (author-reported; no validation shown)

P40 Synthetic persona panel

• Decision how would this kind of person react to this text: scroll, stop, like, annoyed, buy
• State one persona line (age, job, city, interests, temperament, budget) plus the artefact
• Ask one Choice per persona over reactions, many personas per request; optional Nouls about the text itself
• Code generate personas deterministically; draw each reaction from the returned probabilities with a fixed seed so reloads agree; waves, spread rules, audience filters and every count in code. Let an LLM write variants and the panel rank them
• Fit workable as a relative pre-screen (A vs B, which segment is annoyed); poor as a forecast of real reach: @PrajwalTomar_ scored 20% on reach bands vs 76% for a constant guess (Field reports: independent evaluations, critiques, open replicas). Persona reactions are Jev's guess about a described person, not survey data
• Map Cookbook: Parallel questions, Patterns: marketing, sales, GTM, content, support and ops P21 (scores the artefact; this simulates an audience)
• Seen @TheMattBerman: 723 ads × 30 buyer personalities = 21,690 stop-or-scroll decisions for 22 cents. @SimAudience: two launch tweets voted on by 4,000+ survey-participant profiles in 12 s. gaborishka/jevtown (repo): 10,000 computed personas, 100 per request, waves of 600/1,500/3,000/rest; author reports 200 personas in one request answer the same as one, six of a person's own past answers lift Jev's guesses about them from 0.49 to 0.72, a check costs about a cent to ten cents. A panel only knows the attributes you model: its audience filter turned "tech founders of early-stage B2B SaaS" into everyone who works in business. Ships an MCP server so an agent writes variants and the town ranks them

Judgment next to a source of truth

P41 Engine owns truth, Jev owns the judgment: grading and coaching

• Decision given an objective diff between attempt and truth, what kind of error is it and what coaching move fits
• State the reference, the attempt, and the facts a deterministic tool already computed (engine evaluation, word alignment, timestamps)
• Ask Noul "is this the same word/idea?"; Choice over error types or coaching moves; Score on an ordered rubric for "did the meaning survive?"
• Code the engine or aligner finds the mismatches, so only those reach Jev; totals, pace, weights in code; templates or a small writing model phrase the feedback
• Fit strong — Jev never computes the truth or writes the words, only the soft call between. Jev reads text only, so acoustics (pronunciation, accent) stay out of scope
• Map How to build software with System One, Cookbook: Knowledge graph entity alignment
• Seen wquguru/dasheng (repo): streaming ASR hears a read-aloud, alignment picks the mismatched words, Jev judges misread vs ASR spelling variant vs self-correction. Its README calls the stack offline, but its own config sends Jev calls to ZenMux (https://zenmux.ai/api/v1/systemone, model typesafe/jev-1.13, not in TypeSafe's docs; only the ASR self-hosts). JoelLewis/game-coach (repo): "Stockfish owns truth, Jev owns judgment" in a chess coach; status is a scaffold and feasibility spike, not a result

Building and tuning the questions

P42 Tuning questions against your own labels

• Decision which wording, options and rubric make a question agree with your people
• State your dataset rows (selected columns only) and accumulated human labels, optionally with rationales
• Ask whatever the function is: Noul (binary), Choice (multiclass), several Nouls (multilabel), Score (ordered rubric)
• Code run the set, surface the least certain rows plus a random audit sample for labelling, let an optimiser (an LLM) propose a new definition, show the diff, a person accepts or rejects; keep a held-out split
• Fit strong — it moves wording, the main lever against literal reading (#1) and contradictory criteria (#7), out of guesswork. Risks: overfitting a small label set, and one author writing both questions and labels
• Map Testing and evaluating a Jev workflow, Writing instructions and criteria that Jev reads correctly, Confidence vs probability
• Seen sutro-sh/jev-align (repo): CLI using GEPA; 5-20 labels a round, optional 20% held-out set; a higher training score never auto-accepts; can publish functions with their labelled data to a public registry, so never push data you cannot share. Runs direct, via Vercel AI Gateway or Cloudflare Workers AI (gateway routes unverified; not affiliated with TypeSafe). The same habit by hand: wakegate picked its wording from three candidates on a dev set (P38); jevtown drops any question that failed its measurement (P40)

Poor fits worth recognising

P43 Compute knob inside another model

• Decision how much compute a layer or step of a generative model gets (e.g. an attention sparsity rate)
• State whatever describes the layer at that step; for Jev it must be text
• Ask Choice over preset levels (1% | 3% | 5% | 10%)
• Code the model loop, timing and quality checks
• Fit poor — importance of a layer is a numeric property of tensors (#2), Jev sees no tensors, and a network call sits inside a GPU loop. No ablation against a fixed schedule was posted, so the speed-up may come from sparsity alone (inferred). Jev-shaped variant: one call per job on the prompt text ("motion-heavy scene?") to pick a preset
• Map Patterns: agent internals, context and coding agents P02 (effort routing, where the judgment is about the request, not the tensors)
• Seen @sep_is_heim: Jev judges importance per layer (4 steps × 49 layers) of MiniMax H3 video generation and picks the sparsity; reports 6 min 7 s → 3 min 34 s, 41.7% faster on an RTX 4070

P44 Choice probabilities as a text sampler

• Decision which symbol or token comes next in a reply
• State the user's question plus the reply so far
• Ask Choice over an alphabet or token list plus STOP; draw from the probabilities, append, repeat
• Code the whole decoding loop: sampling, temperature, beams, buckets beyond 255 options (the documented cap, verified — Choice questions)
• Fit poor — generation is weakness #9 and costs one request per symbol. Its authors call the results "hilarious" and "kinda works". Use a generative model to write and Jev to pick among finished candidates (P41, Patterns: judging, search, documents, real-time and markets P16)
• Map Jev 1.13 jaggedness: known failure modes
• Seen kyle-pena-nlp/jevchat (repo), erik-dunteman/ChatJev (repo). Tricks that transfer to ordinary Choice use (author-reported, unverified): list options as the finished candidate text, not a fragment the model must splice in (roughly tripled top-1 on character alphabets); shuffle the option order to cancel what the author calls position bias, or average several shuffled copies sent as parallel questions in one request (option-order sensitivity also in Field reports: independent evaluations, critiques, open replicas); for more than 255 candidates, split into buckets each with an OTHER escape, then ask once over the winners

Related

• Decision patterns from the community (with fit verdicts) — index of every pattern ID; Consult guide: could Jev help this project? — the advising procedure
• Patterns: agent internals, context and coding agents, Patterns: judging, search, documents, real-time and markets, Patterns: marketing, sales, GTM, content, support and ops, Patterns: browser, computer use, voice and product UI — P01-P37
• Field reports: independent evaluations, critiques, open replicas, Measurements, access routes and open replicas — numbers and failures behind the verdicts
• Jev 1.13 jaggedness: known failure modes, System One Models — what the fit verdicts are judged against

Sources

• raw/x-repos/shitianfang__wakegate.md (https://github.com/shitianfang/wakegate)
• raw/x/sotak-2100701152824185319.md (https://x.com/sotak/status/2100701152824185319)
• raw/x/StingCloudLab-2102202720341934217.md (https://x.com/StingCloudLab/status/2102202720341934217)
• raw/x/theappcypher-2101095181382721998.md (https://x.com/theappcypher/status/2101095181382721998)
• raw/x/chris_not_busy-2102349211127955925.md (https://x.com/chris_not_busy/status/2102349211127955925)
• raw/x/TheMattBerman-2101439340588974096.md (https://x.com/TheMattBerman/status/2101439340588974096)
• raw/x/SimAudience-2101117203060019431.md (https://x.com/SimAudience/status/2101117203060019431)
• raw/x-repos/gaborishka__jevtown.md (https://github.com/gaborishka/jevtown)
• raw/x-repos/wquguru__dasheng.md (https://github.com/wquguru/dasheng)
• raw/x-repos/JoelLewis__game-coach.md (https://github.com/JoelLewis/game-coach)
• raw/x-repos/sutro-sh__jev-align.md (https://github.com/sutro-sh/jev-align)
• raw/x/sep_is_heim-2101603192664740330.md (https://x.com/sep_is_heim/status/2101603192664740330)
• raw/x-repos/kyle-pena-nlp__jevchat.md (https://github.com/kyle-pena-nlp/jevchat)
• raw/x-repos/erik-dunteman__ChatJev.md (https://github.com/erik-dunteman/ChatJev)