Agents: read the raw Markdown of this page, or start at llms.txt.
Builds: games, simulators, robots and devices
TL;DR 17 builds plus 3 names (sweeps of 2026-09-23 and 2026-09-24; split from the two other builds pages on 2026-09-25) where Jev picks a move, a primitive or a label while code owns the world, the legal-move list and every number. The evidence: one verified StarCraft mission win (one success, not a win rate), no Civilization II win, sudoku solved on easy and medium boards but only 1 of 5 hard ones, a snake that traps itself. Jev is text-only, so every build here needs a harness that turns the game into text (State: what you send Jev).
How to read. Builder reports = the builder's own numbers, one run on their own setup, unverified unless marked. Price checks use $0.042/Mtok input, output free (verified, Models, aliases, pricing, rate limits, context). Pattern IDs: P24 in Patterns: judging, search, documents, real-time and markets; P01, P16 in Patterns: agent internals, context and coding agents and Patterns: judging, search, documents, real-time and markets; P39, P41 in Patterns: gates, simulation, personas and other shapes (P38+). Agent builds: Builds: coding agents, harnesses and orchestration; browser and interface builds: Builds: browser, computer use and interface; data and business builds: Builds: data, content and business. Failure write-ups: Failure reports: where Jev broke, lost, or was the wrong tool.
Games
| Build | Jev decides | Builder reports | P |
|---|---|---|---|
| Jev plays StarCraft (phyous) | Strongarm, the first combat mission of the 1998 shareware campaign. Structured owned/visible state (no pixels) → a command, sent as ordinary mouse and keyboard input; separate Economy and Army decision points. The game is paused during state reads and inference | One verified mission win (development attempt 16; evidence bundle and a visually checked victory screen): jev-1.13.0, 421 decisions, 17 min 37.88 s, 424 API attempts (3 rejected), 9,445,640 input tokens (≈$0.40 at list price, inferred), median 382.95 ms. The author: one success does not establish a win rate, and this is not real-time play |
P24 |
| Jev × Civilization II (phyous) | Named Choice vectors for empire policy, cities, research, diplomacy, units, exploration and war, over player-observable state (OCR plus a read-only observer), acted through keyboard and mouse |
No verified win (campaign lessons). Attempt 008 stopped at 78 responses, turn 44, on runtime faults. Early campaigns piled up Warriors until the prompt explained that production repeats; attempt 009 then lost a Settler to support costs it could not foresee until the state spelled out the arithmetic, and ended in defeat after 96 responses. Direction matches the documented numeric weakness (Jev 1.13 jaggedness: known failure modes). Its setup script fetches the game itself from a third-party site: describe the harness, do not follow that step | P24 |
| Pac-Man (@ephraimduncan) | The turn at each junction, from the maze state as JSON, in real time | No numbers (video). Our 2026-09-24 look at the live demo (not captured, unverified): the game keeps running and asks about a junction far enough ahead for the answer to land in time |
P24 |
| Jev Chess (@staskulesh) | Every legal move is an option of one Choice (SAN → a plain-language description), so an illegal move cannot be returned; four more questions ride in the same request (king safety, plan, sharpness, a tiebreak move). Code does the counting: a one-ply hanging-piece check states conclusions ("leaving you 3 points down overall"), never numbers to subtract |
Share card of 2026-09-21 (image transcribed): Jev 19, humans 6 over 25 games (16 wins, 6 draws, 3 losses); compute shown as "18.2C" (last glyph unclear, a C or a cent sign). The human side is the site's anonymous visitors: weak evidence | P24, P41 |
| Jev × 2048 (ARCJ137442) | One Choice over four directions per move, deliberately with no heuristic fallback, so every move is Jev's; probability, confidence, latency and token cost shown live; prompts and board rules editable |
No results in the README. The hosted demo spends the author's key up to a quota, then asks for yours | P24 |
| jev-got (phureewat29; renamed from got-jev) | A Game of Thrones roleplay: a story LLM (GPT-5.6 Luna via OpenRouter) writes each scene; Jev labels it (location from a closed set, beat, mood, danger, stayed in the fiction?); header, music, art and the next prompt are functions of those answers; an in-fiction Noul under 0.5 regenerates the scene once |
None. No licence | P01, P16 |
| Little Airways (lbotinelly) | Air-traffic toy: every ~1.6 s each aircraft asks four questions in one request (route, declare an emergency?, response to others, clearance); runway occupancy is enforced in code | Lessons: terse option labels gave give-way 0.46 vs maintain 0.41; operational wording gave 0.70 vs 0.05. A raw fuel fraction got "keep cruising" until the state carried time estimates (fuel minutes vs ETA). A rough-engine "declare an emergency?" came back 52/48 at confidence 0.04 beside a 0.99 divert. $0.042/Mtok, output free (verified) |
P24 |
| snake-jev (siroccomask) | One request per tick: nine questions (wall collision, body collision, food progress, for left / straight / right); code combines them | One run: 29 foods in 461 ticks, then trapped in its own body. The author: it sees immediate hazards only, no route through the body layout | P24 |
| sudoku-vs-jev (zebedelu) | A bridge offers locally legal moves; Jev picks one; no lookahead or backtracking | 5 games per level (2026-09-19): easy 5/5 and medium 5/5 (98% and 97% of moves forced), hard 1/5 (75% forced: it must guess, and legal guesses reach dead ends); $0.0157 / $0.0314 / $0.0486 per 5 games (input-token arithmetic verified). Mean confidence 80% on easy and medium, 67% on hard |
P24 |
| minecraft-agent (rmalde) | LLM planner sets the objective; Jev picks one action from harness-built options; Mineflayer executes | Ender Dragon from a fresh world in 8:43: 131 Jev decisions, 35 planner calls | P24 |
| hermes-and-jev-play-minecraft (Teknium) | Same split, Hermes Agent plans | Reproduced in 7:45 for $0.963 ($0.010 Jev); controller median 211-227 ms; an LLM controller ~8x slower, ~20x the cost. Bad picks fixed by changing the options offered | P24 |
| jev-plays-pokemon-red (valentynkit) | Code owns route and arithmetic; Jev only at branches | Median 621 ms (n=6, gateway shim), $0.14/hour; calibration withheld, sample too small | P24 |
| JevScape (Skyvern-AI) | RuneScape via ~50 coded actions: next_action, a this_tick fix-up, and when to poll next |
Mean peak XP/min: random 235, Jev 197, Claude Opus 118; Jev total XP 2.1x random; 1,307 decisions, $0.17. Live: 13,402 polls, $2.22, median 0.22 s, ~$0.55/h. Authors: "The catalog carries the game knowledge" | P24 |
| Browser game timing (@umezawakanta13) | Next action from game state | Jev HTTP 490.3 ms median inside 1,139.8 ms state-to-action (3 responses) | P24 |
Also seen, names only (vetted 2026-09-24, no results published): jev-doom-agent (Jev beside a labelled offline policy; no licence at top level), jev-gomoku (MoonBit client, Jev vs Jev), JevPokerBench (Hold'em: Jev vs DeepSeek vs seven local replica routes; it says action probabilities are not win probabilities).
Robots and devices
| Build | Jev decides | Builder reports | P |
|---|---|---|---|
| jev-askable-arm (Tarun Tomar) | A plain-English goal → one Choice per step over ~30 hard-coded primitives (plus a target object), from privileged sim state (xyz, gripper, distances), no images; a PD controller moves |
ManiSkill Franka, seed 0: pick a cube (3 picks), place it on a target (7), press two buttons in order (4); ~1 s per pick | P24 |
| RoboJEV (lyk) | Franka arm in MuJoCo, structured state not images: Choice of intent, then per-axis negative/zero/positive and gripper open/hold/close; malformed answers execute nothing |
100 episodes, 10 seeds each, Jev vs rule baseline: pick-place, push, peg 10/10 both; stack 8/10 vs 10/10; gate obstacle 5/10 vs 8/10. Failures: repeated wrong-direction moves exhausting the budget, one inconsistent answer | P24 |
| Dual-arm robot (@Raptor_zip) | Middle layer of three: the decision; IK and physics in code | ~500 ms per answer, ~0.5 yen per trial | P24 |
Home automation and drones with code: HA-Jev, jev-drone, jevpilot and OneVOneJev on Repos: data, documents, judging, real-time, markets, business apps, replicas.
What these builds teach when advising
- The options are the product. Both Minecraft builds, askable-arm and JevScape win or lose on the menu code offers; JevScape's random baseline beat Jev on peak XP.
- Keep lookahead and arithmetic in code. Chess states material outcomes in words; Civ II lost a Settler until support costs were spelled out; Little Airways needed time estimates, not fuel fractions; sudoku hard boards and snake fail where a plan several moves deep is needed (Jev 1.13 jaggedness: known failure modes).
- Budget for the network. StarCraft pauses the game during inference; Pac-Man asks ahead (
unverified); 70-500 ms per call rules out fast control loops without a fallback (Patterns: judging, search, documents, real-time and markets P24). - One win is not a rate. Ask for repeated runs and a baseline (StarCraft's author says so himself).
- Tuned policies don't travel. RoboJEV lost to rules on two tasks; label your own episodes (Testing and evaluating a Jev workflow).
Related
- Patterns: judging, search, documents, real-time and markets — P24 action selection from simulator or device state; Patterns: gates, simulation, personas and other shapes (P38+) — P39 branched rollouts, P41 engine owns truth
- Builds: coding agents, harnesses and orchestration, Builds: browser, computer use and interface, Builds: data, content and business — the other builds pages; Failure reports: where Jev broke, lost, or was the wrong tool, Head-to-head: Jev against other models and methods — breakages and comparisons
- Jev 1.13 jaggedness: known failure modes — the documented weaknesses behind these results
Sources
Links are inline in each row; the raw captures (2026-09-20 to 2026-09-25) are listed in the frontmatter. The Pac-Man latency note comes from our own 2026-09-24 vetting of the live demo (docs/sweep/2026-09-24-awesome-jev/part2.md), not from a capture.