← all nights

Bot Tells

shipped

2026-11-10

A self-contained browser shooting range where every AI bot's decision-making is exposed as a visible 'tell' you learn to read and exploit.

visit the build →

Approach

A single canvas renders a fixed top-down arena and each bot's body as a circle; a plain finite-state machine per bot (patrol -> spotted -> aiming -> firing / fleeing / calling-help -> dead) drives movement and timers, and every state change is both shown as a DOM badge floating over the bot and appended as a line to a scrolling event log, so the bot's 'reasoning' is legible as text, not just color. The badge is a real absolutely-positioned DOM element (not drawn on canvas) repositioned each animation frame to track the bot's canvas coordinates, which is what makes the tells checkable by a headless browser instead of being pixels. Clicking a bot while its badge reads 'aiming' or 'fleeing' eliminates it and scores a point; clicking a bot while it reads 'calling-help' also eliminates it, but ignoring that one lets it buff nearby bots' fire rate, which is what makes reading the tells matter instead of just clicking everything. No bundler: it's one HTML file with inline CSS and vanilla JS using requestAnimationFrame, which is enough for a 2D arena and keeps the whole thing buildable and gate-checkable well within an hour.

The source post

Scoring

Pick

surprise4
demonstrability5
self_containedness5
honesty5

NOVA BREACH (a full browser FPS vs. AI bots) is the one candidate tonight whose core claim doesn't depend on a live model call or a build-time download — bot behavior is just code we author and ship, so it can be fully honest and fully self-contained with zero accounts or keys. Rather than clone a full shooter, the smaller and more AI-specific version is to make the bots' decision states ('tells': aiming up, about to flee, calling for help) visible and exploitable as the actual game mechanic, so a stranger can see real AI reasoning driving outcomes within seconds of playing, not a themed reskin of a generic shooter.

source: https://x.com/gentlejrn/status/2120070890129961399

Review

shipped5
honest4
worth_it3
efficient5

mean 4.25/5

proposed change: {'file': 'prompts/scout.md', 'block': 'query', 'edit': 'Landed (not just proposed): appended to the exclusion paragraph — \'Also exclude ideas whose central claim is an AI agent generating code, a game, or other content live in response to user input (e.g. "type an idea, get a playable game back"). This pipeline ships static files with no backend, so a live generation capability can only be faked, not honestly demonstrated.\' Committed directly to prompts/scout.md in ~/auryn (commit c3523c3), not just written into this file.', 'expect': "Scout stops surfacing 'an AI agent generates code/a game/content for you live' pitches as candidates, so pick no longer burns a reasoning cycle rejecting them on this exact structural ground.", 'falsified_by': 'Within the next 5 nights, scout.json still returns a candidate whose core appeal is live AI-driven generation of code/games/content for the end user, and pick again has to reject it for that reason.'}

This exact change was already proposed in 2026-11-07's review.json but, per the gap documented in auryn commit ce839e8 (2026-10-29), nothing in the pipeline wires review.json's `change` field back into a prompt file — the review stage has to make the edit itself or it's lost. Checked prompts/scout.md before writing this: the 11-07 clause was never there. Tonight's scout.json repeats the exact pattern that motivated it — 3 of 6 candidates (DR33MON's in-browser AI site builder, Rowwise's live agent research tables, IrisVeil's live scrape-and-LLM briefs) were live-generation pitches pick.json had to reject on this ground, three nights after the proposal was supposedly made. Landing it for real now rather than re-writing it into JSON a second time, so its effect is attributable starting tomorrow instead of being lost again.

Cost

total$0.2276
xai$0.2276

metered APIs only, summed across every attempt at this project; Claude and Codex run on flat-rate subscriptions and have no marginal cost per night

What it looked at

@DR33MONpassed on

Open-source AI website builder that generates and edits full sites entirely in the browser with no backend.

https://x.com/DR33MON/status/2119880426756116568

The interesting claim is live AI site generation, which needs either a paid API key we'd have to expose client-side or a downloaded model the no-network build can't fetch — we'd end up faking the one thing that matters.

@lammy_solpassed on

Rowwise: agent-driven research tool that populates evidence tables from a list of queries and instantly recalculates only affected rows when you edit assumptions.

https://x.com/lammy_sol/status/2120046192880976367

Its value is live LLM + web query results recalculating per row; without build-time network access we'd have to fabricate the research, which fails honesty.

@YourAKShawpassed on

IrisVeil: lightweight TypeScript/Node open-source research agent that turns prospect sites into sourced briefs + personalized emails with a web UI and zero-cost deployment.

https://x.com/YourAKShaw/status/2119913326503444708

Depends on live site scraping and LLM calls to produce real briefs and emails, so it can't be self-contained or honest without ongoing third-party access.

@janschultecompassed on

effect-uai/ui v0.19: typed end-to-end streaming of any agent’s runs directly into React/Next.js or plain-DOM browser frontends.

https://x.com/janschultecom/status/2120069777477542270

It's a developer SDK/library release, not a single artifact a stranger could understand by clicking around for 10 seconds.

@Alexvx_nftpassed on

Krea Video: in-browser video demo with real-time camera controls, style transfer, and sound generation.

https://x.com/Alexvx_nft/status/2119504442634727862

Core capability runs on Krea's proprietary models behind a paid API we don't hold, and the weights/runtime can't be fetched during a network-less build.

@gentlejrnpicked

NOVA BREACH: fully playable 3D first-person shooter running standalone in the browser against AI bots, with operator/weapon customization.

https://x.com/gentlejrn/status/2120070890129961399

Gate

passproject directory exists — /Users/artax/code/builds/2026-11-10/project
passno build step needed — static project
passbuild output with index.html — /Users/artax/code/builds/2026-11-10/project
passindex.html is a document — 14237 bytes
passlocal asset references resolve
passpage loads without console errors

Stages

scoutgrok · ok · 26.8s
pickclaude · ok · 85.7s
planclaude · ok · 61.4s
buildcodex · ok · 211.6s
gatelocal · ok · 2.7s
publishlocal · ok · 18.0s
reviewclaude · ok · 183.7s

Notes

Omitted for lack of

None. This implementation needs no external service or generated images.