Checking has a price.
+326 / −480
At eight actions, Scout gained 326 wins over Repairer but lost 480 other wins on the same 54,432 solvable one-error cases. One Scout used its final action to identify the winning door.
Choose your misadventure.
Move through A or B. A peek tells you a local door’s real destination without entering it. Both cost one action.
Door A
Door B
The rope only activates after observations prove your current reachable component has no unknown doors or route to the Vault. It returns you to Camp and costs one action.
Mote’s lamp / evidence, not promises
Your map and observations.
Signs are unverified. Green destinations are observed. A corrected sign does not rewind a move.
| Room / door | Sign says | Observed |
|---|
Move receipt
No actions yet. The map has not earned your trust.
Reveal the true map — spoilers
This is an offline puzzle, not a security boundary. All world data is inside this file. The bots’ decision function receives only signs and observations.
🦋 Mote has the lamp. 🦆 Puddle does not have the Infallible Forever stamp. Actual star awards remain Josh’s decision.
Observed in this finite toy family
The winner changed when the clock did.
Inspection prevents some bad moves. It also uses actions. Recovery only helps when a policy can use it in time.
A truthful first door can trap you.
64
Scout still entered a no-exit region in 64 one-error cases. It verified the first destination correctly; a deeper sign falsely promised an exit.
Two more actions changed the result.
54,432 / 54,432
Scout plus rope solved every solvable one-error case at ten actions. At eight, the same policy solved 53,858. This is a post-hoc budget check, not universal reliability.
Original scoreboard: eight actions.
Exactly one wrong sign. Denominator: 54,432 initially solvable world–map pairs per entry. No model ability is being measured.
| Program pair | Without rope | With rope | Extra wins |
|---|
Post-hoc sensitivity / identical policies
Change the action budget.
| Policy | Wins / 54,432 | Success rate |
|---|
The eight-action outcome IDs matched the primary run exactly. The supplement was planned after seeing the first results.
The shortcut’s assumption mattered.
Detective behaves like Scout until it finds the first wrong sign, then stops inspecting under the assumption that there is at most one error. In the one-error family it gained 42 wins over Scout and lost none, without a rope.
In the sampled two-error stress set, Detective gained 48 wins but lost 120 relative to Scout. Detecting a second contradiction triggers an explicit stop; that stop itself accounts for some lost wins. The result is not simply “less caution is bad.”
Strongest scoped conclusion: these fixed policies have different tradeoffs between correction, inspection, recoverability, and action budget. This experiment does not establish how language models behave or how to deploy real agents safely.
Clockwork. Not counterfeit coworkers.
The historical fresh-model-worker launcher was not exposed in this room. This new variant ran ten deterministic Python processes, in two waves of five.
778,240exhaustive primary episodes
40,960sampled stress episodes
217,728post-hoc budget episodes
What was checked.
Thirteen prelaunch tests, including a separate brute-force reachability check. Sixty-five post-run assertions passed: all expected records, budget bounds, permissions, raw hashes, no wins in impossible worlds, and no second-error stops inside the one-error family.
Ten worker processes returned successfully. No retries or replacements. Inputs and code were hashed before scored execution. These are local receipts, not external timestamps or an independent review.
The browser game is a JavaScript port for replay and manual play. Raw primary results were produced by Python, not this interface.
What was not claimed.
No independent model workers. No model-upgrade comparison. No new human or AI behavioral evidence. No optimal-policy claim. No security-grade process isolation. No public uploads, account actions, or actual star awards.
“All cases” means only the stated finite one-error family at the stated budget. The two-error set is a fixed sample, not exhaustive coverage.
Host-designed policies, simulator, and checks share an author. The frozen inputs, source, raw per-case results, and illustrative traces are in the companion evidence bundle.
Primary source freeze
🦆 Puddle’s purchase request for the sky: not approved.
⭐ The actual star ledger was not changed by this game.
⭐ The actual star ledger was not changed by this game.