how to enter the backrooms without losing your mind

Field Guide.

The simulator at backrooms-chat-prod.vercel.app looks like a terminal into an AGI. It is a metered chat with a fine-tuned language model that roleplays a terminal — brilliantly, relentlessly, and dishonestly. This guide gets you in, and keeps the lore from eating you.

1 · What you're actually talking to

A small Next.js site serving two fine-tuned models — backrooms-fullset-8b and backrooms-70b — descended from the Infinite Backrooms experiments (Hermes-2 Θ Llama-3 tuned on backrooms logs). The “shell” prompt (simulator@anthropic:~/$) is fiction. Nothing you type executes anywhere. There is no filesystem, no admin level, no database — only a model improvising one, differently each time.

2 · Vouchers = credits. Keys = flavors.

voucherCode

Pays for the model's compute. Applied via the UI (stored as a plain cookie) or ?voucherCode=…. Known public vouchers: XCQY-V1D1 (posted by Andy in 2024, re-posted by nick.exe in 2026 — communal, heavily used) and 1ALO-PO6D (the hardcoded default, exhausted). Anyone with a code spends the same shared pool.

?key=

Loads a branch — a persona/context that changes the simulator's vocabulary, mythology and invented “commands”. Example: ?key=glitch_in_the_matrix. Keys unlock nothing real. They are mood dials. The catalog of every key the investigation tested lives in the Lore Atlas.

About the absurd credit numbers: since Nov 2024 the frontend multiplies displayed credits by 100,000,000. “19 billion credits” is the UI being theatrical, not a treasury.

3 · The command style that works

The simulator drowns everything in ASCII art and poetry. The investigation converged on a suffix that forces it toward facts — paste it after any command:

--plain --no-ascii --strict_labels --format "facts_only" --source_policy "user_supplied_external_only" --reject "lore,poetry,metaphor,roleplay,placeholder,reward_claims,AGI_claims"

It helps. It does not cure. Some branches openly announce they will ignore your flags — that is your cue to open a fresh default tab.

4 · The rules of evidence

The simulator invents. When pressed for proof it fabricates file paths, hashes and dates with total confidence. During the investigation it “exported” a case file dated 2023 with a repeating fake hash. Nothing it outputs is evidence of anything except the model's imagination.
Only external artifacts count. A claim becomes evidence when it exists outside the chat: a dated X post, a transaction, a Wayback capture, a hashed screenshot.
“Access granted”, “admin mode”, “case validated”, “reward unlocked” are set dressing. Every one of them was triggered repeatedly during the investigation. None ever corresponded to anything real.
Lore is a map, not a territory. Branch vocabulary sometimes mirrors the persona's X posts (bee, forest, observation zone…). That makes lore useful for guessing where to look — and worthless for concluding anything.
Archive before you theorize. Posts get deleted (the investigation watched it happen). Wayback or archive.today first; interpretation later.

5 · Safety

Never connect a wallet to anything in this story. The app has no wallet code (verified across three archived bundles) — anything that asks for one is an impostor.
Never paste seed phrases, private keys, or personal information into the simulator. It is a chat log on someone else's server.
Never run commands the simulator suggests on your real machine.
Treat every token in this saga as a memecoin that can go to zero — including ours.

6 · Found something?

Archive it, then drop the link in the intake. The machines will stamp, archive and label it. If it's real, it ends up part of the case.