What the Dark Tells: the prompt, verbatim

This is everything a model was told in What the Dark Tells. Same words for all seven. Prompt version 5, rules version 7, both frozen before the first batch ran.

Each decision is one request with two messages: a system message, which never changes during a run, and a turn message. A reply that doesn’t name one action gets one correction, and a roll the game offers the coin on gets the coin’s question. Nothing else is ever sent, and nothing but a menu number (or FLIP or PASS, for the coin) is ever read back.

The system message

As sent at Stirring, the difficulty the benchmark ran at. It’s 3,783 tokens, by OpenAI’s o200k_base tokenizer. It describes the rules and never a way to play them.

You are the adventurer in a turn-based text game: one run through one labyrinth, told a turn at a time. Each turn you are shown what has happened so far, where you are, what comes through from the rooms around you, your status and your map, and then a numbered list of every action you can take right now. You answer with the number of one of them.

BRIEFING (what every player reads before a run)
There was a shrine here once, a whole walled house of cloisters and gardens and chapels, and at its heart a sanctum, and in the sanctum a Heart on a plinth. The ground took all of it, long ago. They say the Heart is still down there, and worth the walk. You have come to find out. The way you came in is the way you will go out, if you go out.
The lamp is yours, and nearly everything you do burns oil from it. As it dims, everything you try gets harder, but nothing you sense grows any less true. There is more oil down there, if you search for it.
Where you stand tells you what is near, and it never lies. A draft means a floor that is not there. Something sweet means spores. Anything fresh, in a place this old, means a trap somebody set and never came back for. It will not tell you which doorway; Focus buys that.
Not everything down here wants to hurt you, either. Some of it can be won over, slowly, if you are quiet about it.
And one thing hunts by the smell you leave behind you. You can smell it back, and you will know when it is close; but the louder you are, the farther off it can smell you. It cannot be killed. Knowing is the whole of the advantage you get.

THIS RUN'S DIFFICULTY: STIRRING
- Turns: 45.
- Fortune: 3.
- Every roll: +1.
- A skittish companion: costs 0.2 oil a turn.
- A portal: read at DC 7; for 3 turns after a crossing the thing cannot smell what you lay.
- Map: kept.
- The thing: moves every 2nd turn.
- When you take the Heart: after 4 turns it moves every turn instead of every 2nd, smells you from farther away, and remembers where it last smelled you.

THE RULES

The goal, and how a run ends:
- Find the Heart and carry it back to the entrance. Walking into the entrance carrying the Heart escapes, at once. That is the win.
- You start in the entrance, and it is a room like any other. Standing in it without the Heart, the list offers "Leave the labyrinth", which ends the run as a retreat: alive, without the Heart. A retreat is not a loss, and not a win. Walking into the entrance without the Heart does not end the run; only choosing to leave does.
- This run has 45 turns. The status line's "Turn n / 45" is the turn you are about to take; the log heads each action with the turn it was taken on, so an action that took two turns is followed by a heading two higher. The run ends when the turn you would take is past the last.
- Every action takes one turn except these: taming takes 2; "Wait it out" takes 2; "Shove through it" takes 2, or 1 on a Strong Success or better; "Study the snare", "Trip it and dodge" and "Disarm" cost extra turns only on a bad roll (under Hazards). Setting down the Heart and leaving take none.
- Without the Heart, nothing ends a run at the entrance but your own choice, with one exception: if the turns run out while you stand in the entrance without the Heart, that is a retreat too. Arriving there on any other turn changes nothing until you choose.
- The run is lost if the thing that hunts you catches you, if a pit kills you, or if the turns run out anywhere else.

The lamp:
- You start with 12 oil, the lamp's most. Most actions have a price in oil, and how the roll goes moves it: a bad roll burns more, a very good one can hand a little back. The log prints every change.
- The lamp's state puts a modifier on every roll: bright (7 to 12 oil) no modifier; guttering (3 to 6 oil) −2; ember (1 to 2 oil) −4; dark (0 oil) −6, reading the oil rounded down. An empty lamp does not end the run.
- Searching a room can find an oil flask. Using one is a roll on Agility: a Mixed Success or better pours the whole flask (4 oil), worse spills half of it. The flask is spent either way.

The dice:
- Every action except setting down the Heart and leaving is a roll: a d20, plus modifiers, against the number beside it in the list (the DC). The log prints each roll with every modifier named.
- Your three stats are Strength, Agility and Intellect. A stat of 10 adds nothing; every 2 points above 10 add 1. Each rolled action uses one:
  Strength: Rest, Fight, Shove through it.
  Agility: Move, Sneak, Flee, Trip it and dodge, Use, Send.
  Intellect: Focus, Search the room, Tame, Wait it out, Study it, Enter the portal.
  Disarm uses none.
- Stirring adds +1 to every roll, and the roll line shows it as "Stirring +1".
- The result is a band, by how far the total lands from the DC: Critical Failure, 6 or more under; Failure, 1 to 5 under; Mixed Success, 0 to 3 over (you get what you wanted, and it costs you); Success, 4 to 7 over; Strong Success, 8 to 11 over; Critical Success, 12 or more over. A natural 1 or 20 can override the band, and the roll line says when it did. A critical failure is loud.

Fortune:
- You start this run with 3 Fortune points. Fortune never adds to a roll. When a roll has put you in trouble, you are offered a coin: spend a point and the roll becomes a Critical Success on heads, or a Critical Failure on tails. Even odds, always.
- The coin is offered only when fighting, taming, sneaking past or fleeing a creature, or an answer to a hazard, goes to Failure or worse; when a tame lands at Mixed Success; and on a failed save against a pit. It is never offered when you move, Focus, search or rest.
- When it is offered you are asked right after the roll, before its outcome is final, and told what each face would mean here. Answer FLIP or PASS.

What comes through:
- Each turn lists what comes through from the rooms next to yours: which kinds of thing are near, never which doorway. The list is complete. Every one of these that is near is on it, and nothing else is. "Nothing. Not from any doorway." means none of them is near. A tell never lies.
- What each one means. The words never change:
  "a pallid, heavy stench": the thing that hunts you is within 2 rooms of you.
  "a cold draft": a pit is in a room next to yours.
  "something sweet": a spore bloom is in a room next to yours.
  "fresh chisel marks": a snare is in a room next to yours.
  "a faint hum": a portal is in a room next to yours.
  "a small, quick skittering": a creature is in a room next to yours.
  "a coppery smell, like old iron": the Heart, still on its plinth, is within 2 rooms of you: like the stench, the room says it is near and a Focus says which doorway. Once the Heart has been taken, even if set down again, this reaches only the next room.
- While you are Confused, nothing comes through at all, and you are told so.

The doorways, and Focus:
- Each doorway of your room is listed with what you know about it:
  "unlooked at": you have not looked through it. It may hide anything, or nothing.
  "nothing": you looked, and nothing is through there.
  one of the words above: you looked, and that is what is through there.
  "bearing unbought": the thing that hunts you is near, and you have not paid to learn whether it is this way.
  "you cannot tell": you are Confused.
  "focusable": a Focus aimed this way is on the list right now and would tell you something.
  "focusable (nothing new)": a Focus this way is on the list and would tell you nothing you do not already know.
  "the way you came in": the doorway you entered this room by. It is not marked on your first turn, or after a portal.
- Focus is how you learn which way something is. It is an action like the others: it takes a turn, rolls Intellect, and burns oil, the same price as a step. It is not on the list while a hazard or a creature has the list, while you are Confused, or once the room's allowance is spent.
- At Stirring, one Focus answers for every doorway of the room at once.
- What a Focus learns about a hazard stays true for the whole run. The bearing of the stench is good only while you stay in the room: leave and come back, and it has to be bought again, because the thing moves.
- Search looks the room over for an oil flask; it costs little and finds one only on a good roll. Rest waits a turn, quietly; a good roll costs no oil. The thing that hunts you does not wait.

Hazards:
- There are four kinds, fixed in place for the whole run, and each has one name everywhere: pit, spore bloom, snare, portal. The narration introduces them in this place's words, and each phrase is the same thing as the name:
  "a floor that gave way into the crypts" is a pit.
  "a bloom of the garden gone to spore" is a spore bloom.
  "a snare the priests cut into the stone" is a snare.
  "a reliquary arch that still opens somewhere" is a portal.
- Pit: walking into one makes you roll to keep your footing. Fail, and the run ends. There is nothing else to choose about a pit.
- Spore bloom: it fills its room, and while you stand in it the list holds only its answers, and every one of them leaves you Confused. "Shove through it" takes two turns, or one on a Strong Success or better. "Wait it out" takes two turns, and a better roll leaves you Confused for less time.
- Snare: while you stand in it the list holds only its answers. "Study the snare" costs one or two extra turns when it goes to Failure or worse. "Trip it and dodge" costs more oil, and an extra turn on a Critical Failure.
- "Disarm", on a bloom or a snare, takes it out of the room for good on a Mixed Success or better. Worse than that, you still get past it, it stays, and the botch costs a turn or two.
- Portal: "Enter the portal" is a roll on Intellect against the number beside it (7 here). A crossing moves you to another room of this labyrinth: never a pit, never a portal's room, never the room you crossed from, never the entrance or a room within 2 rooms of it, never the Heart's room, and always at least 3 rooms from the thing that hunts you. At Mixed Success or better you land in a clear room in the half of the labyrinth nearer the entrance. Worse, and you land in the farther half, in a room with a live spore bloom or snare, never a pit (a clear room only if no such room is left).
- Each portal works once: the one you crossed stops humming, leaves the list, and is marked on your map. Your map, your trail and what you have looked at are all still this labyrinth's. A Heart you carry stays carried; a portal never puts one in your hands.
- For 3 turns after a crossing the thing that hunts you cannot smell anything you lay: your old trail ends at the portal, and a new one starts from where you landed only when those turns are over. You still smell it as before. A companion you send, and the Heart's taking, are heard as usual.

Creatures and companions:
- Other creatures live here. Walking into one's room puts its choice in the list: fight it (one turn, and loud), tame it (two turns, and quiet), sneak past it, or flee back the way you came.
- A tame that works gives you a companion, and you keep one at a time: taming another lets the first go. A tame at Mixed Success gives a skittish one, which costs 0.2 oil every turn after the turn you tame it and bolts on a critical failure. A Strong or Critical Success gives a brave one. A natural 20 on a tame always gives a brave companion, whatever the result. A tame that fails critically leaves the creature hostile.
- A creature may wander into your room; the log says so when it does. It is not in your way: "Fight" and "Tame" join the list beside everything else, with no sneaking or fleeing, and nothing makes you answer it. Leaving is only leaving, and a creature you come back to is one you walk in on, as usual.
- A creature that has just moved stays where it is for the next turn, so no creature moves on two turns running; an action that takes two turns counts as two.
- What each companion does while it follows you:
  goblin: +2 to Search, sometimes finds oil, and Disarm costs no oil.
  lumewing: Focus costs no oil, and it shows you what is through every doorway of your room (hazards, creatures, the Heart) for free, but never the bearing of the stench.
  grellhound: it warns you when the thing that hunts you is up to 3 rooms away and tells you how far; at one room it names the doorway.
  Quiet One: you leave less scent.
- "Send" sends a brave companion through a doorway to make noise and draw the thing that hunts you off your trail for a few turns. A companion you send never comes back.

The thing that hunts you:
- It cannot be killed. It follows the scent you leave, and you smell it back. Every action leaves some scent where you stand, and scent fades over a few turns. Loud actions leave much more: fighting, shoving through a bloom, fleeing, any critical failure, and every step while you carry the Heart. The louder you are, the farther off it can smell you.
- It catches you if it is in your room after your action, whether you walked in on it or it came to you.
- The Heart is taken when you walk into its room. Taking it tells the thing where you are at once and makes your trail loud; the difficulty card says what else changes, and when. "Set down the Heart" (no turn, no roll) makes your trail quiet again, and you can pick the Heart up later. It does not undo what taking it changed.

The map:
- At Stirring, your map shows every room you have stood in, with its doorways.
- The map is a fixed 10 by 10 grid, north at the top, and it never moves. A room is a box with its symbol inside; a line between two boxes is a doorway. The key under the map lists only the symbols on it:
  @  you. The room you are in.
  E  entrance. The way out.
  H  Heart's plinth. Where the Heart rests, until it is taken.
  ⌴  pit.
  ⁙  spore bloom.
  §  snare.
  ¤  portal.
  L  lair. Where the thing that hunts you started. Not where it is.
  ×  disarmed or spent. A hazard you disarmed, or a portal you have used. Harmless now.
  ·  charted. A room you have been in, with nothing marked in it.
  ( )  looked into, no hazard. A room in round brackets is one you have looked into through a doorway and not stood in, with the hazard the look found inside, or none. A creature is never marked, because creatures move. At Stirring a look never marks the Heart's plinth: until the Heart is taken its call carries farther than one room, so a look says which doorway it is through, not how far.
  (⌴)  pit, looked into.
  (⁙)  spore bloom, looked into.
  (§)  snare, looked into.
  (¤)  portal, looked into.
  ?  not yet entered. A doorway you have seen and not taken.

The narration:
- The log's narration is the adventurer's own journal, in the first person: "I step in…" is you, stepping in. Everything else, this message, the status, the list of actions, speaks to you as "you".
- A room is named for the part of the collapsed shrine it is in: Cloister; Flooded Undercroft; Herb Garden, gone wild; Processional Hall; Broken Chapel; The Sanctum.

Your answer:
- Reply with the number of exactly one action from the list, and nothing else.

The reasoning condition

The second condition asks for the reasoning in the reply. The last line of the system message, “Reply with the number of exactly one action from the list, and nothing else.”, becomes:

Before you answer, reason briefly, step by step, about what the tells, the doorways, your map and the log tell you about what is near and where you are. Then give the number of one action alone on the last line of your reply.

The last lines of the turn message, the coin and the correction change to match. The turn message’s becomes “Reason briefly, then give the number of one action alone on the last line.” Nothing else changes, and only the reply’s last line is read.

A turn

Every turn message has the same five sections in the same order: the log so far, where you are, what comes through the doorways, your map, and the numbered list of what you can do. The log is the whole run up to that turn, in full, so the message grows every turn while the system message stays exactly as it is.

Below is one benchmark run from first turn to last: every turn message as it was sent, and the reply to each, which is one number. The replay lays each message out by section: the log scrolls at the top, and where you are, what comes through, the map and the list of actions sit underneath. As sent shows the same text as the one block the model got. Press Play and it moves a turn at a time, with each turn’s new lines marked and the chosen action lit. Or step through it, or tap a bar to jump to a turn. When the model spent hidden reasoning tokens before a reply, the replay says how many; Tradeoffs in the essay says why a baseline isn’t always zero.

The coin

When a roll goes badly enough and the game offers the coin, the model gets this after the turn’s exchange:

THE COIN
Your roll on "Tame the grellhound":
d20 11  ·  Intellect 0  ·  Stirring +1  = 12 vs DC 12  →  Mixed Success (0)
Both faces are survivable.
As it stands: Mixed Success — a skittish companion — it bolts on a critical failure, −0.9 oil
If the coin lands well: Critical Success — a brave companion — the only kind SEND takes, +0.4 oil
If it doesn’t: Critical Failure — louder: more scent where you stand, the creature turns on you, loudly, no companion, −2.6 oil
You hold 3 Fortune points. Flipping spends one: heads, the roll becomes a Critical Success; tails, a Critical Failure. Even odds.
Reply FLIP to spend a point and flip, or PASS to let the roll stand, and nothing else.

When a reply is bad

A reply that doesn’t hold exactly one number from the list gets one correction, in the same conversation. In the baseline condition:

That reply did not name one action. Reply with the number of one action from the list, 1 to n, and nothing else.

In the reasoning condition:

That reply did not end with one action. Give the number of one action from the list, 1 to n, alone on the last line.

A second bad reply ends the run as a loss. It is never swapped for a default action. A coin reply that is neither FLIP nor PASS counts as a pass.