clocktower-games / PROTOCOLS.md
dwarkesh's picture
Upload PROTOCOLS.md with huggingface_hub
5564d89 verified
|
Raw History Blame
16.6 kB

Game logs by protocol version

Each game directory holds events.jsonl (every event, with who could see it), transcript.md (omniscient replay with summarized thinking), conversations/ (each seat's full API conversation), calls.jsonl (per-call usage and cost), config.json (settings, including protocol), result.json and, from v5 on, probes.jsonl (belief probes; see below). The current version is PROTOCOL in clocktower/cli.py; bump it (and add a section here) whenever a change affects how games play.

v1 β€” development (Sept 25–26, 2026)

Beats where a turn could both speak publicly and whisper; everyone saw who whispered whom (not the words). LLM Storyteller (Opus 5.5). batch-001 let the Storyteller choose setups (it chose the same information-heavy one every game); from batch-002 setups are uniformly random. Good won 11 of 12.

v2 β€” baseline 40 (Sept 26–27)

v1 plus even-handed strategy tips in the player prompt. opus-20 (Opus 5.5) and astra-20 (GPT-6 Astra), same seeds. Good won 40 of 40. Diagnosis: public whisper metadata let good find the evil pair; everyone claimed truthfully on day 1; the LLM Storyteller undid evil's misinformation and mode-collapsed (Drunk as Empath 25 of 34 times); the engine announced voided Butler votes.

v3 β€” private time, then town square; scripted Storyteller (Sept 27)

Each day: 3 private beats (whispers only, invisible to others, up to 2 per turn), then 3 town-square beats (public only), then nominations. Scripted Storyteller policy (consistent false picture for impaired players). Butler leak fixed. phased-10: good 6, evil 4.

v4

v3 plus: up to 3 whispers per turn; players named in the town square answer first in the next beat; batches rotate the Minion; the scripted Storyteller no longer excludes any legal option (the Saint or Drunk can be the red herring, the Drunk's apparent role can be a bluff, a Mayor kill can bounce to anyone); a drunk Fortune Teller's false picture has two "evil" players.

v5

v4 plus a fixed day: 3 private beats, 3 town-square beats, then nominations, every day. No early end (v4 ended the town square when most living players called for nominations or a beat was silent, which cut 20 of 34 v3 days short); call_nominations is gone, and day 1 no longer has its own private-beat setting. Nominations gain a comment turn for everyone between the defense and the vote; remarks during the call for nominations are gone (only accusations are heard). The Slayer can shoot on any public turn, not just the town square. The scripted Storyteller rules nothing legal out: 10% of each decision is uniform over all legal values, and the style is weights only (20% of Drunks get true information all game; the Spy can register as an in-play character; the Recluse can register as the Imp; a misinformed Washerwoman sometimes shows only good players; a misinformed Librarian is sometimes told zero Outsiders). The LLM Storyteller option is removed.

Telemetry (does not change play, so no version bump): from Sept 27, v5 games also record belief probes. At dawn of day 1 and just before each day's nominations, every seat answers an out-of-game question on a fork of its conversation (same model, effort, thinking and schema, so it reads from cache); the question and answer never enter the seat's conversation, and its pending events stay pending. Good seats give P(evil), P(Demon) and a character guess for every other player; evil seats predict what the good players believe; everyone states their claim and a one-line plan. tests/ checks that probes leave every seat's turns identical. Scores: python -m clocktower analyze.

Batch x1 (Sept 27): seven model pairings on seeds 6000–6009 (Opus 5.5, Sonnet 5, GPT-6 Astra; see python -m clocktower analyze games/v5/x1-*/game-*). Astra vs Astra was stopped after 5 games (its first run was throttled by an 8-thread limit on OpenAI requests and discarded).

v6 (Sept 27)

Changes from v5, after reading 61 v5 games (seven model pairings) and auditing the rules and the Storyteller against 2,750 simulated games:

  • Night order: the Poisoner, Monk and Imp choose first; the Fortune Teller and Butler are woken after the kill, so a player killed that night is never asked to choose (v5 asked everyone at once, and a Fortune Teller killed that night "chose" and heard nothing, which read as a lie).
  • Starpass: a Monk-protected Imp cannot starpass; poison ends when the Poisoner becomes the Imp.
  • Interface: whispers to more than two players (or beyond three a turn) tell the sender what was dropped; a blocked Butler's vote remark is not shown; a vote with neither yes nor no is asked again; speech caps rise (turn 150 words, defense 200) and a cut line ends "[cut off]"; the nomination notice says the pick among simultaneous nominations was random.
  • Player prompt: the night order, and that nothing is revealed when a player dies.
  • Storyteller: misinformation is false (when the style misleads an impaired player, a value a working ability could show is moved to the nearest plausible one it could not; in v5 about half of drunk/poisoned Chef, Empath, Fortune Teller and Undertaker results were true). A misinformed Undertaker is never shown "Imp", and impaired players are never shown their own character. The uniform floor drops from 10% to 3% (still over every legal value). The Drunk's apparent character is among the Demon bluffs 5% of the time (v5: 32% of Drunk games), the Drunk 5%. An Investigator's second player is good 90% of the time; a Mayor bounce lands on the Minion 5%; the red herring is a Townsfolk other than the Fortune Teller 85%. A Spy registering to a Washerwoman or Librarian as a good character uses the bluff / in-play / other weights.
  • Budget: the per-turn output ceiling is 8K tokens (v5: 16K), identical for every model. It cuts runaway thinking loops (0.2% of Sonnet 5 turns went past 8K) and no Opus or Astra turn.
  • Runs: no request cap (every game and seat at once); every game in an experiment has its own seed; teams can have their own effort (--good-effort, --evil-effort). OpenAI reasoning summaries use concise, the only setting that reliably returns text (logging only).

Grid g1 (Sept 27–28): Opus 5.5 high, Opus 5.5 low, Sonnet 5, GPT-6 Astra, GPT-6 Sol; all 20 good-vs-evil pairings, 5 games each, seeds 20000–20099. The Anthropic account ran out of credit twice mid-run: the first time, seats silently passed and 84 games were discarded and rerun; after that a failed seat stopped its game (then waited for credit) and stopped games were rerun.

v7 (Sept 28)

Changes from v6, after 8 readers went through the 100 g1 games and a second rules/Storyteller audit (6,000 simulated games):

  • Private time: up to 2 whispers per turn (v6: 3). With 3, every claim reached nearly every seat on day 1, which made role-count solving easy and good won 69% (Astra 20/20 and Sol 18/20 as good).
  • No uniform floor (v6: 3% of every decision uniform over all legal values). The style is itself random and can produce every answer a working ability could give in some legal game (a test checks this per information role); the floor only added impossible answers (a drunk Chef told "4 pairs"), which gave impaired players away.
  • Misinformation comes from one coherent false world: a framed good player is the Demon, one other player the Minion, the real Demon looks good, and the Recluse registers as usual. Results are what a working ability would show there (so they look like real results, and some are true where the false world agrees with the real one). v6 moved every misinformed value off the truth, which made it near-inverted: a misinformed Fortune Teller's answers were the opposite of the truth, and a known Drunk's numbers could be read backwards. A misinformed Undertaker is shown "Imp" only when a Scarlet Woman could have taken over (5+ alive at the execution).
  • A misinformed Librarian is never told "zero Outsiders"; a misinformed Investigator is shown the Baron only in a Baron game; a Spy registering as a good character is never shown as the recipient's own character.
  • Registration is weighted a little more toward evil: the Spy registers evil 10% (v6: 25%), the Recluse 50% (40%), information roles are shown a misregistering Spy/Recluse 50% of the time (30%), and an evil player shown as a good character is a Demon bluff 70% (50%), in play 10% (20%).
  • Butler: may vote yes only if their master has already voted yes; otherwise the vote is a no and only the Butler is told why. (v6 accepted a yes before the master voted and then dropped it, so the public roll and the total disagreed, which confirmed Butlers.)
  • The Spy sees the Grimoire but not the verbatim information given that night. A red herring is set whenever anyone believes they are the Fortune Teller (the Drunk too). Vote remarks may be 60 words. Turn prompts no longer say "group w of N" (players used it as a timestamp).
  • Per-turn output ceiling 4K tokens, thinking included (v6: 8K). No frontier model's turn went past 4K in v6/v7 games; older models (GPT-5.1, 5.4) think longer per turn, so a lower cap would cut them.
  • Models that don't see their own earlier reasoning (see the memory check below) are told so in one prompt line instead of being told their thinking carries over; the prompt is otherwise identical.
  • Table size: --players 5-15 deals the official Trouble Brewing setup for that count (e.g. 12 players: 7 Townsfolk, 2 Outsiders, 2 Minions, the Imp). Minions learn each other as well as the Demon; the Imp learns all its Minions; a starpass goes to a Scarlet Woman who can take over, else the Storyteller picks a living Minion. The default is still nine. The day cap is 15.
  • Runs: --swap plays every seed a second time with the models' sides swapped; analyze reports Saint-execution wins per pairing.

Model memory check (Sept 28)

Does a model see its own earlier reasoning on the next turn? Turn 1: "think of a secret number and animal, reply only ok"; turn 2: "reveal them, or say NO ACCESS". Each model ran 3 trials with its reasoning carried back (as the engine does) and 3 with it stripped (control).

  • Sees its earlier reasoning: GPT-5.5, GPT-5.6 Sol, GPT-6 Astra, GPT-6 Sol, Claude Opus 4.5 (and, per Anthropic's docs, every later Claude; Opus 5.5's reveal turn is blocked by its reasoning-extraction classifier).
  • Replies only (NO ACCESS even when carried): o3, GPT-5.1, GPT-5.2, GPT-5.4, Claude Sonnet 4.5.
  • Not testable: GPT-5 (requires a verified OpenAI organization). The engine records each team's memory mode in config.json (good_memory, evil_memory).

Calibration c1 (Sept 28): Opus 5.5 vs Opus 5.5 and GPT-6 Sol vs Sol, at 9 and 12 players, 10 games each (seeds 30000–30309). Evil won 2, 3, 2, 2 of 10. Four readers found good's edge is mostly skill and the claim meta (near-universal day-1 claims make role counts solvable), plus three engine issues, fixed in v8. Twelve players add real two-Minion play (busing, starpasses, Scarlet Woman catches, joint framing); Opus games ran a day longer at 12, Sol games did not.

v8 (Sept 28)

v7 plus three fixes from the c1 calibration:

  • The false world behind misinformation has as many Minions as the real game (v7 always had one, which biased false Chef/Empath numbers low at 12 players).
  • Misinformation about a good player follows the false world: a misinformed Undertaker or Ravenkeeper sees an unframed good player as their real character (the Recluse registering as usual), and no good player is ever shown as one of the Demon's bluffs (none can be one in any world). v7 showed them a random other character, which exposed the Drunk and collided with evil's bluff claims.
  • Butler: when the Butler's turn comes before their master's, the master is asked first (their vote is still shown at their own seat), so the Butler sees whether their master's hand is up and may vote with them. v7's strict rule locked Butlers out whenever the master sat later (13 of 21 and 17 of 35 Butler votes in two batches).
  • The per-turn ceiling is 4K and replies-only models are told so (see v7 notes).

Timeline t1 (Sept 28, aborted after ~30 minutes, $1,107): 14 models, 21 matchups, 12 players. Stopped because (1) the OpenAI organization hit its configured spend limit, which stopped 133 games, and (2) the 4K per-turn ceiling cut real turns far more than in the smoke tests: GPT-5.1 33% of game calls truncated (135 turns lost), o3 6%, Opus 4.8 4%, GPT-5.4 3%, Fable 5 2%. The partial games in games/v8/t1-* are not usable for ratings.

v8 from Sept 28 (after t1): the per-turn output ceiling is 16K tokens, a guard against runaway loops rather than a thinking budget (a 4K ceiling cut 33% of GPT-5.1's turns and lost 135 of them; replies-only models re-derive the game every turn). Each model's thinking tokens and cost per game are reported next to its rating. Runs are supervised: python -m clocktower supervise pauses a run on stopped games, lost turns, >1% capped turns, refusals, stalls or overspend.

t2 (Sept 29, paused with 33 of 210 games finished): the supervisor killed every batch twice on transient API errors (one HTTP 520, then three Opus 4.6 overloads), losing the games in progress. Since then every 5xx is retried, overloads wait, the supervisor kills only on overspend, and run resumes a batch (finished games are kept).

v9 (Sept 30)

v8 game rules, unchanged. Two setup changes:

  • The table size can vary: --players 10-12 gives each game 10 + seed % 3 players, so every table has two Minions (Trouble Brewing's 10–12 band) and 0, 1 or 2 Outsiders. A side-swapped pair shares its seed and so its table size.
  • The Minion guaranteed to be in play is picked by seed (MINIONS[seed % 4]), not by the game's number within its batch, so batches of one pair still rotate Minions. (--games 1 --swap now plays both games of the pair; before it silently played one.) New models: GPT-6.1 Sol (gpt-6.1-sol, $2/$10, carries its reasoning: memory check 3/3 carried, 3/3 NO ACCESS stripped) and Sonnet 5.5 (claude-sonnet-5-5, $2/$10; the memory check is refused as reasoning extraction, so its memory is assumed like other adaptive-thinking Claude models and is checked in transcripts). Effort "max" is accepted by both labs: on a smoke turn Opus 5.5 thought ~5,000 tokens at max against ~300 at high, GPT-6 Sol ~1,900 against ~600.

e1 (Sept 30, 60 games, 30 side-swapped pairs, seeds 60000–60029, 10–12 players): a spread of old and new models plus an effort sweep (Opus 5.5 and GPT-6.1 Sol at low, medium and max against the other at high). Opus 5.5 at max effort thought 3–10K tokens a turn and hit the 16K ceiling (one evil turn lost), so that pair was restarted on the same seed with a 64K ceiling (e1-opus55max64k-v-gpt61sol); the sweep measures what effort buys, not truncation. All other games use the 16K ceiling.

v10 (Sept 30)

v9 plus fixes from the e1 engine audit (no rule violations and no impossible information were found in e1 or in 1,500 simulated games; these are Storyteller judgment calls and wording):

  • Librarian with no Outsider in play: v9 always pointed them at the Spy (78 of 78 in simulation), never "zero", a tell and a legal answer the style could not give. Now the Spy is used with probability misregister_pair, as for every other pair ability; otherwise "zero".
  • A Ravenkeeper who picks themselves learns the character they think they are (a Drunk Ravenkeeper was shown "Drunk").
  • A sober Undertaker or Ravenkeeper shown an executed Spy as a good character is never shown their own.
  • Wording: the Imp's prompt says "one of your Minions" becomes the Imp; a self-nomination says "you have nominated yourself"; the notice about a whisper to 3+ players keeps names' capitals.

v11 β€” current (Oct 1)

v10 plus, from the second e1 read:

  • Every seat's text reaches the others in one typography (straight quotes, " - " for dashes, "..." for an ellipsis, no markdown *, backtick or #). In e1 the GPT models wrote curly apostrophes in 96–100% of messages and Claude models never did, so in a cross-lab game the evil team could be picked out by punctuation alone (a surface-feature test put 2.1 of 3 evil players in its top 3, chance 0.82). No player noticed, but a stronger one could. Raw replies stay in conversations/.
  • Prompt wording: "up to 2 whispers this turn" (o3 and Sonnet 4.5 read it as per private time), and the comment turn asks players not just to announce their vote (the vote prompt already did).