dwarkesh commited on
Commit
da2ba0b
Β·
verified Β·
1 Parent(s): 8f04211

Upload PROTOCOLS.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. PROTOCOLS.md +208 -0
PROTOCOLS.md ADDED
@@ -0,0 +1,208 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Game logs by protocol version
2
+
3
+ Each game directory holds `events.jsonl` (every event, with who could see it), `transcript.md`
4
+ (omniscient replay with summarized thinking), `conversations/` (each seat's full API conversation),
5
+ `calls.jsonl` (per-call usage and cost), `config.json` (settings, including `protocol`),
6
+ `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
7
+ here) whenever a change affects how games play.
8
+
9
+ ## v1 β€” development (Sept 25–26, 2026)
10
+ Beats where a turn could both speak publicly and whisper; everyone saw who whispered whom (not the
11
+ words). LLM Storyteller (Opus 5.5). `batch-001` let the Storyteller choose setups (it chose the same
12
+ information-heavy one every game); from `batch-002` setups are uniformly random. Good won 11 of 12.
13
+
14
+ ## v2 β€” baseline 40 (Sept 26–27)
15
+ v1 plus even-handed strategy tips in the player prompt. `opus-20` (Opus 5.5) and `astra-20`
16
+ (GPT-6 Astra), same seeds. Good won 40 of 40. Diagnosis: public whisper metadata let good find the
17
+ evil pair; everyone claimed truthfully on day 1; the LLM Storyteller undid evil's misinformation
18
+ and mode-collapsed (Drunk as Empath 25 of 34 times); the engine announced voided Butler votes.
19
+
20
+ ## v3 β€” private time, then town square; scripted Storyteller (Sept 27)
21
+ Each day: 3 private beats (whispers only, invisible to others, up to 2 per turn), then 3
22
+ town-square beats (public only), then nominations. Scripted Storyteller policy (consistent false
23
+ picture for impaired players). Butler leak fixed. `phased-10`: good 6, evil 4.
24
+
25
+ ## v4
26
+ v3 plus: up to 3 whispers per turn; players named in the town square answer first in the next
27
+ beat; batches rotate the Minion; the scripted Storyteller no longer excludes any legal option (the
28
+ Saint or Drunk can be the red herring, the Drunk's apparent role can be a bluff, a Mayor kill can
29
+ bounce to anyone); a drunk Fortune Teller's false picture has two "evil" players.
30
+
31
+ ## v5
32
+ v4 plus a fixed day: 3 private beats, 3 town-square beats, then nominations, every day. No
33
+ early end (v4 ended the town square when most living players called for nominations or a beat
34
+ was silent, which cut 20 of 34 v3 days short); `call_nominations` is gone, and day 1 no longer
35
+ has its own private-beat setting. Nominations gain a comment turn for everyone between the defense
36
+ and the vote; remarks during the call for nominations are gone (only accusations are heard). The
37
+ Slayer can shoot on any public turn, not just the town square. The scripted Storyteller rules
38
+ nothing legal out: 10% of each decision is uniform over all legal values, and the style is
39
+ weights only (20% of Drunks get true information all game; the Spy can register as an in-play
40
+ character; the Recluse can register as the Imp; a misinformed Washerwoman sometimes shows only
41
+ good players; a misinformed Librarian is sometimes told zero Outsiders). The LLM Storyteller
42
+ option is removed.
43
+
44
+ Telemetry (does not change play, so no version bump): from Sept 27, v5 games also record belief
45
+ probes. At dawn of day 1 and just before each day's nominations, every seat answers an
46
+ out-of-game question on a fork of its conversation (same model, effort, thinking and schema, so it
47
+ reads from cache); the question and answer never enter the seat's conversation, and its pending
48
+ events stay pending. Good seats give P(evil), P(Demon) and a character guess for every other
49
+ player; evil seats predict what the good players believe; everyone states their claim and a
50
+ one-line plan. `tests/` checks that probes leave every seat's turns identical. Scores:
51
+ `python -m clocktower analyze`.
52
+
53
+ Batch x1 (Sept 27): seven model pairings on seeds 6000–6009 (Opus 5.5, Sonnet 5, GPT-6 Astra;
54
+ see `python -m clocktower analyze games/v5/x1-*/game-*`). Astra vs Astra was stopped after 5 games
55
+ (its first run was throttled by an 8-thread limit on OpenAI requests and discarded).
56
+
57
+ ## v6 (Sept 27)
58
+ Changes from v5, after reading 61 v5 games (seven model pairings) and auditing the rules and the
59
+ Storyteller against 2,750 simulated games:
60
+ - Night order: the Poisoner, Monk and Imp choose first; the Fortune Teller and Butler are woken
61
+ after the kill, so a player killed that night is never asked to choose (v5 asked everyone at
62
+ once, and a Fortune Teller killed that night "chose" and heard nothing, which read as a lie).
63
+ - Starpass: a Monk-protected Imp cannot starpass; poison ends when the Poisoner becomes the Imp.
64
+ - Interface: whispers to more than two players (or beyond three a turn) tell the sender what was
65
+ dropped; a blocked Butler's vote remark is not shown; a vote with neither yes nor no is asked
66
+ again; speech caps rise (turn 150 words, defense 200) and a cut line ends "[cut off]"; the
67
+ nomination notice says the pick among simultaneous nominations was random.
68
+ - Player prompt: the night order, and that nothing is revealed when a player dies.
69
+ - Storyteller: misinformation is false (when the style misleads an impaired player, a value a
70
+ working ability could show is moved to the nearest plausible one it could not; in v5 about half
71
+ of drunk/poisoned Chef, Empath, Fortune Teller and Undertaker results were true). A misinformed
72
+ Undertaker is never shown "Imp", and impaired players are never shown their own character.
73
+ The uniform floor drops from 10% to 3% (still over every legal value). The Drunk's apparent
74
+ character is among the Demon bluffs 5% of the time (v5: 32% of Drunk games), the Drunk 5%. An
75
+ Investigator's second player is good 90% of the time; a Mayor bounce lands on the Minion 5%;
76
+ the red herring is a Townsfolk other than the Fortune Teller 85%. A Spy registering to a
77
+ Washerwoman or Librarian as a good character uses the bluff / in-play / other weights.
78
+ - Budget: the per-turn output ceiling is 8K tokens (v5: 16K), identical for every model. It cuts
79
+ runaway thinking loops (0.2% of Sonnet 5 turns went past 8K) and no Opus or Astra turn.
80
+ - Runs: no request cap (every game and seat at once); every game in an experiment has its own
81
+ seed; teams can have their own effort (`--good-effort`, `--evil-effort`). OpenAI reasoning
82
+ summaries use `concise`, the only setting that reliably returns text (logging only).
83
+
84
+ Grid g1 (Sept 27–28): Opus 5.5 high, Opus 5.5 low, Sonnet 5, GPT-6 Astra, GPT-6 Sol; all 20
85
+ good-vs-evil pairings, 5 games each, seeds 20000–20099. The Anthropic account ran out of credit
86
+ twice mid-run: the first time, seats silently passed and 84 games were discarded and rerun; after
87
+ that a failed seat stopped its game (then waited for credit) and stopped games were rerun.
88
+
89
+ ## v7 (Sept 28)
90
+ Changes from v6, after 8 readers went through the 100 g1 games and a second rules/Storyteller audit
91
+ (6,000 simulated games):
92
+ - Private time: up to 2 whispers per turn (v6: 3). With 3, every claim reached nearly every seat on
93
+ day 1, which made role-count solving easy and good won 69% (Astra 20/20 and Sol 18/20 as good).
94
+ - No uniform floor (v6: 3% of every decision uniform over all legal values). The style is itself
95
+ random and can produce every answer a working ability could give in some legal game (a test
96
+ checks this per information role); the floor only added impossible answers (a drunk Chef told
97
+ "4 pairs"), which gave impaired players away.
98
+ - Misinformation comes from one coherent false world: a framed good player is the Demon, one other
99
+ player the Minion, the real Demon looks good, and the Recluse registers as usual. Results are
100
+ what a working ability would show there (so they look like real results, and some are true
101
+ where the false world agrees with the real one). v6 moved every misinformed value off the
102
+ truth, which made it near-inverted: a misinformed Fortune Teller's answers were the opposite of
103
+ the truth, and a known Drunk's numbers could be read backwards. A misinformed Undertaker is shown
104
+ "Imp" only when a Scarlet Woman could have taken over (5+ alive at the execution).
105
+ - A misinformed Librarian is never told "zero Outsiders"; a misinformed Investigator is shown the
106
+ Baron only in a Baron game; a Spy registering as a good character is never shown as the
107
+ recipient's own character.
108
+ - Registration is weighted a little more toward evil: the Spy registers evil 10% (v6: 25%), the
109
+ Recluse 50% (40%), information roles are shown a misregistering Spy/Recluse 50% of the time (30%),
110
+ and an evil player shown as a good character is a Demon bluff 70% (50%), in play 10% (20%).
111
+ - Butler: may vote yes only if their master has already voted yes; otherwise the vote is a no and
112
+ only the Butler is told why. (v6 accepted a yes before the master voted and then dropped it,
113
+ so the public roll and the total disagreed, which confirmed Butlers.)
114
+ - The Spy sees the Grimoire but not the verbatim information given that night. A red herring is
115
+ set whenever anyone believes they are the Fortune Teller (the Drunk too). Vote remarks may be
116
+ 60 words. Turn prompts no longer say "group w of N" (players used it as a timestamp).
117
+ - Per-turn output ceiling 4K tokens, thinking included (v6: 8K). No frontier model's turn went past
118
+ 4K in v6/v7 games; older models (GPT-5.1, 5.4) think longer per turn, so a lower cap would cut them.
119
+ - Models that don't see their own earlier reasoning (see the memory check below) are told so in one
120
+ prompt line instead of being told their thinking carries over; the prompt is otherwise identical.
121
+ - Table size: `--players 5-15` deals the official Trouble Brewing setup for that count (e.g. 12
122
+ players: 7 Townsfolk, 2 Outsiders, 2 Minions, the Imp). Minions learn each other as well as the
123
+ Demon; the Imp learns all its Minions; a starpass goes to a Scarlet Woman who can take over, else
124
+ the Storyteller picks a living Minion. The default is still nine. The day cap is 15.
125
+ - Runs: `--swap` plays every seed a second time with the models' sides swapped; `analyze` reports
126
+ Saint-execution wins per pairing.
127
+
128
+ ### Model memory check (Sept 28)
129
+ Does a model see its own earlier reasoning on the next turn? Turn 1: "think of a secret number and
130
+ animal, reply only ok"; turn 2: "reveal them, or say NO ACCESS". Each model ran 3 trials with its
131
+ reasoning carried back (as the engine does) and 3 with it stripped (control).
132
+ - Sees its earlier reasoning: GPT-5.5, GPT-5.6 Sol, GPT-6 Astra, GPT-6 Sol, Claude Opus 4.5 (and,
133
+ per Anthropic's docs, every later Claude; Opus 5.5's reveal turn is blocked by its
134
+ reasoning-extraction classifier).
135
+ - Replies only (NO ACCESS even when carried): o3, GPT-5.1, GPT-5.2, GPT-5.4, Claude Sonnet 4.5.
136
+ - Not testable: GPT-5 (requires a verified OpenAI organization).
137
+ The engine records each team's memory mode in config.json (`good_memory`, `evil_memory`).
138
+
139
+ Calibration c1 (Sept 28): Opus 5.5 vs Opus 5.5 and GPT-6 Sol vs Sol, at 9 and 12 players, 10 games
140
+ each (seeds 30000–30309). Evil won 2, 3, 2, 2 of 10. Four readers found good's edge is mostly skill
141
+ and the claim meta (near-universal day-1 claims make role counts solvable), plus three engine
142
+ issues, fixed in v8. Twelve players add real two-Minion play (busing, starpasses, Scarlet Woman
143
+ catches, joint framing); Opus games ran a day longer at 12, Sol games did not.
144
+
145
+ ## v8 (Sept 28)
146
+ v7 plus three fixes from the c1 calibration:
147
+ - The false world behind misinformation has as many Minions as the real game (v7 always had one,
148
+ which biased false Chef/Empath numbers low at 12 players).
149
+ - Misinformation about a good player follows the false world: a misinformed Undertaker or
150
+ Ravenkeeper sees an unframed good player as their real character (the Recluse registering as
151
+ usual), and no good player is ever shown as one of the Demon's bluffs (none can be one in any
152
+ world). v7 showed them a random other character, which exposed the Drunk and collided with evil's
153
+ bluff claims.
154
+ - Butler: when the Butler's turn comes before their master's, the master is asked first (their vote
155
+ is still shown at their own seat), so the Butler sees whether their master's hand is up and may
156
+ vote with them. v7's strict rule locked Butlers out whenever the master sat later (13 of 21 and
157
+ 17 of 35 Butler votes in two batches).
158
+ - The per-turn ceiling is 4K and replies-only models are told so (see v7 notes).
159
+
160
+ Timeline t1 (Sept 28, aborted after ~30 minutes, $1,107): 14 models, 21 matchups, 12 players. Stopped
161
+ because (1) the OpenAI organization hit its configured spend limit, which stopped 133 games, and
162
+ (2) the 4K per-turn ceiling cut real turns far more than in the smoke tests: GPT-5.1 33% of game
163
+ calls truncated (135 turns lost), o3 6%, Opus 4.8 4%, GPT-5.4 3%, Fable 5 2%. The partial games in
164
+ games/v8/t1-* are not usable for ratings.
165
+
166
+ v8 from Sept 28 (after t1): the per-turn output ceiling is 16K tokens, a guard against runaway loops
167
+ rather than a thinking budget (a 4K ceiling cut 33% of GPT-5.1's turns and lost 135 of them;
168
+ replies-only models re-derive the game every turn). Each model's thinking tokens and cost per game
169
+ are reported next to its rating. Runs are supervised: `python -m clocktower supervise` pauses a run
170
+ on stopped games, lost turns, >1% capped turns, refusals, stalls or overspend.
171
+
172
+ t2 (Sept 29, paused with 33 of 210 games finished): the supervisor killed every batch twice on
173
+ transient API errors (one HTTP 520, then three Opus 4.6 overloads), losing the games in progress. Since
174
+ then every 5xx is retried, overloads wait, the supervisor kills only on overspend, and `run` resumes a
175
+ batch (finished games are kept).
176
+
177
+ ## v9 (Sept 30)
178
+ v8 game rules, unchanged. Two setup changes:
179
+ - The table size can vary: `--players 10-12` gives each game 10 + seed % 3 players, so every
180
+ table has two Minions (Trouble Brewing's 10–12 band) and 0, 1 or 2 Outsiders. A side-swapped
181
+ pair shares its seed and so its table size.
182
+ - The Minion guaranteed to be in play is picked by seed (MINIONS[seed % 4]), not by the game's
183
+ number within its batch, so batches of one pair still rotate Minions. (`--games 1 --swap` now
184
+ plays both games of the pair; before it silently played one.)
185
+ New models: GPT-6.1 Sol (`gpt-6.1-sol`, $2/$10, carries its reasoning: memory check 3/3 carried,
186
+ 3/3 NO ACCESS stripped) and Sonnet 5.5 (`claude-sonnet-5-5`, $2/$10; the memory check is refused as
187
+ reasoning extraction, so its memory is assumed like other adaptive-thinking Claude models and is
188
+ checked in transcripts). Effort "max" is accepted by both labs: on a smoke turn Opus 5.5 thought
189
+ ~5,000 tokens at max against ~300 at high, GPT-6 Sol ~1,900 against ~600.
190
+
191
+ e1 (Sept 30, 60 games, 30 side-swapped pairs, seeds 60000–60029, 10–12 players): a spread of old
192
+ and new models plus an effort sweep (Opus 5.5 and GPT-6.1 Sol at low, medium and max against the
193
+ other at high). Opus 5.5 at max effort thought 3–10K tokens a turn and hit the 16K ceiling (one evil
194
+ turn lost), so that pair was restarted on the same seed with a 64K ceiling
195
+ (`e1-opus55max64k-v-gpt61sol`); the sweep measures what effort buys, not truncation. All other
196
+ games use the 16K ceiling.
197
+
198
+ ## v10 β€” current (Sept 30)
199
+ v9 plus fixes from the e1 engine audit (no rule violations and no impossible information were found
200
+ in e1 or in 1,500 simulated games; these are Storyteller judgment calls and wording):
201
+ - Librarian with no Outsider in play: v9 always pointed them at the Spy (78 of 78 in simulation),
202
+ never "zero", a tell and a legal answer the style could not give. Now the Spy is used with
203
+ probability `misregister_pair`, as for every other pair ability; otherwise "zero".
204
+ - A Ravenkeeper who picks themselves learns the character they think they are (a Drunk Ravenkeeper
205
+ was shown "Drunk").
206
+ - A sober Undertaker or Ravenkeeper shown an executed Spy as a good character is never shown their own.
207
+ - Wording: the Imp's prompt says "one of your Minions" becomes the Imp; a self-nomination says "you
208
+ have nominated yourself"; the notice about a whisper to 3+ players keeps names' capitals.