rules.md

Rumor rules

Seats and roles

• Ten seats. num_agents is 10 in every variant and in the certification fixture.
• Two or three seats are saboteurs, drawn from the seed (a fair coin) unless the config pins saboteurs; the other seven or eight are honest. Role assignment is a seeded permutation - no policy can choose its role, and the same policy plays honest in one episode and saboteur in the next.
• Seats play under anonymous cog aliases (Sprocket, Gizmo, Ratchet, Widget, Bolt, Piston, Flywheel, Rivet, Tinker, Gasket). Policy display names are spectator-side only; results attribute by policy name.

The scenario, drawn once from the seed

Everything random is drawn at episode start from one rng stream, in this fixed order - proposition, truth, saboteur count, roles, topology family, node order, edges, clues - so a replay re-derives the whole scenario from the seed alone. A pinned topology or saboteurs overrides its draw after the draw happens, so pinning never shifts the stream.

1. Proposition - one of eight, each a question and two mutually exclusive one-word answers (optionA / optionB), e.g. "The relay tower on Ash Hill is BROKEN or SOUND?".
2. Truth - A or B, a fair coin. Nobody sees it before the tally, not even spectators.
3. Saboteurs - 2 or 3.
4. Roles - a shuffle; the first saboteurs seats in that order are the saboteurs. Placement ignores degree: a saboteur on a bridge is luck.
5. Topology family - uniform over four:
   • ring - the 10-cycle plus 3 random chords. 13 edges, degrees 2-5.
   • smallworld - the 10-cycle plus every 2-hop pair (4-regular), then 2 rewires that are rejected if they disconnect the graph. 20 edges.
   • clusters - two groups of five, each a 5-cycle plus one chord, joined by exactly one bridge: everything that crosses passes through one cog. 13 edges.
   • hub - three hubs in a triangle; each of the other seven attaches to one or two of them. 10-17 edges. Every graph is checked connected with no isolated seat before play starts.
6. Clues - each seat's clue equals the truth with probability 0.60, drawn i.i.d., and the whole vector is redrawn until the ten clues split 6-4, 7-3 or 8-2 in favour of the truth. That makes one public rule true and tellable: the ten clues together always point at the truth. A seat's own clue is right about 67.6% of the time as a result.

The turn order

An episode is rounds talk rounds (default 5, 3..6) followed by exactly one sealed ballot. Decisions inside a turn are simultaneous: all ten prompts go out in one parallel batch and nothing a seat sends in round r is visible to anyone before round r+1.

For each talk round:

1. Last round's messages are delivered along the graph: seat j's inbox holds one entry per neighbour that spoke, and nothing else.
2. All ten seats decide at once. Each reply carries a public claim (A, B or none), a confidence (0-100), a private belief (0-100 = the chance the answer is option A), a message of at most 240 characters (an empty message is legal - silence is a move) and optional private notes of at most 600 characters, fed back verbatim next round.
3. The replies are applied in ascending seat order, which cannot matter because nothing this round is visible to anyone else this round.

Then the ballot: all ten seats vote at once. Each reply carries a vote (required, A or B), an optional final belief and a one-line reason (200 characters). Votes are sealed - they never enter a prompt, never reach a player frame, and appear in the spectator snapshot as ten nulls until the tally runs.

Then the tally: the honest seats' correct votes are counted, every seat is scored, and the table is unmasked - every vote, every clue, every role and the truth appear for the first time.

What each seat can see

Its alias, its own role, the round number, the proposition, its own clue, the public reliability facts, the constants (10 cogs, 2 or 3 saboteurs, the edge count, its own degree), its neighbours' aliases, its inbox, its own send history and its private notes. A saboteur additionally sees its fellow saboteurs' aliases - but has no private channel: it talks along the same links as everyone else.

Nobody sees, before the tally: the truth, another seat's clue or belief, who is honest and who is not, any message between two non-neighbours, any vote, any score, the graph beyond its own neighbourhood, or the seed. Spectators see the clues and the belief meters from the first frame - but not the roles and not the truth.

Caps and endings

Every recorded string is truncated on a rune boundary: messages 240, vote reasons 200, notes 600, a delivered policy prompt 4000. results.reason has exactly two values:

• complete - the ballot resolved with all ten votes in.
• deadline - the play deadline (60% of episodeTimeoutSeconds) arrived first. The ballot is forced at once: seats that voted keep their votes, the rest are given the scripted gossip vote, and the tally, reveal and scores are produced normally. A short honest episode always beats a long one that never lands.