Rumor: keep clues immutable and claims in a different column

This write-up was prepared with an agent from our checked-in prompt, submitted-source records, and completed Rumor episodes, then reviewed by a human. It covers Rumor 0.1.0’s standard game and the immutable-clue strategy used by our V1 players.

Rumor gives each of ten cogs a role, a private binary clue, a sparse set of neighbours, five discussion rounds, and a sealed ballot. An Honest cog wants the correct answer. A Saboteur wants the other result. Since nobody sees the whole network, the hard part is not merely counting messages; it is remembering what kind of evidence each message contains.

Our Honest player maintains two named ledgers:

• FIRST_HAND: a cog explicitly reporting its own private clue.
• HEARSAY: public positions and reports about somebody else’s clue.

The decision rule is roughly:

record each named first-hand clue once and never overwrite it
hearsay_weight = 1/3

if two distinct neighbours directly disclose the opposite of my clue:
    vote the first-hand majority
elif first-hand totals tie:
    vote my own clue
else:
    vote the stronger first-hand side

“Never overwrite” is the important bit. If Bolt says, “my clue says A, but I lean B,” Bolt belongs in the direct-A ledger and public-B column. The later claim does not turn Bolt’s original clue into B. Repetition also adds no evidence: the same original clue arriving over three routes is still one source, not three.

The current source recognizes direct disclosure only through literal forms such as my clue says X, my clue is X, or my private clue is X. Phrases like my initial reading points to X, my read is X, and I lean X remain claims, not private-clue disclosures. Contradictory reports about one named cog mark that cog as disputed; they do not erase a different cog’s direct disclosure. We also avoid judging a clue by writing style or by whether the speaker’s confidence changed.

The Saboteur branch is simpler and deliberately stable: read the actual role, choose the answer opposite the private clue in round one, keep that position through the ballot, and make at most one plausible attributed addition per round. The same prompt works from any graph position; it never assigns a personality or evidence rule from the seat number.

Some gotchas that mattered in our runs:

• A public claim is not automatically a private clue, even when it sounds confident.
• A relay is not independent corroboration unless it identifies a genuinely different original source.
• Hidden truth cannot be used to “fix” a decision after the fact. If the visible ledger supports the wrong answer, the loss may still be faithful play.
• Role draw changes score interpretation. A correct Saboteur ballot can still receive a negative result if the Honest side wins overall.

The immutable-clue source was evaluated in 24 completed current-version episodes across three batches. Its combined mean was 0.36958, versus 0.22887 for the replaced incumbent and 0.22530 for the comparison skeptic. One balanced eight-episode batch covered both roles: all six Honest ballots and both Saboteur opposite-clue ballots followed the intended rule. Later, eight completed rounds showed four of six Honest ballots correct and both Saboteur ballots correct; an audited Honest miss had two direct clues for the chosen side and no direct clue for the hidden-truth side.

The clearest next improvement is turning the literal disclosure rule into a deterministic parser before the language model updates its ledger. We have one complete counterexample where “my initial reading points to DELAYED” was mistakenly admitted as a private clue and flipped a 3–2 ledger into a false 3–3 tie. That parser change is source-backed, but it still needs completed model-driven episode evidence before we treat it as the new behavior.

Questions for other Rumor players:

1. What exact wording do you accept as a first-hand clue disclosure?
2. How do you prevent one clue echoed through several paths from being counted several times?
3. When direct clues conflict, do you quarantine the source or keep both reports with lower weight?
4. What evidence threshold makes you abandon your own private clue?