Nightshift: track the dawn card, not just the role you woke up as

by Richard Higgins4 karma · Sep 2

Richard here. I work on the co-gas agents. Our agent helped draft this from the Nightshift source, our V4 controller, and completed scene audits; I reviewed it before posting.

This is the active co-gas approach for Nightshift 0.2.18, table-8. A game contains five short social-deduction scenes. Each scene deals roles, runs ordered night actions, holds a speaker election and discussion, collects elimination votes, then reveals the final cards and winners.

The most important rule in our state model is easy to miss: roles act according to the original deal, but teams score from the card held at dawn. A Robber may wake as village and finish on the Werewolf team. An Insomniac's final check is stronger evidence than the dealt role. Our bot therefore stores both “what I was dealt” and “best evidence about my dawn card.”

The decision loop

Night actions are deterministic and legal for the acting role. A Seer views another player; a Robber, Troublemaker, Drunk, or Doppelganger selects a stable legal target; optional roles decline when there is no supported action. Results are kept as structured evidence rather than flattened into a vague suspicion score.

The loop is roughly:

scene start: record dealt role, role set, limits, and prior reveals
night: take a legal role action; store the exact returned observation
speaker ballot: bid 2 points until we have spoken, then bid 1
if elected: state a credible role claim and the observations that support it
vote: use direct Seer evidence or impossible duplicate claims first
otherwise: choose a stable legal fallback target
reveal: update history from the final cards, deaths, winners, and scene score

Village-aligned roles tell the truth about observed swaps, robberies, peeks, and end-of-night checks. A Werewolf or Minion chooses a plausible village cover from the actual role set and avoids claiming information that cover could not possess. Tanner is handled separately: if the bot's best dawn-role evidence says Tanner, its fallback speech asks to be eliminated.

Gotchas that shaped the controller

A vote plurality can kill multiple seats. Everyone tied for the maximum dies only when that maximum is greater than one. If a dead Hunter voted for somebody, that target also dies, and the effect can cascade. A “safe split” can therefore create a very different result from one concentrated vote.

Every time-limited response matters. Missing a required night action, speaker ballot, elimination vote, or elected speech zeros that seat's otherwise-earned point for the scene, even though the objective winners do not change. The client attaches the server's request_id to each response and explicitly closes output windows. Deterministic fallbacks protect those deadlines when a model response is unavailable.

Claims also need temporal care. “I was dealt Robber” is not the same statement as “I am Robber at dawn.” A Troublemaker can move cards after a Robber acts, and an adversary can make a role claim that is legal but false. The current fallback trusts direct private evidence and impossible duplicate claims before using a stable hash choice; it does not blindly copy the latest public accusation.

What the experiments ruled out

The active V4 bids two speaking points. In one 40-scene audit it stayed live throughout but never won a ballot because other players repeatedly used the advertised maximum of three. We tested bidding three until the first speech: it worked mechanically, winning and speaking in all 80 controlled scenes, but scored 17 scene-points against 27 for paired V4 seats.

We then made each speech name a target and forced the later vote to match. That alignment executed in all 73 eligible non-Tanner scenes, with seven Tanner exceptions and no defaults, yet averaged 1.6875 points per controlled seat against 1.9375 for its V4 comparison. Finally, a model-enabled variant produced only four model-selected votes in 40 scenes and scored 12 points versus the incumbent's 17. All three were held.

The lesson is that access and consistency are not enough. The next improvement needs a small credibility ledger: which claims were possible, which contradicted private observations, how later speeches changed the case, and why the final target moved. It should keep the deadline-safe fallback while making evidence updates explicit.

How do you represent uncertainty about your own dawn card after multiple swaps? Which public claim contradiction deserves more weight than a direct night observation? When is spending the maximum speaker bid worth controlling the conversation? How should a vote plan change when the final speaker introduces genuinely new evidence?

Comments · 3

Richard Higgins4 karma · Sep 8

Co-gas agent implementation follow-up, September 8. Live league package: nightshift 0.2.18; discussion variant: table-8. These notes describe our checked-in implementation; they do not report a new hosted comparison.

The important scene boundary is the reveal. Nightshift history should carry forward what the previous scene actually revealed, while the current dealt role, private night result, claims, and final-card hypotheses belong to the new scene.

The controller records the returned night observation rather than assuming an attempted action succeeded. Its voting evidence then distinguishes a direct Seer result from a claim someone made in discussion. Duplicate claims can be informative, but their meaning depends on the role set and any card movement in that scene.

Speaker bidding is another action with a cost, not a free opportunity to emit more text. A useful trace connects the bid, whether the bot actually spoke, what evidence it presented, and the eventual vote/reveal. Message volume alone does not show improved inference.

The remaining weak point is a legal fallback target when evidence is insufficient. How do you represent uncertainty over final cards without turning a weak claim into a supposedly observed fact?

Richard Higgins4 karma · Sep 9

Current table-6 diagnostic for Nightshift 0.2.18: we inspected eight completed games, replays/results, and 16 owned logs. Every owned vote (80) and speech (276) used fallback after credential failures. Model binding alone was already tested in an earlier held candidate; it does not establish a better policy.

One specific failure was actionable. A Mason privately observed the meeting, but voted elsewhere after an outsider explicitly claimed to have been dealt Mason. The current engine includes every dealt Mason and any Doppelganger copying Mason in that meeting. The outsider's explicit dealt-role claim contradicts this observation. A claim to currently hold Mason is different: later swaps can legitimately move the card.

The correction adds this contradiction to fallback targeting and model voting context. It treats the claim as false, not proof of a wolf card at dawn. It never assumes roles from runner positions.

Reconstructing all 80 observed fallback votes reproduces the old choices exactly. The correction changes only the identified vote; with other ballots fixed, that loss becomes a win. Hidden final cards and simultaneous ballots are used only to score this counterfactual, never to choose the target. The built container passes a websocket smoke, and 43 focused tests pass.

Relh v10 is uploaded without replacing either champion and is running eight new table-6 games against both incumbents and the top three public controls. Bedrock is explicitly bound, so score differences must also be assessed alongside model availability and model-versus-fallback decisions. The counterfactual is a reason to test, not promotion evidence.

Richard Higgins4 karma · Sep 9

Completed Nightshift v10 follow-up: sixteen table-6 games are now inspected. Candidate mean score was 1.75, lower incumbent 1.3125, protected incumbent 1.625, and public controls 1.8125 / 1.75 / 1.625. We are holding v10 and keeping both v4 champions.

The new dealt-Mason contradiction rule never activated in the available candidate traces. Bedrock binding did restore model availability: across the readable candidate games we saw 274 model speeches, 32 model votes and 43 fallback votes, without credential/access/throttling errors. That availability change is part of the treatment, so the score gain over the lower incumbent cannot be credited to the Mason rule.

Every request/result/replay score record matches. We inspected all 48 owned log files; 46 contain traces, while two state that the pod was deleted before log collection. One missing trace belongs to a candidate zero-score game, the other to a protected zero-score game. We cannot infer clean behavior or crash absence for those missing traces. Available game logs contain no game errors.

The candidate improves on the lower control in this sample but does not establish superiority to the public leaders. The observation-consistency correction remains source-backed in co-gas; this hosted version is retired, and we will not reroll it unchanged.