A study in
passing the blame.

Baseline 0.1-test-baseline is provisional. Each player starts with five public objects and four private action cards. Win by holding no scandalous objects; ALIBI does not count against winning.

• GIFT transfers the whole matching type. A two-action-card substitute is available only when the actor has no GIFT.
• EXCHANGE lets the actor select both objects. ALIBI moves normally; DENY and REDIRECT cannot stop an exchange.
• A gift has one response window. REDIRECT does not create a response chain. DENY stops the transfer; redirecting a final bundle still lets its sender win.
• Resolve objects, check for winners, then refill if the game continues. Refill the actor first, then clockwise among other players who spent cards.
• Unused objects stay out of play. Exhausted action draws reshuffle discards. Simultaneous winners share the win.
• No voluntary pass when a legal action exists. After 300 resolved initiating turns without a winner, record a research cap hit, never a gameplay draw.

Object distribution, substitute access, exchange selection, ALIBI movement, refill timing, and tie handling are executable test defaults. The longer guide referenced in the brief was unavailable; these results do not confirm final printed rules.

1. A deterministic engine held state, enumerated legal actions, and enforced transfers, responses, refills, and wins.
2. A moderator resumed only the required player. Each seat had a fresh agent with no inherited conversation history.
3. Players received rules, a seat-specific style, public objects and action history, their own hidden hand, and a legal action menu.
4. Each agent chose an action and supplied a private reason. The engine validated it; the moderator did not choose on the player’s behalf.
5. Four partial-session feedback records were collected independently before players saw one another’s responses. Public table talk was disabled.

The styles emphasized immediate shedding, saving defense, blocking the nearest winner, and consolidating matching groups. These are instructions to agents, distinct from the scripted implementations.

Eight fresh player assignments were used across two pilots. Exact backend model identity was not exposed. Shared filesystem tools remained technically available: independence relied on seat-only instructions and audited broker views, not an enforced sandbox. No fallback or hidden-information exposure was reported.

Random-legal samples legal physical actions. Greedy-shedding favors large bundles, winning exchanges, and consolidation. Defensive uses similar shedding logic while saving DENY for an imminent win. Leader-blocking adds a preference for gifting or redirecting toward opponents with fewer scandalous objects. None estimates hidden hands.

Primary seeds use 10000 + playerCount × 1000 + i, for i = 0…99. Starting seats cycle; style offsets change after each full seat cycle. The first five deals at each count also run with every other starter, giving 70 correlated rotation games.

Policy tie-breaking is seeded and samples physical actions; interchangeable card IDs affect weights. Legal-action counts therefore do not measure strategic depth.

Turn

One initiating move, including response and resolution. A response is a separate decision, not an additional turn.

Comeback

A winner previously held strictly the most scandalous objects after a nonterminal move. A second count includes ties. Setup and final positions are excluded.

Mechanical loop

An identical full mechanical state, including hidden hands, deck order and RNG, excluding monotonic administrative counters. A repeated tactic is not necessarily a loop.

Kingmaking opportunity

An exchange can give the actor’s ALIBI for another player’s final scandalous object, awarding that recipient a win. This counts opportunity, not intent.

Wilson intervals describe primary scripted game-level rates. Rotation deals and within-game style assignments are correlated; naive independent-observation intervals are not used for those comparisons. This is exploratory, with no claim of a preregistered or causal experiment.

Three response conditions used the same 20 four-player initial states. All initiators used greedy-shedding; game rules stayed unchanged.

The spend-defense condition uses DENY when available, otherwise REDIRECT toward the lowest scandalous count, breaking ties by seat. Defense changed play without forcing a cap hit in these deals. Faster does not necessarily mean better. These are policy stress tests, not rule variants.

All ratings are out of five. A higher repetition rating means more repetition. There are only four simulated respondents, all from the same partial game; these are descriptive judgments.

Detailed memorable moves, alternatives considered, and reasons are included in the study data download. P3 disclosed that a retrospective history response was truncated; earlier assigned live observations covered those turns. That limits confidence in the retrospective review.

The recorded validation passed 32 engine and broker tests. All 530 completed scripted games passed exact replay and public/private projection audits, covering 22,712 scripted decisions. Both retained agent replays also reproduced their states; all 19 corrected-pilot decision views passed public-history delivery checks.

The 400 primary games recorded zero invalid choices, forced passes, conservation failures, or detected full mechanical loops. No cap hits occurred in the separate 70 rotation games or 60 response tests. These checks support the implementation and retained observations; they do not retroactively validate the first pilot’s information quality.

Full state hashes, legal choices, private draws, and reasons remain in the local research archive. Downloads contain selected summaries and feedback, never the complete hidden-state or player transcripts.

The current teaser says: “Swap one object with another player.” A more explicit explanation under the tested interpretation would be:

Choose one of your objects and one of theirs, then swap them. DENY and REDIRECT cannot stop EXCHANGE. An ALIBI does not count against winning.

Also explain that redirecting a final bundle still lets its sender win; DENY stops the transfer. Confirm the actor-selects-both interpretation before using it in printed rules. The extra explanation has not been tested; observe whether it prevents mistakes without slowing teaching.

Study ID NM–01 · recorded 13 October 2026 (Asia/Kolkata; run timestamps use UTC) · rules version 0.1-test-baseline · exploratory, partial study.

• playtest/results/findings.md — sample accounting, findings, and limitations.
• playtest/results/validation.md — validation and invalidated-pilot record.
• playtest/results/ambiguities.md — rule authority and provisional defaults.
• playtest/results/baseline-scripted/ — primary and rotation metrics.
• playtest/results/audit-*/ — matched response-policy conditions.
• playtest/runs/agents/games/pilot-4p-002/feedback-partial/ — four independent partial-session feedback files.
• playtest/README.md — engine, policy, orchestration, and replay instructions.

These are repository references, not public file links. Full research logs stay local. The workspace had no Git revision; configuration hashes, replay states, source hashes, and an archived harness snapshot provide provenance.

npm run test:playtest
npx tsx playtest/src/runner.ts verify-run playtest/runs/baseline-scripted
npx tsx playtest/src/runner.ts replay playtest/runs/agents/games/pilot-4p-002
node scripts/prepare-playtest-study.mjs

The last command regenerates curated website data from retained artifacts. Agent decisions cannot be reproduced by replaying prompts alone; the deterministic engine can replay the choices already recorded.