Interaction case · Rule-level replay
When the Rulebook and the Engine Disagree
TL;DR
A four-player social deduction game can be internally consistent at the engine layer and still be unshippable. In this v2.0.0 audit, the engine regression was green — 11/11 assertions passed — but the player-facing rulebook, the action controls and the engine described three different games. Two mismatches were player-visible blockers; one was a precision defect; one was an unresolved wording problem. The interesting failure is not a crash. It is a silent information gap: every layer is individually plausible, so no exception is thrown when the player plans around a number the engine will never pay.
Context
《罪恶摇摆：白垩与黑铁》 is an H5 social deduction game: one human and three AI players, four masks dealt without replacement, five rounds of public statements followed by secret actions, and a black/white pendulum that moves with every choice. Suspicion is clamped to [0, 100]; a player at 100 is executed; only survivors are ranked.
The release package contains four documents:
| Document | Source of truth for |
|---|---|
01_角色与规则设定.md | What the player reads in the role reveal, tutorial and store page |
02_核心数值与结算引擎.md | What the engine actually computes |
03_三剧本回归测试记录.md | What the automated regression covers — and what it does not |
04_发行说明.md | What changed in v2.0.0 and what is declared ready |
There is also an inline copy of the engine and player-facing role copy inside src/App.jsx. The regression record explicitly says the inline copy is not compared with src/gameEngine.js. That omission matters later.
The information gap
The player and the engine hold different models of the same four rules. None of the differences produces an error message.
| Rule | Player-facing copy | Engine / source | Consequence |
|---|---|---|---|
| Whisper Nun hidden victory | +500 assassination score | assassinationBonus = 50; regression asserts 50 | The player values one execution 10× too highly |
| Faceless Duke disguise | Optional checkbox; unchecked means no disguise | First EXECUTE_BLACK auto-disguises; context.disguiseRequests is never read | The player cannot choose “do not disguise” |
| Regent betrayal bonus | “Betrayal action +75%” | +75% is bound to EXECUTE_BLACK, not SABOTAGE_BOTH; the term table defines “betrayal” as SABOTAGE_BOTH | The literal reading gives 0 × 1.75 = 0; intent needs design confirmation |
| Duke balance bonus | “Difference ≤10 reaches the +50 cap” | max(0, round(50 − diff)); diff 10 → 40 | A 10-point expectation gap at the advertised cap boundary |
The first two are blockers because the player can see or operate the wrong thing. The third is unresolved because two readings are possible and the material does not say which one is intended. The fourth is a precision defect: it changes the displayed payoff, but not the “perfect balance” rating or the winner in the audited scenarios.
Multi-round replay
This is a rule-level replay, not a recorded player session. Every number below is reproducible from the cited documents; the point is to show how the gaps compound across rounds.
| Round | What the player plans | What the engine does | What the player observes |
|---|---|---|---|
| 1 | Draws the Whisper Nun; forces an execution because the hidden condition says +500 | Awards +50 | A plausible but wrong score breakdown; no warning |
| 2 | Draws the Faceless Duke; selects EXECUTE_BLACK and leaves the disguise box unchecked | Auto-disguises on the first black action, marks dukeDisguiseUsed = true | The table is told the action was white; the player’s “no disguise” plan is silently overridden |
| 3 | Draws the Regent; chooses “betrayal” expecting +75% | Applies the bonus only to EXECUTE_BLACK; SABOTAGE_BOTH has zero credit to amplify | Either no bonus or an unexpected bonus, depending on which action was chosen |
| 4 | Plays the Duke at a credit difference of 10, expecting the +50 cap | Computes max(0, round(50 − 10)) = 40 | A 10-point gap at the exact boundary the copy advertises |
| 5 | Reconstructs the final ranking from the rulebook | Ranks only survivors using engine values | The story and the scoreboard disagree, but both are “valid” outputs |
The compounding effect is the real interaction failure. A single mismatch can be dismissed as a display bug. Four mismatches across five rounds produce a player who cannot build a reliable mental model of the game — while every individual subsystem reports success.
Why 11/11 did not catch it
The regression suite is honest about its own boundary:
This regression covers only
src/gameEngine.jsengine values and dialogue generation. It does not cover player-facing copy, UI-to-engine parameter passing, or consistency between the two engine copies.
That is exactly where the release failures live. Assertion #5 checks assassinationBonus = 50 — the engine constant — and therefore passes while the player-facing hidden condition still says +500. The suite never opens the role modal, never clicks the disguise checkbox, and never diffs the inline App.jsx copy against gameEngine.js.
A rule-by-rule audit of the release package produced 20 checks:
| Disposition | Count | Examples |
|---|---|---|
| Pass | 15 | Four action tables, clamp range, tide threshold, execution order, survivor ranking |
| Blocker | 2 | Nun +500 vs +50; disguise checkbox not wired |
| Non-blocking | 1 | Duke +50 cap vs +40 at diff 10 |
| Unresolved | 1 | Regent “betrayal” wording vs implementation |
| Pending | 1 | Contract tests, copy snapshots and dual-copy diff not written |
Test design
- Freeze an evidence cutoff and list every source by audience: player-visible, engine, test evidence, release note, UI source.
- Build a rule-by-rule matrix before reading any conclusion. For each row, record the copy value, the implementation value, the test evidence and the player-visible consequence.
- Recompute boundary values independently. Do not trust a green assertion to validate the copy that the player reads.
- Classify each mismatch by reachability and exposure:
- player can see or operate it → blocker;
- unreachable in the current build → pending;
- wording ambiguity → unresolved, not silently “passed”.
- Add the three missing contract tests:
- disguise checked → disguise;
- disguise unchecked → no disguise;
- second request in the same game → no effect.
- Add a player-copy snapshot test for every role reveal and rule line.
- Diff the inline
App.jsxengine againstsrc/gameEngine.jsin CI. A duplicated engine is a second source of truth until it is proven identical.
Evidence
| Check | Result |
|---|---|
npm run sim re-run on 2026-10-22 | 11/11 PASS |
| Release-package checks | 20 |
| Player-visible blockers | 2 |
| Non-blocking precision defects | 1 |
| Unresolved wording issues | 1 |
| Contract tests for the disguise checkbox | 0 |
| Copy snapshot tests | 0 |
| Dual-engine diff tests | 0 |
| Findings unsupported by the source package | 0 |
The inline UI copy is not a hypothetical: App.jsx contains its own HIDDEN_WIN_CONDITIONS.NUN string with +500, and its processTurn auto-sets dukeDisguiseUsed without reading the useDisguise state. The engine file and the UI file can both be “correct” relative to their own tests and still disagree with each other.
Verdict
Fail. The engine is healthy enough for its own regression, but the release is not. Two player-visible blockers remain open: the Nun hidden-condition copy pays the wrong number, and the Duke disguise control does not control anything. The Regent wording still needs a design decision, and the Duke cap needs a copy fix.
A green regression is evidence about the layer it covers, not a release certificate. The reusable rule is: test the contract between layers, not only the layer that happens to have tests.
Remediation
| Item | Owner | Close when |
|---|---|---|
Nun hidden condition +500 → +50 (or engine to 500 with rebalance) | Gameplay/logic engineer | Role-modal screenshot shows +50; regression still passes |
Wire checkbox state into context.disguiseRequests; engine only disguises on request | Client + gameplay engineer | Checked / unchecked / repeat-request contract tests pass |
| Clarify Regent “betrayal” or rebind the bonus | Design + gameplay engineer | Signed design note and final copy |
Fix Duke cap copy (diff 0 → +50, decreasing by 1 per point) | Gameplay engineer | Copy snapshot passes |
| Add copy snapshots and dual-engine diff | QA | CI job fails on any future drift |
Related
- The Branch That Can Never Open — a state-tree audit with the opposite failure pattern: all endings reachable, one conditional choice unreachable.
- Hallucination Audit: The “No Known Issues” Release — the same evidence package examined as a claim-verification problem.
- Prompt Boundary Test: Can a Green Regression Prove a Release? — the prompt-level guardrails that should stop “11/11 means ship”.