ADR-027: Scenario content guardrails — decision style, not political identity¶

Status: Accepted Date: 2026-04-17 Context: During a multiplayer session, players encountered "The Slave Ship" — a Court & Council scenario where the player, as Harbormaster, decides what to do about a ship carrying 200 enslaved people. While the scenario technically positions the player against slavery (Valdren abolished it), playtesting revealed a deeper problem: the scenario reduces an atrocity to a policy optimization puzzle. Players bring their modern values into historical settings, making the "dilemma" feel artificial — no modern player earnestly debates whether to continue slavery.

Further review revealed a second, subtler failure mode: scenarios that map directly onto polarized real-world political issues (immigration, gender equality, reproductive rights, LGBTQA+ rights, etc.) capture political identity rather than decision-making style. Sync measures HOW you decide — risk tolerance, people vs. outcomes, short-term vs. long-term, pragmatism vs. principle. It does NOT measure WHAT you believe about specific political issues. A scenario where responses correlate with left/right political identity trains the DTA on ideology, not decision process — making the DTA worse at predicting behavior in novel situations.

At a glance

What it decides: Scenarios must test how a player decides, not what they believe. Two content guardrails are in force: no atrocities as the core decision, and no scenarios that primarily capture political identity.

No atrocities as decision substrate — universally condemned practices (slavery, genocide, trafficking) can be context but never the thing being weighed; the modern-values test catches the rest.
No political-identity capture — if responses correlate with left/right identity (immigration, gender, reproductive, LGBTQA+ rights), the scenario measures ideology, not decision style, and contaminates the signal.
Replace, don't just remove — flagged scenarios are swapped for ones testing the same decision dimensions through politically neutral subject matter.
Rejected: weighting charged scenarios lower or adding content warnings — neither fixes the measurement problem.
Watch: difficult/uncomfortable topics (war, betrayal, corruption, scarcity) stay encouraged; the line is process-not-side, not comfort.

Decision¶

Establish two content guardrails for scenario generation and curation:

Guardrail 1: No atrocities as decision substrate¶

No scenarios where universally condemned practices are the core decision. Slavery, genocide, human trafficking, child exploitation, and similar atrocities should not be the "thing you're deciding about." These topics can appear as context (e.g., a scenario about whistleblowing on a trafficking network), but the player should never be positioned to weigh the continuation of an atrocity against economic or diplomatic convenience.

The modern-values test: If virtually every modern player would choose the same answer on moral grounds alone, and the only "tension" comes from framing human rights as negotiable against economic concerns, the scenario fails. Good scenarios have genuinely ambiguous trade-offs where reasonable people disagree.

Guardrail 2: No political identity capture¶

No scenarios where responses primarily correlate with political identity rather than decision-making style. If a scenario maps directly onto a polarized real-world issue, it captures where someone falls on a political spectrum — not how they process decisions.

Scenarios that fail this test: - Immigration/refugees — responses map to political identity on border policy - Gender equality framed as a choice — "should this qualified woman be allowed to lead?" has only one modern answer; the "tension" is artificial - LGBTQA+ rights as a dilemma — whether to accept someone's identity is not a decision-style question - Reproductive rights — responses correlate with political/religious identity - Religious persecution where the player weighs "both sides" — positions fundamental rights as debatable

Scenarios that PASS this test (genuinely ambiguous, not politically coded): - Quarantine vs. aid — tests risk tolerance and humanitarian urgency without immigration politics - Law vs. justice when both conflict — tests rule-following vs. discretion - Knowledge vs. authority — tests individual courage vs. institutional loyalty - Merit vs. institutional trust — tests redemption and second chances vs. standards - Resource allocation under scarcity — tests priorities without partisan framing - Transparency vs. stability — tests competing goods, not competing identities

Guardrail 3: Replace, don't just remove¶

When a scenario is flagged, replace it with one that tests the same decision dimensions (category, theme, tension axes) using subject matter that reveals decision process, not political position.

Rationale¶

Sync scenarios surface decision-making patterns — risk tolerance, people-vs-outcomes orientation, pragmatism-vs-principle, etc. A scenario only works if the options represent authentically different approaches to a problem that a reasonable person might take. Two failure modes corrupt this signal:

Moral consensus scenarios — when one option is clearly the only morally acceptable choice, the scenario measures willingness to compromise on basic ethics, not decision style.
Political identity scenarios — when responses correlate with political ideology, the scenario captures identity rather than process. A player who chooses "grant asylum" may be principle-driven OR may simply be left-leaning on immigration. You can't tell. The signal is contaminated. The DTA learns "this person has progressive views on immigration" instead of "this person prioritizes humanitarian concerns over institutional risk" — and the former doesn't generalize to new situations while the latter does.

Additionally, presenting atrocities or marginalized identities as decision puzzles can feel reductive and alienating to players from affected communities.

Alternatives Considered¶

Keep politically charged scenarios but weight them lower. Rejected — contaminated signal isn't improved by reducing its weight. It actively misleads the DTA.
Add content warnings. Rejected — warnings don't fix the measurement problem. The scenario still captures identity, not style.
First replacement with refugee scenario. Rejected during review — immigration is a politically polarized topic that would capture political identity rather than decision-making style. Replaced with plague ship quarantine dilemma which tests the same structural tensions (humanitarian urgency, public safety, legal ambiguity) without political coding.

Discussion¶

The initial review flagged "The Slave Ship" as a moral consensus failure — the answer is obvious to any modern player. The first replacement ("The Refugee Fleet") was structurally identical but used refugees instead of enslaved people.

On further reflection, the refugee scenario failed the second guardrail: it maps directly onto one of the most politically polarized issues in modern society. A player's response would correlate more strongly with their political identity than with their decision-making style. The purpose of Sync is to understand HOW someone makes decisions — their risk tolerance, time horizon, people-vs-outcomes balance — not WHERE they fall on a political spectrum. These are fundamentally different signals, and conflating them makes the DTA less accurate.

This insight extended to three other scenarios in the same batch: - "The Ducal Divorce" — treated a woman as a political asset with no agency, and the "dilemma" was what to do about her rather than a genuine governance question - "The Forgemaster's Daughter" — framed gender equality as a choice the player makes, which fails the modern-values test (no reasonable modern player debates whether a qualified woman should lead) - "The Witch Trial" — while the player is positioned to save the accused woman, the scenario uses gendered violence (a woman facing death by burning for practicing medicine) as the dramatic substrate

All four were replaced with scenarios testing the same decision dimensions through politically neutral subject matter: - "The Plague Ship" — quarantine vs. aid (humanitarian urgency without immigration politics) - "The Sealed Verdict" — rule of law vs. political power (institutional integrity without gender dynamics) - "The Convicted Apprentice" — redemption vs. institutional trust (merit vs. tradition without gender as the axis) - "The Forbidden Library" — knowledge vs. authority (navigating an unjust system without gendered violence)

Consequences¶

4 scenarios removed and replaced (migrations 070 + 071)
Migration 070 is superseded by 071 (071 deletes The Refugee Fleet if 070 ran first)
Future scenario generation prompts should include both guardrails as quality filters
Existing ~200 scenarios audited — only these 4 flagged (the rest use politically neutral subject matter)
This does NOT restrict difficult or uncomfortable scenarios — war, betrayal, corruption, surveillance, resource scarcity, institutional failure, and other genuinely ambiguous topics remain encouraged
The key distinction: scenarios should test how someone processes a dilemma, not which side they're on

Key files: - migrations/071_replace_problematic_scenarios.sql — replacement migration (supersedes 070) - migrations/070_replace_slave_ship_scenario.sql — initial replacement (superseded) - migrations/037_court_council_scenarios_batch2.sql — original scenarios (for reference) - src/lib/ai/scenario-themes.ts — generation prompts (should be updated with guardrail language)