ADR 0053: Keep Gameplay Design Human-Led And Experience Claims Bounded

Status

Accepted.

Context

ADR 0052 added a direct human play gate after a thin complete loop. That fixed one sequencing failure but left an earlier authority problem: the delivery workflow still told the agent to frame the experience hypothesis. An agent could invent a gameplay premise in isolation, implement it competently, and only then ask the human whether the premise was worth making.

Lanternworks and Followspot demonstrate the cost. Both were mechanically healthy enough to support architectural conclusions, yet the owner found them incomprehensible, poorly designed, and not fun. The failure was not only late playtesting. Their gameplay premises and presentation were expanded without sufficient design dialogue with the human who held product authority.

Established game-design and player-research methods provide useful lenses, constructs, simulations, and study techniques. They do not make enjoyment an agent-computable property or make a synthetic policy representative of human experience.

Decision

When work originates a game or materially redefines gameplay:

1. Recover existing human intent and separate it from implementation accidents and prior agent assumptions.
2. Discuss the highest-leverage unresolved gameplay decisions with the accountable human designer before substantial implementation.
3. Let the agent research theory, model systems, expose assumptions, and compare alternatives with costs, risks, and evidence limits.
4. Require a human-approved gameplay hypothesis covering target context, fantasy, verbs, intended experience, the first minute, short loop, success, failure, and stop conditions.
5. If invention is explicitly delegated, permit the agent to draft a bounded hypothesis but still require approval before broad implementation.
6. Use automation for mechanical facts and encoded-policy exploration. Label predictions about clarity, pacing, fairness, balance, accessibility, or enjoyment as hypotheses.
7. Reserve lived-experience claims for direct human play. Owner acceptance can direct a personal project; claims about a broader audience require representative target players.
8. Continue with ADR 0052's thin complete loop and Player Experience Gate after intent is approved.

This is recommended delivery guidance, not a universal Safety Kernel rule. Ordinary fixes, refactors, faithful implementation of approved designs, and settled tuning do not require a new design dialogue.

Consequences

• Human attention moves before implementation has multiplied an unapproved premise.
• The agent remains useful as a researcher, systems modeler, critic, and implementation partner without becoming the silent product owner.
• Design conversations focus on decisions that change the slice rather than imposing a complete questionnaire.
• Owner acceptance remains sufficient for personal direction but no longer widens into an unsupported target-player claim.
• Automated playtesting and telemetry retain high mechanical value while their experiential limitations stay explicit.
• Player-facing autonomous dogfood must pause for design approval and human play even when the engineering is healthy.

Evidence Boundary

The decision is supported by two failed same-owner experiential dogfoods, the contrasting human correction history of The Circussy One, and established game-design and user-research literature. It has not been independently tested as a complete STAGE sequence and does not establish that one dialogue format or theory set is optimal across genres, teams, or audiences.