ADR 0117: Play evidence preserves causal stages

Status: accepted (James, 2026-08-17). Amends ADR 0049 and ADR 0068. Amended by ADR 0160, which reads a retired action as archival evidence rather than a corrupt line.

Context

A turn digest proves that an observation exists, but it cannot show whether an action makes sense or works. Hosted worlds also move while the model thinks. Comparing decision state directly with post-action state attributes that ambient movement to the action.

Decision

Journal V2 stores one bounded causal evidence packet with every turn:

• exact semantic observations presented to the production mind;
• distinct immediate pre-action and post-action observations;
• the complete structured action result, not the bounded legacy detail string;
• available progress, refusal, repeat, stall, objective, notes, timing, and body provenance signals;
• null for a stage that does not exist, including pre/post state on invalid_decision and mind_failed turns.

PNG, PCM, credentials, media bytes, the full room transcript, and generated voice wording do not enter the packet. A narration attempt stores only the exact bounded game event offered to the room beside its speechDeliveryId. The words actually generated and heard remain unknown by policy. A delivery id is a join key, not delivery proof; only a matching played, suppressed, or refused receipt settles that question.

The offline evaluator reads every journal version. V1 evidence remains unknown. It may join the service's canonical CLANKIE_EVENT_LOG/CLANKIE_STATE event trail for terminal lifecycle and the canonical DISCORD_BRIDGE_RECEIPT_PATH receipt trail for narration delivery. A missing summary is terminal only when a durable matching terminal lifecycle event exists; its actual outcome, including lease_lapsed, remains distinct from a completed summary.

Consequences

• Movement start, target, end, and effectiveness are inspectable without frame bytes or inference from hashes.
• Failed decisions and crash-recovered runs remain part of evaluation rather than disappearing from successful-turn averages.
• One journal line is capped at 256 KiB. An oversized line fails as a whole; no partial action result is made to look complete.
• Mechanical evaluator verdicts remain conservative. Missing semantic evidence is unknown, not failure or success.