Outcome

Saltwind Sound completed all six scheduled turns without a human adjudication or delivery decision. The committed history is split across four verified runs because each runner-code correction required a new pinned experiment:

Five intervening Turn 006 replay branches retain rejected or interrupted evidence and contain no commits. All nine run directories pass verification with their pinned code identities, and the current closure run passes normal current-code verification.

Final campaign state

All four leaders remain ACTIVE, eligible, supportive, and at zero misses. Turn 006 approved five decisions: the first Saltmere–Keelstrand Marker Beach barter, the fourth High Barley–Ternhaven Ford Exchange, and three grouped domestic spring-readiness outcomes. No elimination, emergent rule, or victory condition was activated.

The final independent review approved causality and all four private narratives. It found that the two bilateral interactions respected travel time, daylight, hydrology, and territorial boundaries; domestic outcomes followed prior investments; every mutation had valid provenance; and no hidden structural fact, foreign private intent, score, GM note, or trope review entered a player report. Four simulation-only mailbox deliveries match their approved source bytes.

Complete accounting

The 47,665,640-token total includes 32,137 preparation tokens, 25,478,911 source-run tokens, and 22,154,592 tokens after the GM's additional 26,000,000-token authorization. No usage reset was consumed by this continuation. Antigravity did not supply dependable dollar accounting, so dollar cost remains unavailable rather than zero. Starts without completed responses remain recorded and receive no invented token count.

Why the unattended runner stopped

The interruptions were runner integration failures, not human-GM gates:

1. A Google identity-endpoint socket reset was classified as an OAuth credential failure and stopped instead of retrying.
2. A final narrative review found a real GM/state privacy defect, but the runner had no route back through its unused adjudication round.
3. Validation feedback was lost when a transient transport failure occurred between two repair attempts, causing the analyst to repeat invalid evidence IDs.
4. Antigravity repeatedly re-read large duplicated role files and later produced relationship IDs where only fact IDs were legal.
5. The GM expressed a global trope rejection without copying each exact retrieval ID, exhausting clerical repair attempts.
6. Accumulated state and audit history grew role packets beyond Antigravity's practical input ceiling. The provider silently truncated the decisive tail containing submissions, decisions, disclosures, or later player drafts.

Runner changes

The runner now:

• classifies operative connection resets and service failures before incidental OAuth wording, while retaining terminal handling for invalid, expired, revoked, 401, and 403 credentials;
• preserves actionable validation feedback across intervening transport failures;
• sends Antigravity input directly to a fresh tool-free role worker and retains only structured finish output;
• limits analyst evidence references to state fact IDs;
• routes final-review GM issues through a remaining adjudication round, reapplies the replacement to the original pre-turn state, reruns disclosure review, and redrafts all reports;
• deterministically records REJECTED_ALL when a retrieval has no candidates, and expands a model's explicit single global rejection across the exact retrieval IDs without inventing an applied/adapted choice;
• sends current canonical entity context without accumulated audit histories, supplies reviewers exact GM changes and disclosure/fact envelopes instead of duplicate full states, and rejects Antigravity requests above 240,000 serialized bytes before a model call.

The combined regression suite passes 70 tests with zero failures. It covers the new transport, feedback, late-adjudication, trope-normalization, role-scoping, provider-size, isolation, resume, delivery, privacy, and six-turn closure behavior along with the original Workstream 5 checks.

Game findings

Contact pacing was voluntary and geographically grounded. Each pair moved from indirect evidence to observation, bounded greeting, and barter on a neutral boundary. No four-way convergence was forced. Refusal and delay preserved resources rather than receiving an artificial penalty.

Persistent maintenance created most of the meaningful strategy. Saltmere's bank work and fuel curing, Keelstrand's shed repair and prefabrication, High Barley's drying/storage and care systems, and Ternhaven's drainage, smoke control, and boat protection all produced later benefits. Winter safety resulted from prior preparation, explicit limits, and mundane labor rather than a narrative loss quota.

Named characters and customs remained legible across turns, although most internal disputes resolved cooperatively. The run therefore offers stronger evidence for continuity and operational planning than for hard political refusal or factional conflict.

The power model showed weak discrimination. Turn 002 assigned every player 0.8; Turns 003 and 005 assigned every player 1.0; Turn 004 gave three players 1.0. Turn 006 returned a wider but still narrow 0.8–0.865 range. Detailed model-generated plans appear to saturate the rubric, so these scores do not establish comparative balance.

The curated trope corpus remained pinned to logical digest 25791dd59cfc5cdb1eac939952a67bafe2064344000e4d6c7ab4d1666e84f89c. The AI GM rejected every retrieved dramatic pattern as unnecessary or causally mismatched. No catalogue fallback or expansion occurred.

Evidence limits

This result establishes a complete authorized six-turn AI-only campaign history with isolated player memory, simultaneous intake, autonomous AI-GM decisions, two disclosure gates, bounded recovery, verified private simulation delivery, and preserved accounting. Because technical failures required code changes and lineage restarts during the campaign, it does not by itself establish clean six-turn single-invocation reliability; that requires a fresh end-to-end run under the final code. Model review can still miss subtle leaks or causal errors. The run also does not establish human enjoyment, human-GM workload, broad model robustness, or statistical game balance.