Ground bot intentions in live game knowledge and verify outcomes

2026-09-25 13:28:03 +08:00
parent 20ccd0fced
commit 502cb1e4fd
17 changed files with 251 additions and 24 deletions
@@ -0,0 +1,37 @@
+# Grounded decisions and gameplay evaluation
+
+The live model receives up to eight active quest-log entries from this server's
+quest_template table, nearby hostile units within 35 yards, health/mana, fresh
+inventory readiness, eligible nearby party candidates, and recent measured
+outcomes. Static quest titles are cached in ALE. Live facts remain snapshots;
+missing/stale readings are not replaced with invented details.
+
+Fully recovered bots cannot select rest when another allowed action exists.
+Low mana joins low health as a recovery override. Dead/ghost, human-controlled,
+and taxi residents are excluded from intention selection. Native survival,
+trainer, human-party and movement guards remain authoritative.
+
+Socialize requests a guarded native party invitation when an eligible candidate
+is present and neither participant has a pending/recent invitation. Otherwise it
+requests the normal social strategy. Confirmed membership, rather than narrative
+claims, verifies party formation. Trade also requests normal supply purchasing
+and repair, never spawned items or free repairs.
+
+Each resident retains 12 intention checks. Later fresh snapshots can verify
+travel, health/mana recovery, XP/level progress, supply/equipment improvement,
+or the intended companion's party membership. XP is not called a quest turn-in;
+inventory improvements are not called verified purchases. Death or human control
+interrupts a check. Ten minutes without matching observations is unverified,
+not proof the engine failed. Confirmed dialogue has its own existing counters.
+
+The monitor exposes sampled online/active bot-hours, native death deltas including
+normal resurrection, quest turn-ins per bot-hour, and observed/unverified/
+interrupted intention results. Rates start with this evaluation deployment and
+exclude collection gaps over two minutes. Native deaths use retained database
+records; retention rollover can undercount. Quest/XP rate and deaths must be read
+together: idling is not successful survival. Aggregates alone are insufficient to
+compare different levels/classes; inspect per-bot gameplay tables as well.
+
+Run `python3 scripts/evaluate-gameplay.py` on the mini PC for a read-only summary.
+Compare the first full day, investigate failures, and retain the existing model.
+No fine-tuning or paid compute is started by this change.