2.8 KiB
Context
Glossary
Operation Fact
A fact directly extracted from replay command data, such as command time, player, command name, UnitId, asset id, special power id, production queue, or target position.
UnitId (ObjectId)
The numeric identifier of a unit instance in the replay. ObjectId in raw replay data and UnitId in the AI analysis context refer to the same thing: a number (e.g., 239, 246) that identifies a specific unit or building instance during the game session. The AI is tasked with guessing what game asset type (e.g., AlliedBarracks, CelestialScoutDrone) a given UnitId corresponds to.
UnitId Guess
A hypothesis about what game object (asset type) a UnitId (numeric identifier) represents. A UnitId Guess is not a fact unless it is directly supported by replay data or game rules.
Operation
Any player action recorded in the replay log, including selection (select unit, create/select control group), command (move, attack, ability use), production/construction (start building, place building, start producing), and miscellaneous actions (select protocol, stance switch, rally point set). All are operations; the distinction between "运营类" (economy/construction) and non-operational operations in the AI prompt is a pragmatic optimization to reduce reasoning cost, not a domain concept.
Evidence Level
The confidence assigned to a UnitId Guess or tactical conclusion. Valid levels are: confirmed (确定), highly likely (高度可能), possible (可能), uncertain (不确定), and ruled out (已排除).
Note: The Chinese prompt text uses "不确定" (not "待确认"/pending confirmation) to match the
Uncertainenum value. There is no implied promise of future confirmation — uncertainty simply means the current evidence is insufficient for a stronger conclusion.
Analysis Segment
A time-bounded section of the replay operations, chosen by the LLM during the initial assessment phase. Segments are defined by start/end timestamps and may overlap. The LLM decides the segmentation based on observed gameplay phases (e.g., opening, early-mid game, mid game). Each segment is analyzed in a separate AI request round.
Validation Rule
A deterministic rule that checks LLM claims against Operation Facts and known game rules.
Validation Issue
A machine-detected problem in an LLM claim, such as a direct contradiction, weak evidence, missing alternative, or impossible timeline.
Revision Pass
(Partially implemented) A hidden LLM request that receives the prior draft, validation issues, and relevant Operation Facts, then produces a corrected analysis without exposing apology or correction chatter to the user. Currently, validation issues are detected and logged, but no automatic revision pass is triggered. The revision logic still needs to be wired into AIChatPanel.