Deterministic Keyword Scoring Refinement Design

Date: 2026-04-20 Parent Plan: 2026-04-20-scene-skill-102-deterministic-invocation-readiness-plan.md

Intent

Resolve the 9 deterministic dispatch ambiguity gaps found after normalizing the final materialized scene skills for 。。。 invocation.

The previous readiness pass proved that 92 complete packages can be selected by full-scene-name samples, but 9 scenes collide because generated include keywords are too broad for current score rules.

Fixed Gap Set

sweep-026-scene / 县区公司故障明细
sweep-034-scene / 售电收入日统计排程预测
sweep-037-scene / 嘉峪关可靠性分析报告
sweep-038-scene / 嘉峪关周报
sweep-039-scene / 嘉峪关故障明细
sweep-040-scene / 嘉峪关日报
sweep-041-scene / 嘉峪关月报
sweep-044-scene / 国网金昌供电公司指挥中心生产例会报告
sweep-045-scene / 国网金昌供电公司营商环境周例会报告

Scope

Allowed:

Refine deterministic include/exclude keywords for the fixed 9 scenes and direct collision partners when needed.
Run dispatch dry-run checks without browser execution.
Publish refinement decisions and readiness delta.

Forbidden:

Do not execute browser scripts.
Do not repair sweep-012-scene.
Do not change generated scripts.
Do not update official execution board.
Do not modify runtime dispatch code unless this design is superseded by a separate runtime-scoring implementation plan.

Strategy

Prefer manifest-level disambiguation first:

keep full scene names as primary keywords;
remove overly broad standalone tokens from colliding scenes where they create ties;
add distinctive phrase keywords only when present in the scene name;
use exclude_keywords only for direct mutually exclusive cases.

Runtime scoring changes are out of scope for this plan unless manifest refinement cannot make all 9 gaps uniquely selectable.

Completion Criteria

All 9 fixed gaps have final decisions.
Full-scene-name dispatch dry-run uniquely selects the expected scene for each fixed gap.
No new ambiguity is introduced for the complete 101-package set.
sweep-012-scene remains excluded.

2.2 KiB Raw Blame History