Identify wins related to Treasury strategic priorities

[metzgr]

Acceptance criteria

• Given an already-extracted win, the system can compare it against Treasury strategic plan goals, objectives, and key performance indicators.
• The system does not re-extract wins from source documents as part of this ticket.
• The system uses the existing win title, description, rationale quotes, metadata, and tags as input.
• Each win receives one alignment status: aligned, needs_review, ambiguous, or not_aligned.
• A win can be aligned to multiple strategic plan items when appropriate.
• Each aligned win includes the matched strategic plan item type: goal, objective, or key_performance_indicator.
• Each aligned win includes the matched strategic plan item name, ID or reference label, and source section when available.
• Each aligned win includes a short AI-generated rationale explaining why the win supports the strategic priority.
• Each aligned win includes evidence from the win record, such as title, description, rationale quote, extracted number, geography, population served, organization, or tag.
• Each aligned win includes evidence from the strategic plan item, such as goal text, objective text, KPI name, KPI description, or related strategy language.
• The system can identify alignment even when the win does not use the exact same wording as the strategic plan.
• The system uses win metadata to improve alignment, including numbers, populations served, geography, organizations, scope, strategy tags, and proof-point tags.
• KPI alignment is only assigned when the win plausibly supports or demonstrates progress toward that KPI.
• The system does not force alignment based only on broad or generic language.
• The system flags broad, weak, or uncertain alignments as needs_review.
• Ambiguous cases preserve the top possible strategic plan matches for human review.
• Not-aligned cases are stored with a brief explanation.
• AI output is validated before storage.
• Invalid AI output is logged and does not corrupt stored data.
• Raw AI responses are retained for debugging.
• Prompt/model version is stored with every alignment attempt.
• The AI can only select from strategic plan goals, objectives, and KPIs provided by the system.
• The AI cannot invent strategic plan goals, objectives, KPIs, IDs, or source references.
• The system can run on one win or the full batch of extracted wins.
• Batch results include counts for aligned, needs review, ambiguous, not aligned, and error cases.
• Human review status is supported for uncertain, ambiguous, or high-impact alignments.
• A small labeled eval set exists with clear alignments, weak alignments, ambiguous alignments, KPI alignments, and not-aligned examples.
• Tests cover alignment, validation, multiple matches, KPI matching, ambiguous cases, not-aligned cases, malformed AI output, and invented strategic priority rejection.