Replace heuristic task routers with a single relational model

[plur9]

Context

Inspired by Kumo Online Serving (Jure Leskovec / Kumo.ai, 2026-05) and KumoRFM-2 — a pretrained relational foundation model that operates on connected tables natively, no flattening required.

References:

• https://kumo.ai/company/news/low-latency-high-throughput-predictions-with-kumorfm-2-fine-tuning/
• https://arxiv.org/abs/2604.12596

Problem

Several Datacore agents make routing/prioritization decisions using keyword and regex heuristics:

• strategic-prioritizer — task→Intent Graph alignment scoring
• queue-optimizer — nightshift execution order
• learning-classifier — engram routing for new learnings
• gtd-inbox-processor — classify and route inbox items
• meeting-router — daily vs weekly meeting routing
• tag-suggester — tag recommendation
• meeting-router escalation detection

Each ships its own ad-hoc logic. They share inputs (the entity graph: tasks ↔ projects ↔ people ↔ companies ↔ ventures ↔ journal) and they share ground truth (nightshift approvals/rejections, completions, escalations, feedback events).

Proposal

Replace these per-agent heuristics with one fine-tuned relational model that scores (item, context) → action_distribution. Train on nightshift outcome logs.

Pilot path (lowest risk):

1. queue-optimizer first — approval signal already exists in nightshift logs.
2. If lift is real, port strategic-prioritizer next.
3. Then gtd-inbox-processor routing and meeting-router.

Acceptance

• One shared scoring service callable from any agent
• Pilot replaces queue-optimizer heuristic with measurable lift on next-day approval rate
• Heuristic fallback retained when model confidence is low
• Documented as part of the agent registry pattern

Related

• See "/today two-stage briefing" issue for where the model's output gets consumed at morning time
• See "Nightshift as training stage" issue for how this model gets updated