Triage & Insight Scoring¶

A single retrieval can come back with dozens of paths and a handful of motifs. An agent cannot act on all of that; it needs one number that says how much attention this deserves. That number is the triage score, an integer from 0 to 100, and it is the value most agent loops gate on.

How the score is built¶

The score is a weighted blend of five components drawn from the aggregation summary, each on a 0-1 scale:

score = 25·provenance + 25·recency + 25·surprise + 15·motif_density + 10·controllability

Component	Weight	Rewards
`provenance`	25	Well-sourced edges with references
`recency`	25	Fresh, recently-updated evidence
`surprise`	25	Deviation from the baseline / prior
`motif_density`	15	Concentration into repeated patterns
`controllability`	10	How actionable the finding is

The weighting is deliberate: provenance, recency, and surprise dominate because a finding that is well-sourced, current, and unexpected is the kind worth waking an analyst for. With everything maxed the score reaches 100.

Two guards keep it honest¶

Raw weighting alone would let flimsy results look impressive, so two guards pull them back down. If label_coverage falls below 0.8, motif_density is capped at 0.3 and a flat 15 points are subtracted, because poorly-labeled data cannot earn full pattern credit. And if low_support is set because there simply is not enough evidence, the whole score is cut by 40%. Together they ensure a confident-looking number is actually backed by confident-looking data.

A score is never a black box¶

Every triage result carries the breakdown that produced it, so you can always see why a number came out the way it did:

triage = result["triage"]
triage["score"]        # e.g. 87
triage["components"]   # each clamped input, the penalty applied, and the flags
triage["dominant_relation"]

{
  "provenance": 0.82,
  "recency": 0.74,
  "surprise": 0.61,
  "motif_density": 0.30,     # capped by the label-coverage guard
  "controllability": 1.0,
  "label_coverage": 0.71,
  "penalty": 15.0,           # the guard fired here
  "low_support": false
}

When a score comes back lower than you expected, this is the first thing to read: a penalty of 15 or a low provenance usually explains it immediately.

Triage vs. the other signals¶

Triage is the headline, but Odin reports finer-grained signals alongside it. The insight_score is an overall 0-1 quality measure, and ics decomposes it into the evidence_strength and community_relevance that make it up:

Field	Range	Use
`triage["score"]`	0-100	Prioritization: "should I look at this?"
`insight_score`	0.0-1.0	Overall retrieval quality
`evidence_strength`	0.0-1.0	Strength of the supporting evidence
`community_relevance`	0.0-1.0	Relevance to the community scope
`ics`	object	Decomposition of `insight_score`

The rule of thumb is simple: gate on the triage score, and reach for the others when you need observability rather than a decision.

result = engine.retrieve(seeds=[...])
if result["triage"]["score"] >= 70:
    agent.investigate(result["paths"])
else:
    agent.skip(reason="low triage")

That gate is the seam between Odin and the agent: AI Agent Integration builds it out into a full loop, and the Result Schema documents every field these scores live in.