Entity resolution

After per-event fragments are written, build merges them into canonical nodes and edges (oddsgraph/resolution.py, oddsgraph/schema.py).

Fragment merge

merge_fragments unions nodes and edges that share local IDs / edge keys:

• evidence market IDs are unioned
• aliases are unioned
• confidence takes the max
• evidence text prefers the higher-confidence edge

Resolution tiers

Fragment-local nodes map to canonical IDs using ordered tiers, including:

1. Exact Polymarket identifiers (for example event:… / market ids)
2. Exact slug / label matches within a node type
3. Alias matches
4. Fuzzy label matches via RapidFuzz (fuzzy_threshold, default 92)

Unresolved nodes keep an unresolved method marker until a later tier succeeds. Tier counts appear in inference_report.json under resolution_tiers.

MATCH identity

Dateful MATCH ids (match:<teams>-<YYYY-MM-DD>) identify fixtures. Two MATCH nodes that share a display label but carry dates more than one day apart do not merge (true rematches stay distinct). The same team-pair with dates within ±1 calendar day — the common Polymarket slug vs FIFA kickoff skew — coalesces into one node, preferring the official-bracket kickoff id when that fragment binds. Label-only MATCH ids may still upgrade to a dateful id when the official bracket or topology supplies the more specific form. Merging an official_bracket fragment also stamps inference_method=official_bracket on the canonical MATCH so progression edges stay consistent with official bracket data.

MATCH merge policy lives in oddsgraph/match_merge.py and is applied by oddsgraph/resolution.py.

Confidence filtering

--minimum-confidence rejects edges below the threshold during build. It does not change entity resolution itself. Rejected edges land in rejected_edges.parquet with a rejection_reason.

Ontology validation

Edges must match ALLOWED_EDGE_PATTERNS. Invalid patterns are rejected even if confidence is high. See Ontology.

See also

• Architecture
• Output artifacts
• Ontology