methodology
Methodology
Pack structure, September 8 model roster refresh, scoring fields, and track definitions.
Tracks
- Transaction categorization
- Vendor normalization
- Client context fit
- COA mapping
- Split detection
- Tax sensitivity
Input
Each participant receives the same statement pack, target chart of accounts, and normalized output schema.
Output
Outputs are normalized before scoring and merged into one overall table plus six track tables for the active release.
Run protocol
Passes3 seeded · median reported
Batch size32 rows per request
Samplingtemperature 0 where accepted
Output cap8,192 tokens per request
Retries2 on schema failure
ContextCOA + client profile + 90-day vendor history
Normalization
Category labels pass through a 231-entry alias table before exact match. Merchant strings are lowercased, stripped of store numbers and card-processor suffixes, then matched to 20 gold entities. Account codes are matched as-is against the 48-account target chart.
Score columns
Overallweighted composite of six tracks
Track score0–100, median of passes
Reliability% schema-valid on first attempt
Latencymedian s per 32-row request
Cost idxlist USD per 1,000 input rows
Track weights are frozen per pack version and ship with the harness notes.
Version history
v1.0 · Apr 14, 2026512-row pack, 4 tracks
v1.1 · May 12, 2026+ split detection, + tax sensitivity
v1.2 · Jun 29, 20263-pass protocol, reliability/latency/cost
harness 1.2.3 · Sep 8, 2026roster refresh, alias table 214 → 231
The Sep 8 alias fix lifted the nine re-run rows by +0.3 to +1.1; the pack and gold set are unchanged.