FilingsBench / known_failures.md
UraionLabs's picture
Publish FilingsBench v0.1 with raw baseline results
4350aa2 verified
|
Raw
History Blame Contribute Delete
1.11 kB
# Known failures and non-claims
- FilingsBench v0.1 has no human-gold or narrative labels yet.
- Numeric concept selection does not yet cover issuer extension concepts.
- 8-K support is not included in v0.1 because period-duration rules differ.
- Inline-XBRL facts with segment dimensions are intentionally abstained.
- A document can report more than one valid notion of revenue. v0.1 uses a fixed
US-GAAP priority and exposes the chosen concept as evidence.
- Confidence is rule support, not a calibrated probability. Calibration metrics
remain unreported until a sufficient gold-labeled set exists.
- Registrant names are strict-matched in v0.1. Filing legal names can differ from
SEC Submissions aliases (for example `CORPORATION` vs `CORP`), so the strict
score understates entity identity matching; no fuzzy-match score is claimed.
- The deterministic baseline does not evaluate any Uraion model. No model-quality
claim is made.
- The Dockerfile is provided but was not built in the audit environment because
no Docker runtime was installed; CI verifies the native Python package only.