Model card

astris-extraction-self-improving-v1

Self-improving document extraction: density-based routing (M3.1) chooses the parse path, field-level calibration (M3.2) tracks per-field reliability, and a correction-retrieval store (M3.3) injects past human corrections as few-shot guidance into the next parse.

At a glance

Kindreasoning llm
Version1.0.0
StatusActive — contributing to scores
Ensemble weight0
Trained atNot trained yet
Trained onn/a
ArtifactNo artifact (heuristic or pre-training)

Metrics

{
  "components": [
    "density_routing_m3_1",
    "field_calibration_m3_2",
    "correction_retrieval_m3_3"
  ],
  "validation": "synthetic_seed",
  "real_cohort": "pending"
}

Notes

Extraction pipeline, not a retention voter — weight 0. Improves as parse corrections arrive via the feedback loop. See api/ml/extraction_service.py, field_calibration.py, correction_retrieval.py.