data
Data Card
Composition, generation method, provenance, and audit coverage.
Responsible innovation
Responsible-AI practice is built into the project through scope, evidence, failure visibility, and explicit limits on automation and claims.
Design choices
- Metrics from files only: scores are surfaced from committed experiment artifacts rather than invented presentation values.
- Label provenance: labels are rule-derived with documented audit coverage and aren't presented as a large human gold standard.
- Privacy boundary: synthetic organizations and evidence only; no employer, patient, or customer evidence.
- Human oversight: real-world use would require accountable human review; ControlSift isn't an automated auditor.
- Negative findings: within Gemma v1.0, QLoRA didn't beat few-shot and its low parse-success rate remains visible.
- Version honesty: classical v1.1 and Gemma v1.0 scores aren't presented as a controlled same-benchmark ranking.
Assurance documents
model
Model Card
Model context, intended use, and evaluation notes.
risk
AI risk register
Automation bias, leakage, overclaiming, and misuse risks with mitigations.
use
Intended use
Allowed research uses and prohibited production-auditor claims.
limit
Limitations
Synthetic-data gap, class difficulty, output reliability, and evaluation constraints.
audit
Label audit
Structural review coverage and stratified spot-check.