Repository navigation
Qualification and roadmap
The code can check bounded evidence consistency and calculate acceptance. It cannot certify independent annotations or infer institutional authority from a JSON declaration. There are zero qualified held-out entries and no PBI score.
The study of three procurement documents comprising 200 supplied pages is closed incomplete. Three model configurations and three assigned repetitions cover 27 fixed assignments: 14 completed, five failed or incomplete, and eight not dispatched. The additional Google supplement was cancelled before dispatch. Outcomes are reported on the generation results page. Reference decisions are assistant drafts, and unknown charges remain reserved. This development study does not satisfy the external qualification gates.
- Independent source families: scope, rights, exposure and split decisions recorded before evaluation.
- Blinded reference preparation: domain annotators and an adjudicator agree meanings, legitimate alternatives and critical rules before seeing system outputs.
- External commitment: an appointed custodian acknowledges exact commitments through a verifiable record with independent timing.
- Metric calibration: seeded semantic defects, valid controls, detection, false alarms and disagreements are evaluated separately from tuning.
- Comparable execution: frozen configurations, all assigned attempts, failure accounting and complete cost disclosures.
- Uncertainty and human effort: family-aware analysis, correction sessions and limits on any ranking inference.
- Independent decision: identified domain and reproduction reviewers sign the exact scope, digests, conflicts, limitations and appeal route.
These gates are planned work. A public Git tag or a custodian's receipt alone does not validate semantic quality. No external board or custodian appointment is claimed.
Persistent proposition-bank reuse, external registration verification, qualified family-aware uncertainty and a validated human-effort workflow remain open. A paired family bootstrap is available for development diagnostics and does not establish qualification. RFC 8785 serialization is not implemented. The eight-dimensional PBI composite is a proposal, not an available score.
The maintainer also develops PALO Workbench. This conflict is disclosed and requires independent participation; separate branding does not remove it. Model consultations are advisory. The maintainer cannot replace missing independent approval with a self-issued decision.
Study RFC | Launch boundaries | Governance
Validation status | Measurement protocol (draft) | External review requirements
Research pilot | Zero qualified held-out results | Versioned wiki source | Corrections