Intake
Candidate name, primary URL, row ID, and topic enter from the source matrix. The intake topic is preserved rather than inferred by the judge.
Evidence first, then gates, scores, and labels.
The catalog reflects the current CEI source pipeline and its drk_0805 rubric. Retrieval failures and uncertainty remain visible instead of being converted into low scores, and every published row carries assessment provenance.
Retrieve
The pipeline fetches and normalizes page or PDF text, extracts dates and metadata, and writes a crawl-quality report beside the exact evidence used for review.
Gate
Five eligibility gates test retrieval, AI focus, governance relevance, attribution, and whether the source is still in force. A failed gate prevents scoring.
Score
Gate passers receive 0–2 scores for actionability, authority, and currency. Judged claims include evidence rationales; currency is computed from observed dates.
Classify
Perspective, coverage topic, lifecycle stage, layer, track, and granularity are attached as non-scored labels, then the deterministic outcome rule is applied.
Stage 1 · Eligibility gates
G1 is computed from retrieval evidence. G2–G5 are grounded judgments over fetched text. When a required supporting quote is missing or invalid, the verdict becomes unresolved for review rather than an automatic rejection.
- G1 · Retrievable
Does the URL resolve to enough full text to assess?
- G2 · AI nexus
Is AI or an algorithmic system the source’s primary subject?
- G3 · Governance nexus
Does it address rules, oversight, rights, risk, safety, ethics with consequences, or policy?
- G4 · Attributable
Is a named author or issuing body responsible for the source?
- G5 · In force
Is this the current version, with no positive evidence of repeal, withdrawal, or replacement?
Stage 2 · Scored criteria
Only sources that pass all five gates are scored. If one criterion is unknown, the pipeline may pro-rate the available scores; at least two criteria must be scored before a tier can be assigned. Unknown evidence is never silently treated as zero.
Actionability
Distinguishes description or analysis from general direction and from a specific obligation, control, or duty.
Authority
Rates the status of this document and its issuer—not the authority of sources it merely cites.
Currency
Uses the newest observed publication or modification date against a freshness window selected by source perspective. Missing dates remain unknown, never guessed.
Decision outcomes
Scores determine a tier only after eligibility is established. Classification labels describe the source but cannot move its tier. The browser selects Core and Supporting by default; the complete ledger also exposes Context only, Unresolved, and Rejected rows.
Core5–6Highest-priority sources after every gate passes.
Supporting3–4Useful supporting sources after every gate passes.
Context only0–2Relevant context with lower operational or evidentiary weight.
RejectedGate failureFails at least one eligibility requirement and is not scored.
UnresolvedInsufficient certaintyEvidence, gate agreement, or score coverage is not strong enough for a final tier.
Reproducibility and review
Each completed pipeline run is an immutable bundle containing the normalized crawl evidence, quality report, gate judgments, scores, assessment CSV and JSON, rejection log, model usage, and a manifest. The manifest records input and rubric hashes, model configuration, options, timestamps, source-control state, and outcome counts. Per-source revisit dates keep review work explicit.
Published datasetdata/assessments.csv
Current rubricdrk_0805
Run artifactsManifested and hash-bound
Interpretation note
Automated judgments support source triage; they do not replace subject-matter, legal, or editorial review. Unresolved rows and pipeline flags are intentionally retained so reviewers can see where evidence or model agreement was insufficient.