Trust & verification
Prose faithfulness
The deepest question a citable engine must answer: does the developed prose faithfully represent the primary-source law it cites? Not just is there a citation (claim-substantiation) — but does the cited source actually support the sentence. An offline adversarial judge panel(two independent judges, refute-by-default) judged each checkable prose claim against the instrument’s committed provision-excerpt corpus.
Offline adversarial judge panel (two independent judges per instrument). Each judge gives every checkable factual prose sentence a THREE-way verdict against the instrument's committed provision-excerpt corpus: supported (an on-point excerpt establishes it), contradicted (an on-point excerpt is misstated — a genuine defect), or out_of_corpus (no on-point excerpt — unadjudicable). A contradiction is refute-by-default (either judge suffices); supported requires both judges. Judgements are committed here; the serve-time report reads them deterministically (no serve-time model call).
Honest bound. The headline faithfulnessRate is supported / (supported + contradicted) — over the ADJUDICABLE claims only. out_of_corpus claims (whose supporting text is not among the catalog's committed excerpts — a date, a penalty figure, another law, a scholarly citation, or an un-excerpted provision) are EXCLUDED from the denominator and reported separately as a coverage signal, NOT counted against faithfulness. This keeps the rate an honest measure of whether the prose MISREPRESENTS the cited law, rather than an artifact of how much of the statute the catalog excerpts. 2026-07-02 cross-audit: the session’s later citation-inserts + 7-article deepening added checkable claims that were unjudged under the earlier headline; those 26 pending claims were re-judged on CURRENT prose by two independent judges (refute-by-default). 3 initially read as CONTRADICTED on GSA-AI-GUIDE — traced to the catalog’s own provision excerpts OVERCLAIMING the guide’s contents (SIN enumeration / sample clauses); primary-source re-verification (gsa.gov 2024-04-29) confirmed the guide is a considerations/questions resource, so the EXCERPTS were corrected and the claims re-judged (now 0 contradictions). Orphaned judgements from edited-away sentences were pruned. 2026-07-02: the 39 prior 'supported' verdicts were STRICT-re-judged under an explicit multi-limb rule (a sentence asserting several factual limbs is supported only if ALL limbs have on-point excerpts; else out_of_corpus) — 16 were multi-limb over-credits and moved to out_of_corpus, tightening the adjudicable set. faithfulnessRate remains supported/(supported+contradicted); this correction lowers the denominator, it does not inflate the rate.
Independently re-audited (2026-07-01). 0 missed contradictions found — the zero-contradiction result held under adversarial scrutiny. 1 over-affirmation corrected (FEDRAMP-AI-2024 supported→out_of_corpus). The 5 judge-split ‘uncertain’ verdicts are retained in the review queue (not resolved on a single auditor’s opinion).
Machine-readable: /wiki/prose-faithfulness.json. The cell-level sibling is /wiki/faithfulness.
Faithful (of adjudicable)
100%
Contradictions found
0
Out-of-corpus (coverage)
142
Claims judged
170/170
Review queue — prose that misstates an on-point provision
Of 23 adjudicable claims (an on-point excerpt exists to check against), 23 are faithful and 0 misstate the cited provision. A further 142 claims cite facts beyond the committed excerpt corpus (dates, penalties, other laws, scholarship) — unadjudicable here, a coverage signal that motivates excerpt backfill, not a faithfulness defect. The queue below is the genuine review set: contradictions and split verdicts.
- CA-SB-942uncertain
“SB 942's latent-disclosure mandate (§ 22757.3(b)) places California within a global drift toward provenance-by-watermark, but its design diverges in instructive ways.”
The §22757.3 excerpt directly establishes SB 942's latent-disclosure mandate ('A covered provider shall include a latent disclosure in AI-generated image, video, or audio content'), the load-bearing factual assertion of the claim.
- CN-DEEPSYN-2022uncertain
“Zhang and colleagues prove that strong watermarking of generative outputs is, under broad assumptions, impossible against a motivated adversary (arXiv:2311.04378), so the Art. 16 embedded-identifier d…”
The claim's assertion is a scholarly impossibility result about watermarking against adversaries; the Art. 16 identifier duty and Art. 18 anti-removal prohibition are named (excerpts [3],[4] support their existence) but the claim's substance about circumvention survivability is not established by any excerpt.
- CN-DEEPSYN-2022uncertain
“Compared with the liability-and-privacy lattice that Novelli and colleagues map across EU law for generative AI (10.1016/j.clsr.2024.106066), the Provisions are administratively front-loaded: filing (…”
Excerpt [6] names Art. 19 filing and Art. 20 security assessment, but no excerpt establishes Art. 9 real-name verification nor the ex-ante-through-regulator vs ex-post-through-courts characterization that is the claim's core comparative assertion.
- DOD-RAI-2022uncertain
“The June 22, 2022 Strategy and Implementation Pathway (S&IP) is an internal DoD policy statement, not a statute, that operationalizes the five Ethical Principles for AI adopted Feb. 24, 2020 — Respons…”
The four excerpts are the Ethical Principles Governable, Reliable, Responsible, and Traceable, corroborating the named principles; the excerpts confirm these are DoD principles for AI (the date and 'not a statute' framing are extra but the on-point principle-naming assertion is established).
- FEDRAMP-AI-2024uncertain
“Agencies and the authorisation pathway are directed to weigh model-specific risks — training-data exposure, prompt-injection, and unintended output disclosure — when scoping the System Security Plan a…”
The excerpt covers disclosure of provenance/evaluation/model documentation, but does not direct agencies to weigh prompt-injection, unintended-output-disclosure risks, or select NIST 800-53 control overlays, so that specific assertion is not represented.