Methodology and limitations
SF-005 release contract · SLOPFILTER_SPEC_V1.1 · report schema slopfilter_report_v0_2
What the service measures
The runtime segments prose claims, assigns a speech-act role, maps available evidence posture, checks provenance state, identifies selected inference and rhetorical patterns, and produces separate support, provenance, rhetoric, and inference components. The displayed aggregate is a weighted summary, not a probability that the text is true or false.
Evidence and citations
A citation-shaped token does not create evidence credit. Resolution of an identifier or reachable host confirms identity or reachability only. A claim receives confirmed support only when an explicitly resolved source also passes a claim-specific local text-alignment check. That check is heuristic and can miss paraphrase, context, qualification, tables, images, or contradictory language.
Document segmentation
Headings, contents, reference sections, high-confidence footnotes, and link targets are excluded from claim and hedge denominators while remaining available for provenance inspection. Repeated claims remain repeated so rhetorical saturation is visible.
Risk posture
An unsupported or overstated claim produces a high review posture regardless of document length. This prevents a severe finding from disappearing inside a long document. The posture is a review control, not a legal or factual verdict.
Known limitations
- Rules can generate false positives and false negatives.
- Source alignment is lexical and polarity-aware, not full semantic entailment.
- Document extraction can alter headings, tables, footnotes, and reading order.
- Outbound resolution is capped and may fail because a source is unavailable, protected, dynamic, or not machine-readable.
- Scores are not validated for consequential decisions and must not be used alone for them.
Release gate
The mandatory release gate includes adversarial semantic probes, frozen excerpts from Federal Reserve, CDC, NIST, and NASA documents, privacy-boundary tests, API tests, and release-package verification. These tests establish behavior on the covered fixtures, not universal accuracy.