This page is the standing disclosure behind PRISM and every Coverage Audit: the full 15-technique taxonomy, the fixed 0–10 Manipulation Index bands, the quote-verification protocol that separates a finding from a guess, and the exact models doing the work. When any of it changes, the changelog at the bottom says so.
Every analysis draws from the same fixed library — the same names, the same definitions, whether the source leans left or right. The framework flags the persuasion mechanism, never the party. If a technique isn't on this list, we don't invent one.
One number per article: how much persuasive pressure the text applies, on a fixed scale with fixed band cut-points. The index measures pressure, not truth — a factually accurate piece can still score high, and a wrong one can score low. The verdict on the facts stays yours.
The cut-points are identical on every StrataLens surface — the free PRISM tool, Coverage Audits, and standing monitors — so a 6.8 means the same thing everywhere, and two reports are directly comparable.
The step that separates a StrataLens finding from a chatbot guess: every finding must rest on a quote from the source, and every quote is checked by code before it ships. A quote that cannot be verified is flagged — it is never asserted.
The analysis sees the article's actual words on the page — not a model's memory of what an outlet "usually" says.
The model is required to mark each quoted span explicitly as it writes. Unquoted assertions about the text don't qualify as findings.
Deterministic code — not a model — verifies every marked span, character-for-character after typographic normalization (curly vs. straight quotes, letter case, punctuation spacing), so formatting can never fake a match or fake a failure. The quote's full word sequence must appear in the source text. Spans shorter than 12 normalized characters are too ambiguous to verify and never count as verified.
Anything that fails the check is flagged as unverified or removed. It is never smoothed over into a finding, and the counts — quotes checked, quotes verified — are reported, not hidden.
Every Coverage Audit includes quotes.json — a machine-readable ledger listing, for every analysis layer of every article, how many quotes were checked and how many verified against the source — plus the exact text of any quote that failed verification, tied to its article URL. Failures are printed, never smoothed over; they're what the guarantee pays on. See a sample ledger → · the published schema →
quotes.json here, ,
or try the sample ledger
Which model runs which stage, by exact ID. A model never grades its own citations — verification is deterministic code.
When a model ID changes, the change is recorded in the changelog below, and reports produced after the change cite the new ID.
A score means more with a baseline behind it. A fixed public reference corpus of scored articles lets any article's Manipulation Index also be read as a percentile against the wider press.
The v0 baseline is published: 121 articles from 11 mainstream US outlets across the political spectrum (left/center/right), each scored on the same 0–10 Manipulation Index, median 3.5. Full distribution, per-lean medians, and composition are on the State of the Baseline page.
Every Coverage Audit now reads its average score as a percentile against this corpus (and a "manipulation debt" vs the reference average). A percentile only appears once the corpus has at least 30 scored articles; below that the report shows absolute 0–10 scores only, never a fabricated percentile.
A methodology that maps persuasion techniques should survive contact with its own sales copy. So the same PRISM read that scores news coverage runs against StrataLens's own Coverage Audit page — and the result publishes here unedited: score, band, and every technique it catches us using.
2 candidate quote(s) did not verify and are excluded above, per the same protocol every audit uses.
Every substantive change to the taxonomy, the bands, the verification protocol, or the models lands here, dated. Corrections included — especially corrections.
Entries are append-only. If a change would make older reports non-comparable with newer ones, the entry says so explicitly.
Try a single headline in PRISM free, or put the full protocol to work across your beat with a Coverage Audit.