How the scores are made.
Every number on the wire is a measurement with a method behind it, and this page is the method: how a source earns a credibility score, how coverage becomes a narrative, how a claim is assessed against the sources that carry it, how motion is measured, and how a person or organisation is scored. It also says, plainly, what those numbers are not.
/methodology A score is earned, or it is not shown.
Every outlet on the wire is scored from 7 signals, each measured separately and refreshed on its own schedule. The score is a weighted combination of them on a 0–100 scale, banded into tiers, and it is published only when at least 5 of the 7 have real measurements behind them, the outlet has published at least 10 items in the last 30 days, and its first item is at least 7 days old.
Below that threshold the source shows no score and is marked as having insufficient data. The rule exists because a formula fed defaults converges on the same middling number for every unknown outlet, which would look like a measurement and be nothing of the kind.
Authority
How much weight the outlet carries: its estimated reach and its standing in the topics it covers.
Fact-checking record
What independent fact-checkers have found in its published claims.
Correction rate
How often it corrects, and whether corrections are published at all.
Bias
Where its coverage sits relative to the outlets covering the same stories, on a −1 to +1 scale.
Transparency
Mastheads, ownership, funding and editorial policies that can be found and read.
Editorial standards
Stated standards and the evidence of them in the copy.
Cross-source corroboration
How often what it reports is carried, independently, by others.
Fewer than 5 measured signals, fewer than 10 recent items, or fewer than 7 days of coverage: no score, marked insufficient. A score once published is withdrawn if a re-run no longer qualifies.
Credibility is a property of an outlet's record, not of any one article. An article from a low-credibility outlet may be true; the score says how much corroboration to want before believing it. Every source's page shows which signals were measured.
Coverage becomes a story.
Each item on the wire is embedded as a 1,024-dimensional vector and clustered with the items making the same claim, whatever the outlet or the wording. The cluster is the narrative; its centroid — the mean of its items' vectors, weighted by each item's signal score — is where the narrative sits in the space. As new items arrive the centroid is recomputed and the history is kept as a trajectory.
A narrative is then decomposed into its atomic claims, and for every claim the coverage is grouped by framing — the model's own phrase for how a cluster of outlets tells it — and counted in independent networks, not outlets: twelve affiliates of one owner count once. That is what the Perspectives view shows as consensus, contested and selectively told, and what the selective lens filters on.
c(t) = normalise( Σᵢ wᵢ · embed(itemᵢ) for itemᵢ ∈ narrative at t ) wᵢ = signal(itemᵢ) a per-item signal-quality score ‖c(t)‖ = 1
Items are weighted by their signal score — a measure of how clear and informative the item is — so noise moves the centroid less than a substantive report. Source credibility enters in the assessment (§03) and in trend ranking, not in the centroid.
Do the cited sources support the claim?
An assessment is a language model's reading of a narrative's cited items, asking one question: do these sources support this claim? It weighs the sources by their credibility, looks for corroboration across independent networks, and returns one of four labels with the evidence passages it leaned on.
The labels are deliberately about the sources. "Supported by cited sources" is not "true"; "not supported by cited sources" is not "false", and neither is a statement about the person or organisation the claim concerns. Every assessment carries the disclosure that it is AI-generated from cited sources and may contain errors, and links the items it read so a reader can check.
A claim about a private individual is not assessed. A complaint about any assessment goes through the privacy request form and neutralises the label everywhere while it is reviewed — see §07.
Louder, or different?
Momentum is how fast coverage is arriving, measured against the narrative's own baseline rather than a global one. A story is emerging, accelerating, at its peak, declining or stabilised according to that velocity. A Watch agent fires when a narrative's coverage velocity rises past the number of standard deviations you set, or when an entity's reputation, sentiment, momentum or mention velocity crosses a threshold you set.
Drift is how far the story has changed shape: the cosine distance between the narrative's centroid now and at the previous snapshot. Volume can be flat while drift is high — the same story, reframed — which is usually the more important signal.
Entropy is how spread out a subject's coverage is across topics. A focused story has low entropy; a subject whose coverage is pulling apart into competing frames has high entropy, which tends to precede a change of state.
drift(t) = 1 − cos( c(t), c(t−1) )
= 0 → same story, same frame
→ 1 → the centroid has moved to an unrelated region Computed on every trajectory snapshot; the value is what the narrative page plots.
θ̄ = 0.04 rad
Pharma CEO, routine quarter. The narrative direction barely rotates
day-to-day; velocity is non-zero but the angle stays put. Bucket:
stable.
θ̄ = 1.14 rad
Same CEO, contested acquisition leaks on day 10. The velocity
rotates sharply — the discourse is now
in a new part of the space. Bucket:
pivot.
pᵢ(t) = share of credibility-weighted coverage on topic i at t H(t) = − Σᵢ pᵢ(t) · log₂ pᵢ(t)
In bits. Eight topics spread evenly is 3.00 bits; a subject with one dominant story sits well under 1.5.
H = 1.63 bits A steady quarter. One topic dominates — earnings — with everything else a thin tail. Low entropy; a narrow narrative.
H = 2.94 bits Same CEO, mid-crisis. The distribution flattens — coverage splinters across antitrust, regulator, lawsuit, and internal drama. High entropy; a narrative under pressure.
Scores describe coverage, not character.
A person or organisation in the corpus carries four scores on a 0–100 scale, each computed from the coverage about them over a stated window, each served with its basis — the window, the volume and the component breakdown — and each with the same sentence attached: scores are computed from media coverage volume, sources and sentiment; they describe coverage, not character, and must not be used for decisions about individuals.
Profile private people
A person is scored and shown only in a public role — an office, a public company, a byline, a stage. Someone named in the news without such a role is not profiled, and minors never are. Anyone can ask for a review.
Make decisions about people
Quantiveo is not a consumer reporting agency. Its scores and assessments must not be used to decide credit, employment, housing, insurance or any eligibility. The acceptable use policy forbids it.
Keep articles
Article text is held for the short window analysis needs and then removed; what remains is an excerpt, the metadata and the derived measurements. Publisher images are stored for analysis, never displayed.
Fill gaps with fiction
A source without enough evidence has no score. A claim that has not been assessed says so. Nothing on the wire is estimated to look complete.
The binding statements are the privacy policy, the acceptable use policy and the crawler page.
When we are wrong.
A complaint about a claim goes through the privacy request form. The assessment is neutralised on every surface while it is reviewed, and we reply within two days. A request about a profile, a portrait or a score is answered within thirty. A suppression is applied everywhere, including the API and cached pages, and checked again daily.
Quantiveo (2026). Methodology. quantiveo.com/methodology
@misc{quantiveo2026methodology,
title = {Methodology},
author = {Quantiveo},
year = {2026},
url = {https://quantiveo.com/methodology}
}