How Lenstrum Works

Lenstrum extracts text from a URL and passes it to Claude (Anthropic) using a structured non-partisan prompt. The prompt first asks Claude to determine whether the content is political or non-political, then rates it accordingly: political content is rated Left, Centre, or Right; non-political content — theological, scientific, economic, philosophical, or otherwise ideological debates that aren't about party politics — is labelled with the two or three perspectives actually at play in that debate, so a framing axis is never forced into a Left/Centre/Right frame it doesn't belong in. Every rating comes with key framing choices, perspectives absent from the coverage, and a confidence level (High, Medium, Low). Lenstrum analyzes individual pieces of content — not outlet reputations. The same outlet can produce Centre-rated and Right-rated content on the same day, depending on the specific piece.

The AI Involvement tab estimates the probability that a piece of content was generated or significantly assisted by AI. Lenstrum looks for structural signals common in AI-generated text: consistent paragraph length, explicit transitional phrases, vocabulary patterns, and a particular kind of fluency that differs from human writing. For video content, synthetic voiceover patterns are also flagged. Results are expressed as a probability score (0–100%) and always framed as analysis, not verdict — AI detection is an inexact science and Lenstrum treats it accordingly.

The Source Context tab draws on Brave Search to find related coverage of the same story from across the political spectrum. For known publishers, Lenstrum references AllSides and Ad Fontes Media bias ratings. For independent creators and social accounts, Lenstrum classifies the account type based on content patterns and stated editorial position. Source context is designed to show you what else is being said about the same story — not to endorse any outlet.

For Pro analyses, Lenstrum also profiles how a piece is built, placing it on up to five of the spectrums below, each on a five-step scale, with the evidence behind each reading. Neither end of a spectrum is good or bad. This describes how a piece is built, not whether it's trustworthy.

These spectrums are in beta. They are not yet validated against the accuracy corpus described below, so treat them as a starting point.

Reporting ↔ Opinion

Reporting: Describes events and attributes claims to sources. Opinion: Argues a view, judges, or recommends action. We look at: Attribution, evaluative language, calls to action, and whether interpretation is labelled as such.

Measured ↔ Emotionally charged

Measured: Calm, precise wording. Emotionally charged: Loaded words, urgency, outrage, fear or exaggeration. We look at: Word intensity, emotional appeals versus evidence, and emphasis devices.

Well-sourced ↔ Thinly sourced

Well-sourced: Named, checkable sources, documents or data. Thinly sourced: Vague ("experts say"), anonymous, or no sources. We look at: The number and specificity of sources and how claims are attributed.

Editorial ↔ Commercial

Editorial: Independent content. Commercial: Sponsored content, affiliate links or product promotion. We look at: Disclosures, promo codes and product focus — only signals visible in the content itself.

Established ↔ Speculative

Established: Claims presented with support as settled. Speculative: Predictions, rumours or unverified claims. We look at: Hedging, the evidence offered, and prediction language.

Individual ↔ Systemic

Individual: The issue is explained through specific people or stories. Systemic: It is explained through causes, policies or patterns. We look at: Whether the issue is framed as one case or a broader trend.

Nuanced ↔ Absolutist

Nuanced: Acknowledges trade-offs and uncertainty. Absolutist: All-or-nothing, "always/never", us-versus-them. We look at: Qualifiers and acknowledgement of other views.

Data-driven ↔ Anecdotal

Data-driven: Statistics, studies or measurements. Anecdotal: Personal stories or single examples. We look at: The type of evidence relied on.

Context-rich ↔ Context-light

Context-rich: Background and what led here. Context-light: An isolated claim or clip without background. We look at: The presence of history, timeline and prior events.

Lenstrum's analysis engine was tested against a 100-item validation corpus before public launch. The corpus spans five categories: political articles (40 items), political YouTube videos (25 items), non-political YouTube (15 items), non-political articles (10 items), and centre/balanced sources (10 items). Results were validated against AllSides and Ad Fontes Media ratings where available, and against internal reference ratings for content types with no established third-party benchmark.

Lenstrum performs strongest on political content — 95% accuracy on political articles and 100% on political YouTube, validated against AllSides and Ad Fontes. Non-political content — theology, economics, philosophy, and science — is scored on a 25-item sample (15 YouTube videos, 10 articles) within our validation corpus, since no published third-party benchmark exists for these categories; results are checked against internal reference ratings only. Across that sample, Lenstrum currently scores in the 73–80% range. These are v1 baseline figures, not a settled accuracy claim — we don't yet have a per-category breakdown within that range, and we'll publish an updated figure as the sample grows. High-confidence outputs match at 92.9% — when Lenstrum is sure, it is right the vast majority of the time.

The corpus will be expanded to 200 items post-launch using real URLs submitted by users. A user feedback loop — thumbs up/down on every result — will be incorporated as an additional validation signal. The corpus will be re-run after any significant prompt change. We will publish updated accuracy figures here as the methodology evolves.