How the map and report are built.

What enters the topic map

Ten short answers yield a distribution across twenty broad topic categories. “Other / none of these” contributes no topic evidence. If every answer is neutral, we report insufficient evidence instead of presenting an artificial diversity score. Optional X follows and YouTube subscriptions add distributions inferred from public account or channel descriptions. A follow or subscription is a possible route to information, not proof that an item appeared on screen or was read.

With only informative assessment answers, those answers supply the whole map. With one connection, initial weights are 45% assessment and 55% connection. With two, they are 40% assessment, 30% X and 30% YouTube. If the questionnaire supplies no topic evidence, it receives no weight: one connected source supplies the map alone, or X and YouTube each supply 50%. These are disclosed product assumptions. They are not fitted coefficients or calibrated probabilities. Report feedback is kept separate: “more” and “less” affect future ranking, not the underlying topic map.

What the number means

Topic shares add to 100% within this model. The descriptive diversity number is normalized Shannon entropy: −Σ pᵢ ln(pᵢ) ÷ ln(20), scaled to 0–100. It changes with the category scheme and smoothing. It is not a score of knowledge, openness, media quality or measured exposure. Small differences should not be interpreted as meaningful change.

How five stories are chosen

A shared pool first filters for recent, source-linked candidates. A deterministic ranker combines topic fit, recency, an importance heuristic, a source-quality heuristic, source repetition and overlap with stories already selected. It seeks one piece in each role: familiar, adjacent, bridge, less represented and wildcard. “Adjacent” now requires a meaningful topic that actually neighbors one of the map’s stronger topics; bridge and less represented roles also have minimum topic-mass rules. If the pool cannot fill all five roles, the report is withheld rather than inventing a story or mislabeling a role.

Our source-quality number is a domain-level heuristic, not a factual accuracy probability. AI topic classification and short summaries can be wrong. Links, publication dates, original evidence and independent editorial checks remain necessary. A different viewpoint is only valuable when it has supporting evidence.

How a news card earns approval

Discovery uses a small curated RSS list and, when configured, Exa search. DeepSeek is the default text model. A generated card stays pending until an authorized reviewer opens the sources, checks every claim and number, confirms the date and topic labels, and checks either original primary evidence or independent supporting reporting. The domain score never grants approval. Review decisions, edited text, source links and evidence excerpts are recorded together with a version number.

Each card separates what happened, why it matters, an open question and what remains uncertain. Drafts must provide exact source excerpts for their supported statements. Automatic checks catch missing quotes, incomplete dates, invalid topic labels, excessive length and long sentences. These are editing aids, not proof of truth or a validated measure of comprehension. Source excerpts can omit important context, and exact quotes can still be misinterpreted; the reviewer must check their meaning and the original source.

Reports use only approved, recent cards and avoid repeating the same URL or story cluster. A withdrawn card is hidden in the report history, and its pending email is blocked. Already delivered email cannot be recalled. Live model quality and reader comprehension still require evaluation with real sources and independent reviewers; simulated tests do not establish those outcomes.

What the first live pilot found

On 10 October 2026, an engineering pilot froze fifty source excerpts from thirty-nine domains before generation. Eighteen summaries passed the then-current automatic editing checks; seventeen also had valid topic-weight output. These are format and editing results, not factual accuracy rates. Reviewing the accepted drafts against their excerpts still exposed incomplete evidence mappings, numerical scope problems, research outcome wording and overly broad limitations.

The revised draft rules require a separate evidence statement for each summary sentence and matching numerals in its quote. Research cards must preserve the study design and what was actually measured. All cards still require source and editorial approval. This pilot did not complete independent human labeling, reader comprehension testing or validation of the information map, and its findings do not establish quality across other weeks or populations.

What research supports—and does not

Research separates social connections, feed ranking and clicks as different stages of exposure. Survey answers can differ substantially from digital traces, and people differ in how much challenging content they want. Diversity can be a deliberate design goal for news recommenders, but there is no single universally correct mix. These findings support measuring signals separately and giving users choice; they do not validate OffFeed’s current numerical weights.

• Bakshy, Messing & Adamic (2015): network, feed and click exposure were measured separately.
• Cardenal et al. (2022): self-reported news exposure was compared with web tracking.
• Helberger, Karppinen & D’Acunto (2018): opportunities and limits of diversity-sensitive recommendation.
• Vrijenhoek et al. (2021): news recommendation diversity needs explicit, multidimensional evaluation.

Validation required before stronger claims

We need consent-based reading diaries or comparable observed data to test whether topic shares match what people actually encounter, with results reported by source type and subgroup. Human raters should independently label topic, evidence quality, geographic coverage and whether each of the five roles fits; disagreements must be measured. We should compare the ranker with simple baselines on usefulness, novelty, source diversity and factual errors, and test whether people find the reports useful over several weeks. The dataset, metrics, thresholds and failure cases should be documented before changing weights to fit results. Until then, OffFeed is a transparent, testable hypothesis, not a validated diagnostic instrument.

We do not infer political ideology, religion, health status, intelligence or identity from this map. Start with optional connections →