How we make recaps

CognioNews looks for the knowledge behind the news: what people already knew, what records preserve, and how those ideas help with a present-day problem.

A recap connects that knowledge to the published reporting. It names its sources, credits the people involved and explains the limits of what has been reported.

From source to recap

• We work from published reporting and research. Each recap links to its sources and identifies the original outlet or institution.
• We write a CognioNews headline that brings the knowledge connection forward. The publisher’s original headline is shown in the source details.
• We credit people, communities and institutions in the roles supported by the reporting. We do not add identities or endorsements that the source does not support.
• We keep quotations brief and attributed. Illustrative photographs are labelled and have a separate image-credit control.
• Recaps require editorial clearance before publication. Where AI tools assist with preparation, responsibility for what we publish remains with CognioNews. Read our AI stance.

Understanding the Knowledge gauge

The Knowledge gauge reads the evidence in the original article for knowledge being brought into view and put to use. That might be living community knowledge, an archive, earlier research or a connection across different ways of knowing. Knowledge can be alive with its holders even when others overlook it.

The reading runs from 0 to 10. It does not measure how much our recap improved the article. It is not an accuracy rating, a percentage or a verdict on a community’s knowledge. A 10 means the article reaches the top of this rubric; it does not mean every claim is verified or that a project’s benefits will last.

Six questions behind the reading

Each question receives 0, 1 or 2 from the article’s text. A higher value needs stronger, specific evidence in that text.

• Previously overlooked: does it describe knowledge or a problem that was lost, ignored, unread or unresolved for a long time? Weight: 1.5.
• Knowledge already held: does it identify earlier holders, records, practices or research? Weight: 1.5.
• Brought to light: does it report a concrete result, rather than only a thin claim? Weight: 2.
• Put into use: is that knowledge proposed for use, partly used, or actually returned to use and built on? Weight: 2.
• Checked: does it describe no checking, checking limited to the authors or one source, or checks on the ground, by hand or by independent experts? Weight: 1.5.
• Credited: does it identify the holders, and explain their roles? Weight: 1.5.

One extra point recognises a supported connection between two or more kinds of knowledge, with at least one beyond research alone. These can include community practice, lived experience, archives, another discipline or region, and machine-surfaced connections.

For each question, we multiply its value by half its weight, then add the connection point if supported. The result is rounded to the nearest whole number, with halves rounded up, and capped at 10. Without both prior or overlooked knowledge and evidence of surfacing or use, the reading cannot exceed 4. An announcement without the strongest checking evidence cannot exceed 7. If the source explicitly describes the result as unverified, the checking component is zero.

Evidence, review and limits

Knowledge readings are assessed automatically. Two different AI models read the same stored article independently, without seeing each other's answers. They must agree on all six component values, the connection point and the limiting conditions. Software checks their quoted evidence against the article and calculates the total. A separate challenge step checks the proposed assessment for unsupported interpretations and contradictions before it can be published. The challenge uses one of the two models with a different task; it is not an independent human review.

Routine readings that pass these checks can be published automatically for an editorially cleared recap. The explanation identifies the reading as automated and shows its component values, short reasons, source link, assessment date and method version. A person has not reviewed the reading unless we explicitly say so. CognioNews remains responsible for the publication.

Disagreement, unsupported evidence or unavailable source text sends the reading to a private exception queue. We omit its gauge; absence does not mean zero. If a previously checked source temporarily becomes unavailable, its existing reading may remain for up to 72 hours after the last successful source check. A detected source change requires reassessment. General items in In other news carry no Knowledge reading.

This is a new editorial rubric, initially running on a small set of cleared recaps. Agreement between models is a consistency check, not proof that they are right: models can share mistakes. The rubric has not completed independent blind accuracy calibration. Different evidence profiles can reach the same total, especially at the upper limit, so read the component explanations and source limitations alongside the needle. Checking reported evidence does not independently verify every underlying claim, and what an article omits may be invisible to the reading.

Earlier scores used different definitions and are not comparable with this gauge. We retain the method version with each published reading and record changes rather than silently replacing its history.

Sources and corrections

Read the original reporting alongside our recap. You can expand each recap’s source details to see its credits, context and limitations.

If we have missed something or made a mistake, request a correction or credit change. Published updates appear in our Changes list.