How these summaries are written

CivicLens publishes a short daily summary of the news for every Indian district and state. Each summary is written by an AI model from public news headlines and published with links to the articles it came from.

The point of the site is coverage, not scoops. Most Indian districts have no single place that tells you what happened there today. CivicLens tries to be that index — and then get out of the way and send you to the people who did the work.

The whole pipeline is automated and runs once a day. District summaries are live by about 12:30 PM IST and state roll-ups by about 1:00 PM IST.

• Collect. For each district and state we fetch public news headlines from Google News’ RSS search, restricted to recent items — in practice roughly the last 48 hours. We take headlines and article links only. We do not fetch, store or republish article text.
• Filter. Near-duplicates and headlines with no local relevance are dropped before anything reaches the model.
• Summarise. The surviving headlines go to an AI model together with the standards in the next section. The output is structured, not free prose: the main issues, one line on what they add up to, and something a resident could actually do.
• Publish. The summary appears on the district or state page, dated, with its source articles listed and linked.
• Record. Every summary stores whether it was really AI-generated or fell back to a plain template, and which model produced it. A broken model shows up in the published data rather than quietly producing something that looks the same.

These are the rules the summarisation pipeline is given on every run — not a statement of intent written afterwards. They are also the rules we check a summary against when somebody flags it.

• Never present an allegation as an established fact.
• Never infer guilt, corruption, criminal responsibility, motives, political intentions, causation or wrongdoing unless the source explicitly establishes it.
• Label every claim as a confirmed fact, a reported event, an allegation, a claim by an official, a claim by a politician, a point on which sources disagree, or unverified.
• If sources disagree, say so rather than quietly picking the more interesting version.
• If a source is unclear or unreliable, leave the claim out. An omission is better than a guess.
• Never invent names, dates, statistics, quotes or government actions. If it is not in the source material, it does not go in the summary.
• Preserve the uncertainty in the original. “Police are investigating” must not become “police found”. Hedged reporting stays hedged.

If you find a summary that breaks one of these, that is a bug and we want to know. See corrections below.

The reason the rules above are worth anything is that the labels survive into what you read. Claims in a summary carry one of these seven tags, so you can tell at a glance whether something is established or merely said by somebody:

Confirmed

Established as fact by the reporting source.

Reported

Reported as having happened, without further verification here.

Official statement

Stated by a government official or department — a claim, not an independent finding.

Political claim

Stated by a politician or party — a claim, not an independent finding.

Allegation

An allegation. Not established, not proven, and not a finding of wrongdoing.

Sources disagree

Sources reported this differently. Treat with caution.

Unverified

The source was unclear or incomplete. Verify before relying on it.

A missing label is not a claim of certainty. Older summaries, and any summary where the model ignored the instruction, simply go unlabelled — we would rather show nothing than apply a tag we are not confident in.

A summary names a person only where a source named them, and describes them in the role the source gave them. Where a person is connected to a complaint, an allegation or an investigation, the summary has to say that it is a complaint, an allegation or an investigation — and must not describe, imply or hint at guilt.

An investigation is not a conviction. A complaint is not a finding. An opposition leader’s accusation is a political claim, and a department’s denial is an official claim; neither is a fact because it was said loudly.

If you are named in a summary and it is wrong, unfair, or reads as an accusation, tell us and we will take it down while we look at it — not after. That applies whether or not you send a legal notice, and you do not need a lawyer to ask.

Designing against failure is not the same as preventing it. These are the known weaknesses of this approach, stated plainly so you can read the site with them in mind:

• Headlines are thin. We summarise headlines, not full articles. The qualification that lived in paragraph nine is invisible to us.
• Places get confused. Many Indian districts share a name with a city, another district, or a person. A headline can be attached to the wrong one.
• We inherit coverage bias. We can only summarise what was published. A thin summary means thin coverage of that district that day, not a quiet district — and under-covered places stay under-covered here too.
• Recent is not the same as important. The model ranks what was published in the window, not what matters most in the long run.
• Names travel badly. Transliteration differs between publications, so the same person or place can be split into two, or two merged into one.
• Compression loses caveats. Three lines cannot carry everything a 900-word article carried, and what gets dropped is usually the nuance.
• Fluency is not accuracy. Language models write confidently whether or not they are right. A well-written sentence here is not evidence of anything.
• Nothing is independently checked. No human verifies a summary against reality before it is published.

Every summary lists the articles it was built from, with the publisher named and the article linked, so the primary source is always one tap away. Summaries published before we started storing article links show the headline text only — those rows predate the change and were not rewritten after the fact.

We do not host, mirror or reproduce article text, and we claim no ownership of anyone else’s reporting. If a source got it wrong, the summary built from it will be wrong too, which is the main reason we ask you to read the original before relying on anything here.

Publishers who would rather not be summarised or linked can say so — see publishers, copyright and attribution.

Honest version: summaries are published automatically, and nobody reads each one before it goes live. There are more than 700 districts a day and one person behind this site. Anyone claiming otherwise at this scale would be exaggerating.

What a human actually does:

• reads and acts on every flagged summary
• spot-checks summaries against their source articles
• can edit any summary, and an edited summary is recorded as human-edited rather than passed off as the model’s output
• can hide a summary immediately, from any page, at any time
• reviews the daily run’s own health report, so a failing model is caught as a failure instead of quietly degrading

Two routes, and the first one is much faster because it tells us exactly which summary you mean:

• “Flag this summary” — the button under every summary. Pick the closest reason, add a line of detail, and leave an email address if you would like a reply.
• Email hello@theciviclens.tech — better for a pattern across several summaries, a publisher request, or anything legal. Include the page link.

What happens next

• We read the flag and compare the summary against the source articles it links and the standards above.
• If it presents an allegation as a fact, names someone wrongly, or could be defamatory, it comes down first and is reviewed after. We do not leave it up while we think about it.
• Otherwise the outcome is one of three: we correct or rewrite the summary, we hide it, or we leave it and tell you why if you gave us an address to reply to.
• Expect a few days. One person, no support desk, and we would rather answer properly than quickly.

We do not currently publish a public log of corrections. If that changes, it will be linked from this page.