How this works

We work only from public, on-the-record sources: published news reports, official press releases, parliamentary records, broadcast interviews, and official social media accounts. Every statement we record is stored with a link to its original source and the date it was made. We do not use private communications, leaked material, or anonymous tips as the basis for any published claim.

1. Ingestion. We collect public statements from monitored sources and store the original text, the date, and the source link.
2. Translation. Where a statement was made in Sinhala or Tamil, we keep the original and add an English translation, clearly labelled.
3. Automated analysis. Software compares statements by the same person over time to flag possible contradictions. This step produces candidates, not conclusions.
4. Human review. A person reviews each candidate against the original sources. Weak, ambiguous, or out-of-context items are rejected here.
5. Publication. Only approved items appear on the public site, each shown alongside the underlying statements and their source links so you can read the originals yourself.

No contradiction reaches the public page on automated analysis alone. That gate is enforced in the query itself, not by editorial promise: the public list and the public count both read only approved rows, so a candidate still awaiting review cannot appear on the site or be counted in a headline figure. A substantial share of detected candidates is sitting unreviewed at any given time, and those are invisible to you by design.

The Promise Tracker works differently, and you should know that before you read a verdict. Promise statuses are assigned by AI and published without a human approving each one. There is no equivalent of the contradiction review gate. What stands in its place is a set of structural constraints, plus a visible appeal route on every verdict.

• Two separate calls, not one. One model assembles an evidence file from statements and official documents, labelling each item as supporting the promise or cutting against it. A second model, seeing only that fixed file, assigns the status. A single model that both gathers and rules would label its own conclusion as corroboration.
• No evidence, no verdict.A promise with no evidence attached stays “not yet rated” rather than being rated on the model’s general impression.
• A politician’s own claim is not proof. A promise whose only supporting evidence is the politician or party saying they delivered it cannot be rated “kept”. Otherwise a press release launders itself into a fact.
• Every guard only weakens a verdict. None of these rules can make a verdict stronger or more damning, only hold one back.

Every rated promise shows the date it was assessed, states plainly that it was not editorially reviewed, and carries a link to dispute it. If you think a verdict is wrong, that link is the intended response and we would rather hear it than not.

The Accountability Tracker consolidates major incidents, controversies and delivered outcomes from published reporting, official statements and parliamentary material. Its discipline is a strict separation between a bad outcome, a government handling failure, an allegation, and proven misconduct. An allegation, arrest or indictment is never treated as guilt, and positive outcomes sit in the same record as failures.

Records without a citable source link are held out of publication, enforced in the database rather than by editorial promise, so gaps in the public reference numbering are normal. The full label definitions, the inherited-problems rule, and the takedown policy are set out in how the labels work, and the relationships between records are shown in connections and patterns.

When we show two statements side by side, we present the primary sources first and our analysis second. The analysis is our interpretation of how the statements relate. Reasonable people may read the same two statements differently, and context matters. A politician may have changed position for good reason, or may have been describing different situations. We encourage you to read the linked originals and reach your own view. Our commentary is opinion and analysis based on disclosed facts, not a statement of proven bad faith.

• Same-source pairs. Two sentences from the same interview or article are usually context, not a contradiction. We screen these out.
• Context and timing. We show dates and sources so a shift in position over time is visible and verifiable.
• Attribution.A national statistic or a quote about someone else is not the same as a person’s own commitment.
• Tone. We aim to describe what the record shows, not to assign motive beyond what the sources support.

This is an evolving system and it is not infallible. Automated detection can produce false positives; translation can lose nuance; sources can themselves be wrong or incomplete. Promise verdicts in particular carry no human sign-off, which is why they are shown with their evidence, their assessment date, and a dispute link rather than as settled fact. That is also why we publish our sources and maintain an open corrections process. If you believe something we have published is inaccurate or unfair, please tell us: see our Corrections & Right of Reply page.