How the compass works

Depolaris shows how outlets across the Indian political spectrum frame the same story. This page explains exactly how placements are made — because a rating you can't interrogate is just another opinion.

Why not left–right

A single left–right axis fails in India: parties and outlets mix cultural, economic, and federal positions freely. We use four independent axes, and show only the axes a story genuinely engages. A story that engages none gets no compass at all.

The four axes

• Cultural — Majoritarian ↔ Pluralist: how the piece frames the relationship between the religious-cultural majority and minorities.
• Economic — Statist ↔ Market: framing of the state's role in the economy: welfare, regulation, privatisation.
• Structural — Centralising ↔ Federalist: framing of Union–state relations and where power should sit.
• Institutional — Populist ↔ Institutionalist: framing of mediating institutions — courts, Election Commission, RBI — versus direct popular mandate.

Neither end of any axis is the "bad" end. Scores run −3 to +3 between the two named poles; 0 means the piece takes no discernible position or balances positions.

What we score — and what we never score

We score framing choices in the text: which stakeholders are quoted and which are absent; what the headline foregrounds; loaded vocabulary; what is treated as cause and consequence; editorialising markers. We never score an outlet's ownership, an author's identity, a story's subject (covering a party's event is not endorsing it), or whether conclusions favour any party. Truth of claims is a separate matter from framing and is not what the compass measures.

How a placement is made — and checked

1. Articles covering one story are clustered together.
2. The relevant axes for the story are selected (and can be: none — a story that engages no axis carries no compass).
3. Each article with readable full text is scored per axis, with a written rationale and verbatim evidence recorded internally.
4. An adversarial pass argues for the opposite placement. Unstable placements are flagged, as are all extreme scores and a random sample.
5. Flagged placements face two independent checks: every cited evidence span is machine-verified to exist verbatim in the article (fabricated evidence can never pass), and a second AI model from a different provider blind-rescores the axis. If the two independent models agree within one point and the evidence verifies, the flag resolves as machine consensus, recorded as such. If they disagree, a human editor arbitrates.
6. A placement still awaiting resolution may appear publicly, but always marked contested, with a notice explaining that our own review read the framing differently and the position is provisional. We would rather show you our uncertainty than hide it.

How the edition is ordered

Each story's divergence index measures how far apart outlets' placements sit on its axes — a documented spread, not an editorial vibe. The edition leads with India stories whose framing diverges most, because that is where seeing multiple lenses matters most. Stories where the spectrum substantially agrees are given their own section — agreement is news too, and a depolarisation platform should say so as loudly as it documents division. World stories appear with a stated India stake; procedural and statistical stories appear under "The record".

Tone indicators

Separately from the compass, articles are checked for heightened language: sensationalism, inflammatory phrasing, loaded vocabulary, and emotive appeals, each with a severity from 1 to 3. Tone flags describe how something is said, never whether it is true and never which side it favours — measured prose about a divisive topic earns no flag, and calm-sounding misinformation earns no pass. Hover any tone chip to read the specific reason it was raised.

Beyond the press

Coverage also lists video reporting and commentary from a spectrum-spanning set of YouTube channels. In the current phase, video items are listed for completeness and linked, but not compass-scored — scoring requires text we can verify, and transcript analysis is planned. Video can join a story's coverage but never determines whether something counts as a story, and never supplies a story's title.

Outlets get distributions, not labels

We never assign an outlet a fixed label. Over time we show the distribution of an outlet's article scores per axis, with sample sizes. Outlets drift; distributions capture that.

Limits, honestly stated

• Some outlets block automated reading of their full text; their articles are listed for coverage but not scored. We disclose this per-article.
• Scoring uses AI assistance with layered checks: adversarial review, machine evidence-verification, cross-model consensus, and human arbitration of disagreements. Some flags are resolved by machine consensus rather than a human, and are recorded as such. Errors will still occur; the corrections log is public and placements are versioned.
• This methodology is versioned (currently v0.1) and will change; changes are logged.

Disagree with a placement?

Any outlet or reader can challenge any score via the grievance page. Appeals receive a reasoned response, and outcomes are logged publicly.