How the numbers are made
Every figure on this site comes from the steps below. The same rubric, the same credit rules and the same checks apply to every party and every era.
Each action is scored on five criteria, from −5 to +5 each. Each criterion is an integer from -5 (clearly harmful on this criterion) to +5 (clearly beneficial). 0 means no net effect or no evidence either way.
An action's points are the sum of its five criterion scores, so they range from −25 to +25. A negative number means the evidence shows net harm on that criterion. A positive number means net benefit. Zero means no net effect, or no evidence either way.
How many people did the action materially affect?
Scores between the anchors are used when the evidence sits between two descriptions.
Do official or independent data show the intended result (or its opposite), compared with the era baseline and peers?
Scores between the anchors are used when the evidence sits between two descriptions.
Is the fiscal cost proportionate, and is the effect durable?
Scores between the anchors are used when the evidence sits between two descriptions.
Was it made with debate, consultation and legal soundness, and did it respect rights and court rulings?
Scores between the anchors are used when the evidence sits between two descriptions.
Was it implemented as announced, on time and at scale?
Scores between the anchors are used when the evidence sits between two descriptions.
Two independent scorers rate each action blind to party. Final criterion score = mean of scorers. A criterion where scorers differ by 2 or more is flagged for review. The range shown on the site uses the lowest and highest scorer totals.
Before scoring, each action is stripped of party names and of the name of the government. Scorers see the facts, the stated reasons, the criticism, the benefits and concerns, the evidence and the era baseline. They work separately and do not see each other's scores. Both sets of scores, and each scorer's note, are published on the action's page.
Scores are AI-assisted. Entries are drafted by AI models and both sets of scores are produced by separate AI scorer sessions working from the public rubric. A person reviews each batch of changes before it is published, but the numbers are not the product of a panel of human experts.
Blind to party, not blind to content. The scorers are not told which party or government acted, but they do read the facts text, and that text can reveal the era, a minister's name or a government indirectly. Blinding therefore reduces party bias; it does not remove it. A separate neutrality review checks the wording of every entry for parity between parties, and its findings are in the repository.
Every entry lists its sources with the publisher and a type: primary, official data, court, reputable media or academic. Source quality varies. Some entries rest mainly on official documents and court judgments; others rest on secondary summaries, encyclopaedia pages and press reports because the primary text was not opened. Entries with no primary source, or with notes about unopened documents, carry a low confidence label. The source-quality list shows which entries to verify first.
Government start and end dates were checked against Wikipedia lists of office-holders; a pass against primary sources (Gazette, Election Commission) is still pending. Seat-share figures, which set how credit and years are split inside coalitions, are approximate for several governments: some small-party seat counts come from general knowledge or partial tables and are flagged as unverified. See the data notes. Shares for the parties in a coalition should not be read as exact.
The government in office when the action was implemented earns the full score. A different government that originated or legislated the idea earns 25% of the same score. Coalition governments split credit by each partner's share.
The same rule is applied whichever party holds office. Positive and negative points are credited in the same way, so a harmful action lowers the credited government's total as much as a beneficial one raises it.
Some decisions are made by an independent institution, such as a court, a regulator or the Election Commission, and not by a party government. These entries carry credit: none and the badge Independent institution: no party credit.
There are currently 1 such entry.
A dossier is a fact file on an institution, organisation, party, company group or live topic. It has a dated timeline, a claims table, the positions of each holder, measured data, a list of what the facts do not show, and numbered sources. Dossiers are not scored and earn no party points. Each shows the date its facts were last checked and says that later developments may exist. A dossier states what is on record; it does not conclude who is right, and a listed fact or an allegation is not a finding unless its type says so.
As-of date and how sources were read. Every fact in a dossier is as of its "as of" date. Many sources were seen only in search results or summaries and were not opened in full; the source title says so (for example "seen in search results, not opened"). Such a source is typed as reputable media even when the underlying document is official, and a status taken from it is "as reported". Where only a year or month was documented, the date is shown as a year or month and not as a precise day.
Every timeline item and every row of a claims table is labelled with one of these types. The label is shown in words, with a symbol and a border pattern, so it never depends on colour.
Credit for a coalition government is divided among its partners by each partner's share of the governing coalition's Lok Sabha seats at formation. The same shares are used to divide the years in power. Parties that supported a government from outside, without joining it, receive no credit and no years, and are listed on that government's page.
Caretaker governments are included and counted by days in office, like any other.
The table used to lead with points per year in power. That figure is the product of two things: how many actions were entered per year, and how many points each action earned. The first is a choice made by the maintainers about which actions to include, not a fact about a party's record. A party in a period with many entries shows more points per year than one in a period with few, even if each entry scored the same. In the data at the time of the change, BJP had about 2.7 times as many entries per year in power as INC, and half the points per entry; its higher points-per-year figure came from the entry count, not from higher-scoring entries.
So the headline is now the average points per action, which does not depend on how many entries there are per year. Two further figures sit beside it: points per year (shrunk), which keeps the idea of a rate but is pulled toward zero for short tenures, and the plain total. A small control above the table switches the sort and the highlighted figure between the first two. The raw points-per-year figure is still used in the category grid and on party pages, where each is labelled as depending on how many entries are included.
These numbers are recomputed from the published data each time the site is built.
What these figures still do not do. The average per action treats a large reform and a small scheme as equal items, and it depends on which actions were chosen: adding several minor entries lowers an average, adding several major ones raises it. For a coalition partner, credit is multiplied by its seat share, so the average for a small partner is small for that reason; such parties have under 2 years in power and are listed apart. The shrunk figure is not a measured rate; the constant 5 is a choice.
This describes the raw per-year figure, kept for the category grid and party pages. A party's total is the sum of its credited points. Its years in power are the sum, over every government it belonged to, of (days in office divided by 365.25) multiplied by its coalition share. Points per year is the total divided by years in power. It exists so that a party that governed for forty years is not compared with one that governed for five on raw totals alone.
Short tenures. A party with under 2 years in power can show a very large or very small figure from a single action, because the divisor is tiny. Such parties are listed in a separate group below the main table and are left out of the category grid.
This is a summary of the evidence entered so far. It is not a verdict on any party, and the order of the table follows a figure, not a judgement. Parties with short terms or few scored actions produce wide and unstable figures, which is why ranges and “low data” labels are shown with them. A party's total also depends on how many of its actions have been entered, which is a selection and not a census, so a higher figure can reflect denser coverage of a period as well as stronger results.
Scorers judge each action against what was normal at the time, not against today. Each entry states a baseline in one sentence: what growth, peer-country performance or the previous decade looked like. The Outcome criterion asks whether results beat that baseline and peers. Because the economy, the state and the data available differ widely between 1950 and 2020, the baseline is part of every entry.
Some actions show their effects years later. Each entry gives a lag in years before outcomes could reasonably show. Scorers weigh outcome evidence only up to what that lag allows, and recent actions inside their lag period are scored with low confidence. Outcomes are attributed to the government that implemented the action, even when they appear years later under a different government; the lag is stated so that readers can judge how much of an outcome the implementing government could have affected. Only the originator rule above gives credit to any other government.
Ranges between parties often overlap. When they do, the data do not separate those parties.
A party–category cell with fewer than 2 scored actions is labelled “low data” and drawn with a hatch in the grid. Category pages show how many entries exist for each party and decade, so gaps in coverage are visible and are not mistaken for gaps in activity. The aim is the same depth for every party and era; where that is not yet true, the page says so.
The home page lets you set a weight from 0 to 3 for each of the 24 categories. A weight multiplies that category's points and range before they are added up; years in power do not change. The page recomputes in your browser with the same formulas used at build time, and nothing is stored or sent. You can also limit the table to a period, such as before or after 1991; the period applies to the date of each action and to years in power.
These are known weaknesses in the method. They were found in an independent neutrality review, whose full text is in the repository (docs/NEUTRALITY-REVIEW.md). None has been fixed in the scores yet.