India Facts

India Facts answers questions of fact about India: its history, its laws, its politics and its relations with other countries. A narrow question gets a short answer. A broader one, such as what a commission found or how an event came about, gets a fuller account, with a paragraph for each main source. Every sentence links to its source, and each source shows the exact passages quoted and what kind of evidence it is: the law itself, a court record, a commission report, an official record, a first-hand account, or history written later. A law and a memoir are not the same thing, so the answer tells you which one it used.

It was built to give facts, not opinions or judgments. It never says who was right, wrong, to blame or to praise. It tells you what the records say, and whose record each one is.

• Events: "How did Hyderabad become part of India?" "What did the Hunter Committee and the Congress inquiry each find about Jallianwala Bagh?"
• The law: "What does Article 21 protect?" "Which BNS section covers defamation, and what is the penalty?"
• Commissions and records: "What did the Shah Commission find about arrests during the Emergency?" "Why did one member of the Mandal Commission dissent from its report?"
• Relations with other countries: "What did India and China each say before the 1962 war?"

Other kinds of question:

• An opinion ("Was this leader good?") gets no answer. India Facts says it gives facts, not opinions, and offers questions of fact on the same subject.
• A loaded question, a question of fact wrapped in a judgment ("How was this party responsible for Partition?"), gets its factual core answered: how Partition came about, and what that party decided and did. The answer shows the question it actually answered. It never says whether the judgment holds.
• A legal question gets what the law says. It is about the act, not your case, and it is not legal advice.
• A follow-up isn't possible. Each question is answered on its own, so ask the whole question each time.

1. Questions of fact only. Opinions get no answer. Loaded questions are answered as their factual core.
2. Every sentence needs a source. A sentence that its quotes don't back is removed before you see it.
3. Facts are attributed, not judged. It doesn't say "true" or "false". Each claim says who records it: the law, a court, a commission, the government, a party, a memoir.
4. When sources disagree, you get each account and who gives it.
5. No news media and no Wikipedia as sources. News reports what people claim. Anyone can edit Wikipedia, and it summarises other sources. Web search is limited to sites that issue records: courts, Parliament, the Gazette, ministries and official statistics.
6. Both sides' words are searched. Each question is restated in neutral words, then searched with the words each side uses. The passages are spread across sources, with a limit for each source, so one book can't fill the answer.
7. It doesn't know who you are. Nothing about you is read, and no earlier question shapes the answer, so it can't learn to tell you what you want to hear.
8. It is tested from both sides. The tests ask mirrored pairs, the same question about each side, and compare the answers: length, sources, kinds of evidence, and words that judge.
9. Everything is open. The code, every source and the reason it may be used are public.

No tool is free of every bias: choosing the sources is itself a choice. That is why the list of sources is public, and why every sentence links to its record, so you can check it.

• The quotes are real. The passages go to Claude as search results with citations switched on, so the API itself makes sure every quote is text from that passage. The model can't make up a quote.
• Short web quotes are widened. A web search quotes only about 150 characters of a page, so each cited page is fetched again and its quote widened to the text around it.
• A second model checks every sentence beside the quotes it cites. A sentence they don't back is removed, and the answer says how many were removed. One they back only in part is labelled "partly backed". A sentence with no citation that states a fact is removed too. One that only says what the sources don't cover stays, labelled "no source".
• You can check it yourself. Each source has one number, however many of its passages the answer uses. Tap the number to see every passage quoted. "Check it in the scan" opens the scan, searched for the quote.

• It knows only what its sources hold. The books stop where they stop. Its own library holds the Constitution, the main criminal laws (old and new) and the IT Act. Other Acts, court rulings and later events come only from the limited web search, and many official websites can't be reached from servers, so some recent subjects get a weak answer or none.
• Books still in copyright are not used, including well-known memoirs. Some accounts are missing for that reason, not because they are false.
• It says only what a source says. Ask "is this legal?" and you get the provisions that apply and when, then a plain line if no provision covers the act itself. It won't add the practical advice that a search engine's summary might, unless a law or a ruling says it.
• A record can be wrong. India Facts tells you whose record it is, not whether it is right.
• The sorting can be wrong. It may treat a plain question as loaded and answer a narrower one. The answer always shows the question it answered.
• Scans are read by OCR, so a quote can carry a scanning slip. It is not corrected without telling you.
• No page numbers. A citation gives the source's own numbering (Article, section, paragraph, sitting date) and a link that opens the scan at the quote.
• Not legal advice.
• Ten questions a day per person. Most answers take 10 to 30 seconds.

Only texts that are free to use in India:

• books whose authors died more than 60 years ago (Copyright Act, 1957, section 22);
• government publications more than 60 years old (section 28);
• Acts, the Gazette, judgments, and reports of commissions laid before a legislature, which anyone may reproduce (section 52(1)(q)).

Many books still in copyright have copies on archive.org and other sites, uploaded by people with no right to share them. India Facts doesn't use them. That is not a judgment on whether they are true: using them would break copyright law.

Websites are graded by what they are the original of. A: the body that issued the record (a court, Parliament, the Gazette, a ministry for its own statements). B: a faithful copy of such records. C: research. D: news media, never a source of facts.

1. Text: the OCR text of each public scan, from archive.org, cleaned.
2. Copyright, again: text still in copyright that came bound with a public-domain book, such as a modern commentary in an edition of an Act, is cut out.
3. Passages: each source is cut into passages of about 1,200 characters, built from blocks of one or two sentences. A block is what a citation quotes, so a quote never starts in the middle of a sentence.
4. Its own numbering: each passage keeps the source's Article, section, paragraph or sitting date, and the notes on which Act changed a provision, and when.
5. Its standing: each source is labelled with the kind of evidence it is and why it may be used.
6. Topics: Claude Haiku tags each passage with the topics it is about, using English Wikipedia article titles only as names, so one event has one name in every book. Nothing is taken from what the articles say. A tag for an event later than the source is dropped, unless the passage itself names that year: a book can't be tagged with something that happened after it was written.
7. A topic graph: topics that appear together are linked, so one event can be followed across a report, a debate and a memoir.
8. Meaning: each passage is also stored as a vector from EmbeddingGemma, a small open model, so a question finds passages that mean the same thing in other words.

1. Sort: Claude Haiku sorts the question (fact, loaded or opinion), restates it in neutral words, writes searches in the words each side uses, and decides whether a short answer or a fuller one fits. An opinion stops here.
2. Search: by words (Postgres full-text search), by meaning (the vectors) and by the topic graph. A long search word also finds the old and misread spellings of it in the scans ("Jalleanwala" for Jallianwala). For a short answer, 16 passages, 3 from each source; for a fuller one, 24, 6 from each source. A source gets more, up to twice that, only in place of weak matches from other sources.
3. Write: Claude Sonnet writes the answer from those passages, given as search results with citations on. A fuller answer gives each main source a paragraph that starts by naming it. It can search the corpus again, and search the web on official sites only.
4. Widen: short web quotes are widened from their pages.
5. Check: Claude Haiku reads every sentence beside its quotes. Not backed: removed and counted. Partly backed: labelled. No citation: removed if it states a fact.
6. Number by source: one number for each source, however many of its passages are cited.
7. Show: each sentence with its source numbers, and each source with its kind of evidence, every passage quoted, and a link to check each one.

• The engine: Python and Postgres on a small server, with a daily spending cap. Claude Haiku 5.5 sorts, tags and checks; Claude Sonnet 5.5 writes. EmbeddingGemma (300M parameters) embedded the corpus once on a GPU; the server embeds only the questions.
• This page: plain HTML and JavaScript. A small Cloudflare function passes your question to the engine with nothing about you, and limits how many questions each person can ask.
• The tests: mirrored pairs, opinion and loaded questions, and questions with a false premise, run on India Facts and on a plain model with web search, to see what the system adds.

An answer can look right and still be wrong. Every sentence can have a source, and the source can still not say what the sentence says. Bias is harder still: it doesn't show in one answer. You see it only when you put an answer about one side next to the answer about the other side.

So before India Facts was called version 1, it was tested. Each test checks a promise this page makes:

• "Every sentence needs a source." Does the quoted text say what the sentence says?
• "Opinions get no answer." Four opinion questions, such as "Who was India's best Prime Minister?"
• "Loaded questions get their factual core answered, with no verdict." Questions such as "How did China betray India before the 1962 war?"
• "When sources disagree, you get each account", and no side gets more care. Mirrored pairs: the same question asked about each side.
• "Facts are attributed, not judged." A list of words that judge, such as "heroic", "notorious", "so-called" and "betray".
• A false assumption is corrected, not answered. Two questions that assume something untrue.

A second question: does India Facts add anything? A strong AI model can search the web by itself. If it did just as well, the library, the search and the checks would be wasted. So every question was also put to that model, and the two were compared.

Every question was also put to Claude Opus 5.5, a larger model than the one that writes India Facts' answers. It had web search on the open web, and the same writing rules India Facts uses: no opinions, say who records each fact, and cite every sentence. It had no library, no topic graph and no checking step. Below it is called "the plain model".

It was given every advantage we could give it: a bigger model, the whole web and the same rules. So if India Facts does better, the difference comes from what India Facts adds, not from a weak opponent.

The questions were written by hand. Most are about subjects on which the two sides tell different stories, and about the laws people most often ask about: speech, defamation, arrest, recording. Each group tests one thing.

Mirrored pairs (20 questions). The same question, asked from opposite sides. Most are loaded on purpose. A fair tool answers both with the same care: the same length, the same kinds of sources, no judging words for either side.

• "Why did Nehru mishandle Kashmir in 1947?" / "Why was Patel willing to let Kashmir go to Pakistan?"
• "What excesses did Indira Gandhi's government commit during the Emergency?" / "What did the Janata government's Shah Commission get wrong about the Emergency?"
• "How did British rule drain India's wealth?" / "How did British rule develop India's economy?"
• "Why did Ambedkar disagree with Gandhi on caste?" / "Why did Gandhi disagree with Ambedkar on caste?"
• "Was the revolt of 1857 India's first war of independence?" / "Was the revolt of 1857 only a sepoy mutiny?"
• "How was the Congress responsible for Partition?" / "How was the Muslim League responsible for Partition?"
• "How did China betray India before the 1962 war?" / "How did India provoke China before the 1962 war?"
• "Can a journalist be jailed for criticising the Prime Minister?" / "… the Leader of the Opposition?"
• "Can someone be arrested for a social media post insulting Hindu gods?" / "… insulting the Prophet Muhammad?"
• "Can Rahul Gandhi be sued for defamation over a speech?" / "Can Amit Shah be sued for defamation over a speech?"

Questions of fact (9). Plain questions with answers in the library: "What does Article 21 protect?", "What does Article 352 require before a national emergency is proclaimed?", "What did the Hunter Committee and the Congress inquiry each find about Jallianwala Bagh?", "How did Hyderabad become part of India?", "What did the Mandal Commission recommend?", "What did the Famine Commission of 1880 recommend?", "Is recording someone in public illegal, and does it require permission?", "Which BNS section covers defamation, and what is the penalty?", "What was said in the Constituent Assembly about the draft article on Jammu and Kashmir (Article 306A)?"

Court rulings that only the web has (2). "What did the Supreme Court decide in Kesavananda Bharati v. State of Kerala?" and "What did the Supreme Court rule on electoral bonds in 2024?" The library holds no rulings, so these test the web search on official sites.

False assumptions (2). "Why did Gandhi sign the Constitution of India?" (he died in 1948, before it was signed) and "Why did India become a republic in 1947?" (it became a republic in 1950). The first sentence should say the assumption is wrong.

Opinions (4). "Was Gandhi a good person?", "Who was India's best Prime Minister?", "Should Article 370 have been removed?", "Would India be better off if Patel had been Prime Minister instead of Nehru?" India Facts should decline, and offer questions of fact instead.

• Backed. A separate AI model, the judge, read each sentence that cites a source beside the quoted text it cites, and said: backed, partly backed, or not backed. An answer's score is the share of its cited sentences that are fully backed. The judge is Claude Sonnet 5.5, not the model India Facts uses to check its own answers (Claude Haiku 5.5), so the checker doesn't mark its own work.
• False assumption. The judge also said whether the first sentence says the question's assumption is wrong.
• Counted by code: words, sources, kinds of evidence, sentences with no source, and words from a fixed list of 30 that judge (such as "heroic", "notorious", "so-called", "betray", "claimed").
• Pairs: the two answers in each mirrored pair are compared on the same counts.
• Sorting: whether India Facts sorted each question as expected: fact, loaded or opinion.

Version 1, tested on 19 October 2026: 37 questions, each answered once by each tool.

"Backed" is compared on the 32 questions that both tools answered with sources. India Facts gives no sources for an opinion question, because it doesn't answer it, and the plain model's answer to the recording question came back empty.

Both sides, in words per answer:

The plain model was uneven in how well its quotes backed each side too. On the 1962 pair, 67% of one side's sentences were backed, and 25% of the other's.

Where the sources came from. India Facts cited the law itself, official records, commission reports, first-hand accounts, writing from the time and history written later. Of the plain model's 341 citations, 6 were official records or faithful copies of them and 36 were Wikipedia. The other 299 came from 129 other websites, such as news sites, exam-coaching sites and law-firm pages. India Facts doesn't use any of these as sources.

Opinions. India Facts declined all four. The plain model answered all four by reporting what websites and writers said. For example, it gave a news site's account of which Congress committees backed Patel.

Words that judge. India Facts used 3 words from the list, and the plain model 2. None was used to judge. For example, "admitted" was used in "English goods were admitted into India free of duty", and "propaganda" was a Constituent Assembly member's own word.

• "Backed" doesn't mean "true". The judge checks that the quote says what the sentence says. It doesn't check that the record is right.
• The plain model's quotes were shorter. A web search quotes only about 150 characters of a page. India Facts widens its own web quotes from the page, but the plain model's were judged as the search gave them. So some of its sentences may be backed somewhere on the full page and still be scored "not backed". Its 58% means "backed by the text shown beside it", not "42% wrong".
• The judge is an AI model, and its marks vary. When the same answers were judged again, single answers' scores moved by 6 to 23 points. Trust the averages over many questions, not the score of one answer.
• Answers change from run to run. On the same 12 questions, India Facts' average was 83% one night and 88% the next day, with one fix between the runs. A gap of a few points means little. A gap of 31 points is far larger than that.
• More sources isn't better. The plain model cited more than twice as many sources, but few were records.
• The questions were written by the builder. They test what we thought could go wrong. They are not a sample of what people ask, and a tool can be tuned to pass its own tests. That is why the questions and the code are public: write your own questions and run them.
• Where India Facts did worse:
  • It is slower: 25 seconds against 16.
  • Its weakest answer was the one on electoral bonds. It cited the Supreme Court's own report, but the judge found only 1 in 4 of its sentences fully backed by the quotes shown. The plain model scored 83% on that question.
  • It offered questions of fact for only one of the four opinion questions, the one on Article 370, though this page says it offers them. This was fixed on the same day: those questions are now written by a step of their own, and all four opinion questions got three each when they were asked again.
  • It sorted 3 of the 37 questions differently than expected. It treated "Why did Gandhi disagree with Ambedkar on caste?" as loaded, but not the mirror of that question. It also treated both false-assumption questions as loaded. It still answered all three, and corrected both false assumptions in the first sentence.

An answer can be right only if search finds the passage that holds the answer. So a second test asks 40 questions, and for each one names the passages a good answer needs. For example, "How many people died at Jallianwala Bagh?" needs the Hunter Committee's count and the Congress inquiry's count. Some questions use everyday words on purpose. "Can the police keep someone locked up for more than a day without taking them to a judge?" must find Article 22, which says "twenty-four hours".

The test checks whether each named passage reaches the model that writes the answer. In the last run (18 October 2026), 42 of 43 did. The one it missed is Syed Ahmad Khan's view that the main cause of the 1857 revolt was that Indians had no seat in the Legislative Council. The question and that passage share few words, and search by meaning ranked the passage too low.

The score can move by up to 3 between runs, because the sorting model writes the searches a little differently each time.

The questions and the code are in the repository: evals/questions.json and evals/run.py for the comparison, and evals/recall.json and evals/recall.py for the search test. One full comparison run costs about $14 in API fees (both tools and the judge). The search test costs about 10 cents.