Why trust this.

Every source in the library is rated. Tier 1 is a primary text — the source itself, in print or its faithful translation ( Aryabhatiya, Brāhmasphuṭasiddhānta, Lilavati, palm-leaf manuscripts, stone inscriptions, archaeological surveys). Tier 2 is peer-reviewed scholarship that cites Tier 1. Tier 3 is established encyclopedias that themselves cite the literature. Tier 4 is open web — used only when something at Tier 1–3 corroborates it.

The badge on every claim card tells you which tier the headline quote came from. Most claims you see here trace to Tier 1; the corpus currently has 13 Tier-1 sources.

When we extract a claim from a chunk of text, we keep the exact sentence the chunk used. The quote on the claim card is the sentence — unedited, with original punctuation, in the source’s language of publication. We do not paraphrase to make it sound more readable. If the original sentence won’t carry a thought, we do not publish the thought.

The verbatim-quote check is programmatic. The pipeline takes the asserted quote and substring-matches it against the normalised chunk text. If it does not match, the claim does not move to verified. Fuzzy matching is not allowed; the matcher uses the same Unicode normalisation rules that produced the chunk in the first place.

An LLM extracts candidate claims from a source chunk. A second, independent LLM in adversarial mode receives the candidate plus its quote and looks for reasons to reject it: ambiguous attribution, mistranslation, paraphrase drift, unsupported inference, missing context. Anything the adversary catches goes back to the queue; anything that survives goes to editorial.

An editor reviews every candidate before it is verified. The editor can mark a claim as contested — meaning there is evidence on both sides — and that contested status is visible on the public page.

The pipeline runs an integrity audit every night at 02:00 UTC. For every verified claim, the audit re-pulls the source chunk and re-checks the verbatim quote. A drift — in either direction — sets the claim to audit-failed and surfaces it to the editor in the morning.

We publish the consecutive-clean-days counter in the homepage trust strip. We do not publish a number we cannot defend; if the streak hasn’t yet reached seven days, the strip stays hidden. The corpus currently has 50 verified claims.

Every public claim page has a “Flag this claim” link. A flag must come with a counter-source — either a URL or a free-text citation; we do not accept “this is wrong” without a pointer to where to look. The platform requires at least 50 characters of argument so the editor has something to engage with, and throttles submissions per IP so no one can drown the queue.

An editor reviews every flag. Outcomes are flag invalid(the counter-source doesn’t support the objection), duplicate (already raised), claim updated (we changed the claim), or claim unchanged with note(we considered and stayed put). Every claim page shows the aggregate count of flags it has received and how many led to updates — counts only, never the flag content. That is the platform’s transparency boundary.

We do not let an LLM author a claim. The model extracts and rewrites for readability; an editor decides whether the extraction is faithful. The model never produces the source text. AI-generated photographs of historical figures and events are forbidden; rendering an imagined Bhāskara II is dishonest about what we actually know. Article generation draws only from already-verified claims; nothing new sneaks in under article cover.

We do not run a comment thread on the website. If you disagree, file a flag with a counter-source — we will read it.