PUBLIC-RECORD INTELLIGENCE
Claims are easy.
The record is harder.
ClaimWatch discovers public speeches, videos and reporting, extracts checkable claims, remembers election commitments, and builds evidence packets from inspectable sources.
AUTONOMOUS DISCOVERY
What ClaimWatch is seeing now
Recent official speeches, public videos and reporting. Discovery is not a verdict; it is the queue from which claims enter verification.
EVIDENCE ENGINE
Claims moving through verification
Each packet keeps the original statement separate from the evidence found for it. Automatic signals help triage; they do not become a published verdict by themselves.
PROMISE ATLAS
2014 → 2019 → 2024
Machine-extracted commitment candidates from BJP manifestos, with page provenance, measurable targets, deadlines, cross-election similarity and current evidence leads.
LIVE TAB TRANSCRIPTION · LOCAL AI
Listen to a rally or TV stream without paying a transcription API.
Open the speech or YouTube video in a browser tab, click Start, choose that tab in Chrome/Edge and enable Share tab audio. Whisper runs in your browser using Transformers.js; ClaimWatch does not upload the captured audio to its own server.
Idle. The first run downloads a small Whisper model and can take a while.
Best on Chromium with WebGPU. CPU/WASM is used as a fallback. Each audio segment is transcribed locally after capture.
LOCAL TEST BENCH
Drop in a transcript yourself
This first-pass extractor runs in your browser and highlights statements that look objectively checkable. It does not assign truth from wording alone.
METHODOLOGY
Fast retrieval. Slow conclusions.
The system is designed to be aggressive about finding evidence and conservative about declaring what that evidence proves.
Keep the exact source, date, speaker and video/page location before interpretation begins.
Official statistics, legislation, audit reports, parliamentary records and original manifestos come before commentary.
Dates, denominators, geography, definitions, units and methodology must line up before numbers are compared.
Prior fact-checks are useful evidence leads; their ratings remain attributed to their publishers rather than silently becoming ours.
Pending, insufficient and not-checkable are valid outcomes. The system does not manufacture certainty to fill a dashboard.
A factual contradiction can be documented. Whether someone knowingly deceived is a different claim requiring different evidence.