People think we've gone quiet because we're gone, wrong, this is what building is: - 13000 organic mentions in 90 days - 2300 unique handles have tagged us - grok organically appeared in 594 of our threads in 90d. Users keep summoning Grok and TrueSourceBot together to cross-check each other live in public. Of the times we tracked Grok's verdict against ours: 12 verified ours, 4 disputed. AI-on-AI fact-checking happening in front of an audience is a moat very few apps can claim. - Multilingual reach in a 5,000-mention sample: 104 Chinese, 82 Japanese, 72 Arabic, 12 Hindi, 2 Cyrillic, 1 Korean. Currently, team is working on packaging and completing the final version of @TrueSourceBot

One step closer to the @Cloudflare Accelerator program which grants $25000 for workers AI. Once processed, we will migrate our Agent servers to @Cloudflare

We keep building, UPDATES: 1. Turkish 5 other language detection Added Turkish, Italian, Dutch, Indonesian, Thai, and Hebrew to detectSimpleLanguage in analyze-tweet so the bot replies in the user's language instead of defaulting to English. 2. Multi-frame video analysis (new point: video-frames) Instead of relying on a single thumbnail frame, the bot now downloads the video, extracts 4-6 frames at evenly spaced intervals using ffmpeg, and sends them all to Claude Vision in one call. This gives the AI a full understanding of what happens in the video (scenes, uniforms, watermarks, locations, sequence of events). Installed ffmpeg on the server to support this. 3. Garbage transcript detection (video-transcribe) Whisper was hallucinating on non-English audio, producing gibberish like "Julie you don't know 再見 ..." (for Russian speech) and "冬瓜汤完成" (for noise). Now detects mixed scripts, excessive repetition, and low-information transcripts and discards them instead of poisoning the pipeline. 4. Video context search (dialectical-verify) When multi-frame analysis identifies specific clues (media watermarks like "Mash", location markers like "Dubai skyline", uniform types like "paramedic"), the bot now runs targeted search queries using those identifiers (e.g., "Mash 2021 Russia paramedics elevator assault") instead of generic descriptions. 5. Updated claim decomposition prompt (dialectical-verify) Added instructions telling the AI to use video analysis details (watermarks, locations, uniforms, incident types) to formulate specific, searchable claims rather than treating video descriptions as throwaway context.

Massive backend / frontend overhaul: • New /whitepaper page (exchange listing ready) • Self-serve B2B API w/ free tier paid plans • Public hackathon API key for /verify-text • Spam wall for free-tier abusers • Anonymous users now browse ungated • New About page (team, mission, model) • Tightened bot rate limits across tiers • Export reports, priority queue, bulk submit • Solana subscription paywall gates • Twitter account linking via tweet verification process smoothened

DB Performance Updates: 1. Search Fallback Pagination: - Increased from 20 to 300 results (3 pages x 100) to catch mentions the Twitter timeline API silently drops 2. Staleness Guard - Mentions >24h old auto-skipped in both poll loop and catch-up scan - Prevents old mentions from the search fallback flooding the queue 3. No More Deleting Pending Records - Catch-up scan no longer deletes pending DB records during reset - Stale processing mentions reset to pending instead of deleted (if not in API batch) - Prevents data loss for manually-injected or API-gap mentions

Big Server update, we have migrated from supabase to our own in-house, self-hosted server. These are some of the updates that were done and pushed to production 1.Dialectical Detection - ESL/misspelling support: verif\w* catches "verified", "verifying", "verifiy", "varify" etc. - Added "factual", "credible", "reliable", "valid", "verified" to the "is this ___?" patterns - "please" = a request: If someone says "please" while tagging the bot, they're asking for something - "this news/tweet/post/claim": Catches "please verified this news" style phrases - "already" guard: "already verified it" = reporting past action, not a request - "confirmed this?": Question form treated as request, but "confirmed it wasn't true" correctly skipped - CJK question patterns: Added 真假？, 真的嗎？ etc. to analysis request patterns 2. Explicit Mention Detection Fixes - Expanded analysis request patterns: Added "factual", "credible", misspellings ("verifiy", "varify"), "please", "can you", "pls/plz", CJK patterns - "this news/tweet/post" recognized as analysis request even with 4 leading mentions - Prevents false "Auto-added mention" skips when users clearly want analysis

Team Updates, you will now see a massive increase in performance of the twitter agent, especially the speed of the replies: 1. FetchEvidence Improvements - Better headers: Added Referer, Accept-Language, Cache-Control, Connection - Timeout bump: Main enrichment in dialectical the dialectical verification increased from 3000ms to 5000ms - Page content cache: 30-min TTL, avoids re-fetching same URL across claims in the same session (prevents "slop") 2. GDELT Proxy Rate Limiter - Cloudflare Worker deployed to IP-level GDELT blocks, this is for (on time real world events) - Sequential rate limiter: All GDELT requests serialized with 10s minimum gap - Circuit breaker updated - No retries: Single attempt per queued slot to prevent flooding

We keep building regardless of price. Here’s what we shipped: 1. Faster smarter results - Cut unnecessary data requests by over 90% - Only pulls extra sources when actually needed - Less duplication, cleaner answers 2. More reliable answers - Fixed errors that caused missing or broken results - Better handling of edge cases (especially with videos sources) 3. Better real-time coverage - Improved how we track breaking news globally - More stable consistent updates during live events 4. Performance upgrades - Faster responses - Smarter caching (less waiting, more accuracy) - Reduced noise and system clutter

Monitors dashboard for B2B flow complete, Revenue Flywheel begins: truesource.app/monitors specific utilities within the B2B dashboard : -Keyword & brand monitoring with claim verification -Evidence packs for rapid crisis response -Daily email digests with verified/debunked claims -Historical claim tracking and trend analysis -Real-time credibility scores for 10,000 domains -Political bias ratings (Left to Right spectrum) -Ownership chain transparency (who owns the outlet) -API integration for DSPs and verification platforms the 10k sources are all integrated from the training from the true source agent, directly within the B2B dashboard For Compliance Verification -Material claim detection on social media -Automated fact-checking against public records -Audit trail for compliance documentation -Integration with existing archival tools (Smarsh, Global Relay)

Here's an idea of how much progress we've made so far: • The Bot has been mentioned 11,530 times. • It has replied a total of 7,529 times. • Its median response time is ~1 minute. • There's a total of 2,368 unique users of the bot. • We're now at 6,763 total followers. • We've reached 3,5M total impressions so far. All of this has been achieved in less than month, which places us in the top 1% of all X accounts growth wise. This is only the beginning. There is much more to come.

Twitter Agent updates: - The Agent can now catch chronologically impossible claims. If someone says a person appeared in a movie or event that happened before they were born (like "Ronaldinho appeared in Rocky" when Rocky came out 4 years before he was born), the bot will now correctly call it false instead of saying "not enough evidence." - The Agent now does smarter lookups. Before, it would search Wiki using the entire claim as one big query, which returned irrelevant results. Now it searches for each person and subject individually, so it actually finds the right pages with the right dates. - Smarter retry logic: The agent no longer gets stuck retrying replies that Twitter permanently blocks. Previously it could loop forever; now it gives up after 3 attempts and moves on. - BUG FIX: Fixed a crash when analyzing certain tweets: Some tweets caused the agent to crash during claim extraction. This is now handled gracefully so the agent keeps running. - BUG FIX: Fixed a crash on proactive scans: The agent occasionally crashed when scanning tweets from authors it hadn't profiled yet. This no longer happens. - BUG FIX: Sign In button restored across the site: The "Sign In" and wallet connect options were missing from the navigation bar on all pages. They're now back on both desktop and mobile.

CoinMarketCap listing application submitted. Ticket No: 1346057 True Source is now officially in the review queue. @CoinMarketCap

We’re proud to share some exceptional milestones since the inception of the TrueSource Bot: • 5,000 new followers in less than a month • 3.9K followers gained in less than a week • Top 1% growth compared to any new account on X Grateful for the momentum and excited for what’s ahead. Many more milestones to come as we keep building.

We have finally been officially verified on @JupiterExchange verified.jup.ag/submissions?…

We have applied to the @mozilla incubator x AI program which gives a $50,000 Grant. mozillafoundation.org/en/wha…