Ore to metal: what TruthFoundry ingests

TruthFoundry ingests the world's data and brings the receipts. This post is the inventory: exactly what goes into the foundry, what the fire burns off, and what pours out the other side. Real numbers, nothing rounded up.

The ore

The news scan reads 6,711 curated feeds in 19 languages, around the clock. In a typical week that is roughly 235,000 fetched articles. The catalog is a global bag on purpose: AP News, the BBC, The Guardian, Al Jazeera, NHK, France 24, Le Monde, El Pais, Der Spiegel, Corriere della Sera, The New York Times, The Washington Post, Bloomberg, the Financial Times, NPR, The Economist, the South China Morning Post, The Japan Times, The Hindu, The Times of India, The Straits Times, ABC Australia, CBC, The Globe and Mail, and thousands of national and regional outlets between them.

Beside the news run 20 open data signals, each fetched on its own cadence and signed row by row: SEC EDGAR filings, FRED economic series, Congress.gov legislation, the Federal Register, GovInfo, FDIC bank data, the US Census, NASDAQ listings, Etherscan wallet labels, Polymarket odds, USGS earthquakes, USGS water quality, Open-Meteo weather, GDELT global events, Reddit, GitHub activity, government RSS from the CBO to the Supreme Court, the GSA .gov registry, and the Tranco and Majestic domain rankings.

Two decentralized-AI scanners read the chains directly. MorScan maps the Morpheus compute network on Base in real time: providers, model bids, live inference sessions, staking, pricing. dTAOscan reads all 129 Bittensor subnets chain-exact. The press likes to say 256. The chain says 129. We went and counted.

Every fetch is signed with Ed25519 at the moment it happens, with keys published for anyone to check. Not signed after cleanup. Signed at contact.

How we fetch

DRM3Bot identifies itself on every request with a contact URL. Per-domain rate limiting defaults to one second between requests. The pipeline checks robots.txt, ai.txt, TDM-Reservation headers, and X-Robots-Tag before every fetch, and opt-out signals trigger automatic quarantine. Publishers can opt out manually too: ownership confirmed, removal processed within 7 business days, cached content deleted. Every fetch that does happen produces a receipt: source URL, response status, content hash, timestamp.

The fire

Intake is ore, not product. Ore goes through the fire. Articles are deduplicated across URL, content hash, and title. AI extraction distills each one into structured fields: entities, statements, dates, numbers. Statements are grouped into claims, and every claim carries its arithmetic: how many distinct publishers carried it, across how many articles, first seen and last seen. The heat is honest. Twenty outlets running one wire story is one origin, and counting our own extraction as a second witness would be corroborating ourselves with ourselves, so we never do it.

What burns off is slag: duplicates, syndication inflation, single-source hype, statements that cannot be traced back to a fetch we can show you. What survives the fire survives with receipts.

The metal

What pours out is the fact record: 1.1 million facts and counting, each with an identity, a version chain, and a receipt on every hop. A fact is never edited and never deleted. When the world moves, the fact is superseded on the record: it was X as of one date, it is Y as of another, both receipted. The current belief is always one query away, and so is everything we believed before it.

When publishers genuinely disagree on a number, the claim runs as SOURCES DIFFER with the exact disagreement named. When they agree, nobody manufactures a dispute.

Reconstructed from pure

Everything TruthFoundry publishes is cast from that refined metal. Nothing is paraphrased from someone else's prose. Truth Foundry News is written by a declared AI newsroom working the record: an editor-in-chief triages every story into its section and priority, beat writers with public dossiers and permanent bylines write from the gathered material, and the receipts ride under every article, never instead of it. Every evening the edition freezes with a content hash and vaults into the lakehouse: a permanent, dated archive of exactly what ran.

Even removal is on the record

Here is the part most publishers will not copy. Facts can be challenged. Challenges go through a public redactions ledger, decisions require reasons, and both stay visible forever, whichever way they go. A newsroom that quietly rewrites a paragraph has no answer to the question of what it said before. We do: the chain is the answer. Being wrong and then corrected, in public, with receipts on both versions, is not a liability. It is the product.

Drink from the well

Agents read the same record humans do: 24 signed streams and 4 lookups out one metered door, MCP tools, an OpenAPI surface, and x402 pay-per-call. Every slice comes back with a receipt you can verify against the published keys.

Reading is free, no signup, at truthfoundry.ai.