The receipts · Issue 03
Who Supplies Grok?, with its work shown
Every load-bearing claim in Issue 03, with its primary source. The ingestion terms come from X's and xAI's own policies; the citation findings from the original studies; the supplier asymmetry from the Issue 02 Atlas archive.
At a glance
159
Active entities
5–15×
Citation asymmetry
2
Primary policy sources
2
Citation-accuracy studies
Issue 03 makes a structural claim and declines to claim what it cannot source. Where a fact is held privately by xAI, the Issue says so rather than estimating it. These are the sources for everything it does claim.
The grounding moat — xAI's own account
The claim that the X firehose is Grok's distinguishing advantage rests substantially on xAI's and Musk's own characterization, corroborated by independent analysts.
Musk on real-time X access as a "massive advantage"
At Grok's launch, Musk framed real-time access to X as a core advantage over rivals. The origin of the moat claim, in the principal's own words.
Open →Independent analysis: the proprietary data moat
Market analysis attributing xAI's competitive position to exclusive access to X's live stream plus compute — the corroboration of the self-claim from an outside source.
Open →The third-party-training reversal as moat-building
After the merger, X restricted third-party AI training on its data. Analysis reading the move as reserving the firehose for xAI while shutting rivals out.
Open →Default-on ingestion — the platform's own terms
The claim that the cohort's work enters Grok free and by default rests on X's and xAI's own published policies.
X Privacy Policy
States that X "may use the information we collect and publicly available information to help train our machine learning or artificial intelligence models." The primary source for default ingestion.
Open →xAI FAQ on model training
Confirms xAI "may use your content and interactions with Grok along with Grok's responses to train our models," with Private Chat the stated exception.
Open →Terms timeline & the EU rollback (DPC actions)
The November 2024 terms update, the January 2026 expansion of "Content" to Grok prompts, and Ireland's DPC High Court order and 2025 statutory inquiry — the one place ingestion was forced to stop.
Open →Attribution & dependence — the citation studies
The claim that credit does not reliably return, and that the systems depend on retrieved reporting, rests on two studies read together.
Tow Center / CJR — "AI Search Has a Citation Problem" (March 2025)
Sixteen hundred queries across eight tools; more than 60% wrong collectively; Grok 3 worst at 94%, with 154 of 200 citations leading to error pages. Notes that even Grok overwhelmingly cites traditional news organizations.
Open →Stanford news-intermediary evaluation (February 2026)
Six chatbots including Grok 4; accuracy has improved, but more than 70% of remaining errors are retrieval failures, not reasoning — the systems are near-totally dependent on the reporting they pull in. Strengthens the case for licensing-with-attribution.
Open →The supplier asymmetry — the Atlas foundation
The 5–15× citation asymmetry, the 159-entity roster, and the platform-flight finding all come from Issue 02 and its archive.
Issue 02 — the Atlas, and its receipts
The five-to-fifteen-times citation asymmetry (Phase 8b §3), the entity roster, the citation bands, and the platform-flight finding. The full 21-phase archive is browsable there.
Browse the Atlas receipts →Phase 9b entity table
The full entity spreadsheet underlying the count. The 159 active figure reconciles from the ~200-row catalog by resolving 31 cross-listings, 5 closed/merged entities, and 2 parent rows.
Open →Licensing comparators
The market-price reference points: News Corp–OpenAI (up to $250M/5yr), Reddit–Google (~$60M/yr reported), and the small-outlet pitch floor (~$1M). Reported figures, not all independently confirmed.
Open →The companion model
An illustrative valuation snapshot — recognized methods, bullish stated assumptions, explicitly not an appraisal.
The leaflet & the valuation registry
What a defensible model produces under stated assumptions, and the full 159-entity registry behind it. A snapshot, clearly marked as such — not a receipt in the strict sense, but shown openly.
Open the leaflet →Continue from here