1n2.org · Research · June 2026 · Infographics

Six images that explain Grounded AI

A visual companion to the field report. See the full paper for the prose argument.

The six figures below condense the argument of the Grounded AI field report into images. Each corresponds to a specific section of the paper. The paper is the reference; these are the proof that the argument can be seen at a glance.

Figure 1 of 6

The two modes of AI use

Ungrounded operation has no visible inputs. Grounded operation does. That is the entire point.
UNGROUNDED "It makes mistakes" no source material in reader question language model (black box) ? training memory confident output accuracy uncorrelated with the truth no inputs — no way to verify GROUNDED Reading, not guessing corpus supplied at query time reader question CORPUS primary sources language model (reading) output "quoted text" [source:VOL_053] verifiable citation traces back to source

Corresponds to §1 (the misconception) and §2 (what a language model actually is). The left panel is the failure mode most readers have seen. The right panel is the workflow this paper documents.

Figure 2 of 6

How substring verification works

Every published quotation must substring-match its source row. The build fails otherwise.
1 SOURCE TRANSCRIPT archive row in brain.db source: bitcoin_report · row_id: VOL_053 · date: 2013-06-26 ...this is just kind of what's going on in the community, people like Let's Talk Bitcoin, the real people like MadBitcoins, and the new shows that have just been incredibly impressive... [transcript continues · ffprobed runtime: 14:32] extract candidate quote 2 PUBLISHED OUTPUT rendered quotation with citation the real people like MadBitcoins [bitcoin_report:VOL_053] Brother John, Volume 53 · June 26, 2013 — rendered into /journalism-report/brother-john.html verify 3 VERIFICATION STEP substring match against source row every published quote must substring-match its source row. build fails otherwise — mechanical, not interpretive. matched substring

Corresponds to §4. The check is small, mechanical, and fast — not an interpretive judgement. Its effect is to make hallucinated quotations impossible to ship. Real fragment from The Bitcoin Report, Vol. 53, June 26, 2013.

Figure 3 of 6

The corpus layer

Inputs to the archive, outputs from it. The model sits between, operating on records rather than on training memory.
INPUTS CORPUS LAYER OUTPUTS Whisper transcripts ~5,000 episodes MB · WCN · TBG · bitcoin_report Public records social archives · doc exports tweets · PDFs · open letters Structured metadata ffprobed durations, info.json dates read from file, not inferred Measured quantities runtimes, counts, dates measured, never estimated brain.db corpus ~570,000 records SQLite · local · indexed FTS5 · addressable by source label AI summarisation operates on retrieved rows never on training memory Site builds analytics + journalism pages /journalism-report/ /jason-king/ /shows/ Citation lookups source label → row_id addressable by URL Verified quotations substring-matched at build time ship only if check passes The model sits between. Its job is to operate on the records, not to recall the world. §3 — The archive layer is the primary input to most analytic work.

Corresponds to §3. The archive is a single SQLite database of roughly 570,000 records augmented by transcripts of ~5,000 episodes. It is local, private, indexed, and addressable by source label.

Figure 4 of 6

When the data contradicted the brief

The author assumed Today in Bitcoin ran daily 2017–2020. The archive said otherwise. The published version reflects the data.
THE BRIEF “TIB ran daily 2017–2020” implied: ~365 episodes/year 365 275 180 90 0 2017 2018 2019 2020 2021 assumption: flat · daily · uninterrupted THE DATA Episode counts from brain.db measured, not estimated 365 275 180 90 0 161 59 6 1 6 2017 2018 2019 2020 2021 launched tapered effectively dormant Jan 2021 coda reality: daily through 2017 only, then taper, then silence When the data contradicted the brief, the published version reflects the data. §5.1 — The author was wrong about his own show. The corpus was the corrective. /today-in-bitcoin/

Corresponds to §5.1. The brief was wrong; the corpus was the corrective. This is methodology in action: when the data contradicts the assumption, the data wins.

Figure 5 of 6

When the record is ambiguous, the publication is ambiguous

The Jason King case study: documented charitable activity on one side, documented community concern on the other, no resolution in the record, no resolution in the publication.
DOCUMENTED CHARITABLE ACTIVITY What the record shows on one side 60,000 meals served Sean's Outpost, Pensacola, FL Satoshi Forest land purchase ~9-acre parcel for homeless camp Run Across America cross-country awareness run, 2014 Press, supporter testimony contemporaneous coverage 155 archive episodes · 27 external sources DOCUMENTED COMMUNITY CONCERN What the record shows on the other Doernberg's open letter public call for accounting ~387 BTC accounting gap unreconciled in available records “He was a fraud” — Mandrik attributed quote, on the record No court records, no obituary no recent public posts found present-day proof-of-life search returned no resolution When the record is ambiguous, the publication is ambiguous. §5.4 — The report does not adjudicate. /jason-king/

Corresponds to §5.4. A reliable summary of a contradictory record sounds contradictory. The report does not adjudicate because the record does not.

Figure 6 of 6

Three rules for the non-technical reader

Three questions any reader can ask of any AI-assisted output, without writing code.
Grounded AI · Three Reader Rules How to read AI output Three questions any reader can ask. No code required. 1 Was the model given source material? a transcript, a document, a corpus — anything to read YES → output is checkable against the material NO → treat output as a confident guess 2 Does the output cite its sources? quotations, statistics, claims — can the reader follow back? YES → output is verifiable by the reader NO ask why. 3 Is ambiguity preserved? contradictory record → contradictory summary; or smoothed? YES → the summary is honest about the record NO → certainty was invented or smoothed from the record 1n2.org/reports/grounded-ai-2026-06/ · §7 Practical notes

Corresponds to §7. These are forms of ordinary editorial skepticism. They happen to be sufficient. The takeaway poster of the paper.

These figures accompany the field report Grounded AI: A Field Report on Primary-Source Methodology From One Year of Bitcoin Archive Research. The author can be reached at @MadBitcoins.