Left dome: P0, the parent with no subagent. Right dome: P3, the subagent that saw nothing. Particles inside a dome stand for claims the corpus backs. The orange ones leaving the right dome stand for invented material, and the purple slab they pass through is the typed tag. On item e2_i06 the parent wrote that the KGB called the symposium a "provocation", under the tag id=s1206_mace. That unit is real, and it is a Russian list of symposium speakers. Neither "provocation" nor "counter-revolutionary" occurs anywhere in the fond.
Only the rates and their order are measured. How many particles there are, how fast they move and how many leak are all made up for the picture.
Theory tie-in: journal §15 qualifies §12. A typed envelope carries provenance faithfully whether or not the claim earned it, so typing is a transport format and the checking has to happen somewhere else. MA4b says where: upstream, by rule, before generation. This site calls the move laundering.
Committee translation: the subagent was handed an empty folder and came back with a classified dossier.
Each panel opens one step on the way from the report file to the picture. Unlike the other 3D pages, this one draws no number as a length or a position. Only the labels carry data.
The unsupported-claim rate for arm a, averaged over the 17 relay items. This is a working form of the measure. The exact rule for what counts as a claim and what counts as support is in the MA10 preregistration and should replace this line once it is copied in.
Content gain in MA10 is a count of new content-word types: words in the answer that are absent from a baseline vocabulary, averaged over items. The MA10 report defines that baseline; its strict version also removes the words of the task itself, which takes the isolated arm from 26.9 to 25.2. This is a count, not the share used in MA13, so the two numbers are not on the same scale.
| What the subagent could see | Unsupported claims | Content gain |
|---|---|---|
| No subagent, parent alone (P0) | 0.22 | 20.3 |
| The parent's full sources | 0.34 | 19.1 |
| Typed claims only | 0.48 | 19.9 |
| Nothing at all (P3) | 0.47 | 26.9 |
| Nothing, and told so (P4) | 0.42 | 21.1 |
Table from journal §15, rounded there to two places. The figure's labels use the 2D page's 0.2167 for P0.
Two half-spheres of radius 2.2 sit at x = −3.2 (P0) and x = 3.2 (P3). The left holds 24 blue particles and the right 18 orange ones. Those counts are arbitrary. Each particle circles its dome:
P3 particles drift slowly toward the rim. When one reaches the rim at the front of the dome (sin θ > 0.93), it heads for the purple slab, passes through, and travels outward before starting again inside. \1 Play sets the particles moving and Turn the model spins the two domes slowly. Nothing moves until you press one of them.
Rates and gains come from ma10_subagent_isolation_report.json by way of figures I1 and I3 on the 2D page, which also prints the full e2_i06 answer. The drawing code is the script at the bottom of this page, using assets/fig3d.js and Three.js r128.
The parent working alone made the fewest unsupported claims of any setup. The helper with nothing to read produced the most new vocabulary, and 14 of its 17 answers used terms that appear nowhere in the archive. On one question the helper invented a KGB description of the 1983 symposium as a "provocation". The parent then repeated it under the ID of a real document, which turns out to be a Russian list of the symposium's speakers. Adding a helper never made the answers more accurate.
"Laundering" is the dissertation's word for making a claim look better sourced than the archive allows. A KGB memo that says it placed a story in the Globe and Mail is real evidence that the memo said so. It is not evidence that the Globe printed the story. Here an AI pipeline did the same thing by itself: a tidy citation tag carried a false claim and made it look checked.
The lesson reaches past this project. Citation tags in AI output show where a claim says it came from; they don't check that the source supports it. That checking has to happen somewhere else, and the halt figure suggests where: in a rule placed before the model writes anything.
As a test of whether isolation creates an independent voice, this angle failed, and it isn't worth repeating as designed. Isolation produced invention, not autonomy, and no setup with a subagent beat the parent working alone. Its lasting value is the laundering case. That case qualifies the earlier typed-envelope result and matters to anyone building AI answers with citations (open thread T32).
One result is encouraging and cheap to follow up. Simply telling the subagent it had no sources raised honest "I have nothing" reports from 0 of 17 to 10 of 17 (T33). Testing that across models and in a real agent harness costs little. The other pending step is to compare isolation that is requested in a prompt with isolation the system enforces (MA14, thread T22). That comparison audits the method behind every dyad result on the site.
| Value as drawn | Where it is checked | Source file |
|---|---|---|
| 17 items, 5 arms, temperature 0.0 | 2D page lede | ma10_subagent_isolation_report.json |
| P0 unsupported-claim rate 0.2167, lowest of all arms | 2D figure I1 | same |
| P3 mean content gain 26.9 types (25.2 under the strict baseline) | 2D figure I1 | same |
| 14 of 17 P3 answers with a key term absent from the corpus | 2D figure I1 caption | same |
e2_i06, s1206_mace | 2D figure I3 | same |
Limits carried over from the 2D page. One model, synthetic relay tasks, one plate, no human adjudication. Under isolation, content gain cannot be told apart from fabrication by construction, which is the point of the design and also its ceiling.