One column per delegation, left to right in report order. Height is the compression ratio on a log scale, and the floor sits at a ratio of exactly 1, where a report would be the same length as its instruction. The red bar across the columns is the median. Press Play to step through them, or drag the slider.
Length is the easy part. What the subagent put in that length is covered on the 2D page: most of the report's vocabulary was new (median content gain 0.8387), and the parent reused only a little more of it than a matched null would predict.
Theory tie-in: the same 26 delegations are the blue ball in Lotman's rooms. Journal §13 uses the volume result to rule out Lotman's I-s/he cell, and §15 (MA10) takes up the question this corpus cannot answer, which is how much of the new material was read and how much was made up.
Each panel opens one step on the way from the report file to the picture, with the equation it uses.
For delegation i, c(·) counts characters. ρ below 1 would mean the subagent sent back less than it was given, and ρ above 1 means more. Across the 26 delegations ρ runs from 1.2456 to 23.65, the mean is 5.6511, and the median is 4.3843, which is the average of the 13th and 14th values in sorted order (4.3694 and 4.3992).
The floor sits at ρ = 1, where ln 1 = 0, so a report the same length as its instruction would have no column at all. A ratio of 30 would reach 3.4 units. On a log scale every doubling adds the same height (3.4 · ln 2 / ln 30 ≈ 0.69), which is why the tick marks for ×2, ×5, ×10 and ×20 are not evenly spaced. The shortest column is h(1.2456) ≈ 0.22, the tallest h(23.65) ≈ 3.16, and the red median bar sits at h(4.3843) ≈ 1.48.
Columns stand left to right in the order the report lists them, 0.38 units apart. That order has no meaning. Only the height is data.
Content gain as the 2D page describes it: the share of the report's content-word types (after a stoplist frozen before the run) that never appear in the instruction. The median is 0.8387, so most of a typical report's vocabulary was new.
Retention is the share of a child's new vocabulary that later shows up in the parent's own prose. On its own, 0.9356 looks like nearly everything survives. The matched null replaces each child's new vocabulary with new vocabulary drawn from the other delegations, 200 draws per delegation, and still gets 0.8732, because the parent writes so much that almost any word turns up somewhere. The real effect is the difference, 0.0624. It is small but consistent: 25 of 26 delegations sit above their own null.
Play highlights one column every 0.8 seconds in report order and prints its exact ratio. The slider does the same by hand. Turn the model spins the row of columns slowly. Nothing moves until you press a button, including when your system is set to reduce motion, and the scene draws nothing while it is still.
Not one report came back shorter than its instruction. Even the shortest ran about a quarter longer. The prediction made before the run was the opposite.
Engineers often describe handing work to a subagent as a filter that narrows things down. In volume, at least, it did the reverse every time. That matters for the dissertation's use of Lotman. In Lotman's scheme, ordinary sending keeps a message the same size, so this is not that. Talking to yourself can make a message grow, but only through form, with no new facts, and most of these reports' words were new, so it is not that either. The rooms figure shows where that leaves AI delegation: in a corner his typology never named.
One caution travels with every number here. The transcripts don't record what the subagents read, so this figure can't tell new material that was looked up from new material that was made up. MA10 takes that question on directly.
The volume result is solid and settles one question: calling a hand-off "compression" is wrong for this corpus. Using content gain as evidence that subagents contributed something is a dead end on these transcripts, because the tool results are missing, and MA10 showed that gain can be pure invention.
The corpus is also narrow: one agent, one operator, one family of tasks, and the person measuring wrote the instructions. More, and more varied, delegation logs would help a great deal, ideally from other harnesses and other people's work, with tool results kept so read material can be separated from produced material. A cheaper alternative while those are gathered: re-run a sample of the same tasks with tool logging switched on and compare the gain with and without access to sources.
| Value as drawn | Report key | Source file |
|---|---|---|
| 26 column heights, in report order | delegations[].compression_ratio | ma13_real_subagent_report.json (same array as D1 on the 2D page) |
| 26 delegations, 3 excluded as incomplete | corpus.n_delegations, corpus.n_excluded_incomplete | same |
| shortest 1.2456, median 4.3843, longest 23.65 | aggregate.compression_ratio.min, .median, .max | same |
| content gain 0.8387 (guide text only) | aggregate.content_gain_ratio.median | same |
Limits carried over from the 2D page. One parent, one operator, one task family, and the instructions were written by the agent that later measured them, so this is participant observation. Tool results and the child's intermediate prose are missing from the transcripts, so nothing here separates what a subagent read from what it produced. The play order is report order and carries no meaning.