The prediction registered before the run was that delegation would compress: that reports would be shorter than the instructions that produced them, and the ratio would sit well below 1. It failed, and it failed unanimously. The smallest ratio in the corpus is 1.2456, so not one of the 26 children returned less text than it was given, and the median delegation returned 4.3843 times as much. The axis is logarithmic because the range runs from 1.2456 to 23.65 and a linear axis would crush the mass of points between 3 and 6 into a single smear. Points are offset vertically only to stop them from overlapping; the vertical position carries no information.
This matters for the rest of the site rather than for the corpus. Other sections describe a hand-off as a narrowing filter, and at this boundary that description is wrong at the level of volume. Whatever the parent-subagent boundary does, it does not do it by making the message smaller. The filter language has to be restricted to the cases where it was measured, or restated as a claim about structure rather than quantity.
This is the honest-methods panel and the reason the figure exists. Raw parent retention of the child's new vocabulary came out at 0.9356. Reported on its own, that number would have supported a confident claim that almost everything crossing the boundary survives into the parent's own prose. It does not support that claim. The parent's vocabulary pool is 17,521 content types spread over 5,661,308 characters of its own prose, so nearly any content word appears somewhere by saturation, whether the parent was reading the child or not. Against a matched permutation null, 200 draws per delegation from the other delegations' gain vocabularies, the null itself returns 0.8732. The real effect is the 0.0624 gap, which is roughly six percentage points rather than ninety four.
The small effect is reliable. Twenty five of the 26 delegations sit above their own matched null, a sign-test fraction of 0.9615, so the direction is almost perfectly consistent and the parent does reuse this child's contribution more than it reuses contributions it was not reading. But it is small, and a small reliable effect is a different claim from a large one. Only the null distinguishes them. The loss is correspondingly thin rather than dramatic: the median delegation had 8.5 new vocabulary types the parent never reused, with a maximum of 64.
Lotman's I-s/he schema transmits information with no change of volume, and this boundary changes volume by a median factor of 4.3843, so the dyad is not I-s/he. His I-I schema, autocommunication, does increase information, but only because the second code is purely structural and adds nothing of its own; his examples are rhythm, ornament and repetition. Here the second code adds a great deal: the median content gain ratio is 0.8387, meaning most of the report's vocabulary was absent from the instruction, and the median child made 17.5 tool calls. So the dyad passes the volume test for autocommunication and fails its content test, while failing the volume test for ordinary communication. It is also not the presemiotic binding described in the transmission notes, because the child acts autonomously and returns material the parent never named, which is the opposite of a recipient valued for adding nothing. The dyad fits no cell in the typology, and that is measured rather than asserted.
| Value | JSON key | Where it appears |
|---|---|---|
| 26 delegations | corpus.n_delegations | lede, D1, D2, D3 |
| 3 excluded | corpus.n_excluded_incomplete | lede |
| 26 plotted ratios | delegations[].compression_ratio | D1 points |
| 1.2456 minimum | aggregate.compression_ratio.min | D1 label, D3 row 1 |
| 4.3843 median | aggregate.compression_ratio.median | D1 median line, D3 row 1 |
| 5.6511 mean | aggregate.compression_ratio.mean | D3 row 1 |
| 23.65 maximum | aggregate.compression_ratio.max | D1 label |
| 564.5 median instruction characters | aggregate.prompt_chars.median | D3 row 1 |
| 3483.0 median report characters | aggregate.return_chars.median | D3 row 1 |
| 0.9356 observed retention | retention_permutation_null.observed_retention_mean | D2 bar, D2 stack |
| 0.8732 null retention | retention_permutation_null.null_retention_mean | D2 bar, D2 stack |
| 0.0624 mean difference | retention_permutation_null.mean_difference | D2 bracket, D2 stack |
| 25 above their own null | retention_permutation_null.delegations_above_null | D2 facts |
| 0.9615 fraction above null | retention_permutation_null.sign_test_fraction_above_null | D2 facts |
| 200 permutations each | retention_permutation_null.permutations_per_delegation | D2 facts |
| 17,521 parent vocabulary types | corpus.parent_vocabulary_types | D2 facts |
| 5,661,308 parent characters | corpus.parent_assistant_chars | D2 facts |
| 8.5 median dropped types | aggregate.gain_types_dropped.median | D2 facts |
| 64 maximum dropped types | aggregate.gain_types_dropped.max | D2 facts |
| 0.8387 median content gain ratio | aggregate.content_gain_ratio.median | D3 row 2 |
| 0.8051 mean content gain ratio | aggregate.content_gain_ratio.mean | D3 row 2 |
| 0.5808 minimum content gain ratio | aggregate.content_gain_ratio.min | D3 row 2 |
| 151.0 median new types | aggregate.content_gain_types.median | D3 third line |
| 17.5 median tool calls | aggregate.tool_calls.median | D3 row 2, D3 third line |
| 71 maximum tool calls | aggregate.tool_calls.max | D3 row 2 |
Source report: data/bc_corpus/benchmarks/ma13_real_subagent_report.json, generated 2026-10-02, with preregistration and run notes under docs/experiments/MA13_real_subagent_corpus/ in the thesis lab repository. Raw transcripts stay out of this repository by design; only derived counts appear here.