← Journal §14 · 3D orbit model · Multi-agent §10 · MA10 · Figures home

Round-trip asymmetry (MA12)

Lotman's test, stated bluntly: if you translate out and back and get the same text, you are inside an artificial language (logic, not poetry). High recovery is evidence of transmission. Low recovery can mean a creative relation, or it can mean corruption. Thirteen corpus stretches, five conditions, temperature 0.0, reverse passes audited so the original never leaked into the return call.

Figure R1 · Recovery by condition

Mean token Jaccard vs the original (n = 13 each)

Observed order (qwen2.5:3b / gemma2:2b, 2026-10-02): R0 (0.877) > R3 cross-model (0.740) > R2 cross-tier (0.386) > R1 register (0.274) > R4 three trips (0.251). Archived Llama 3.2 sizes (3B/1B) gave R0 1.000 and R3 0.891; see journal §14.

Figure R2 · Where the distance actually lives (with a caveat)

Loss of recovery relative to R0 = 1.000

Loss relative to this stack's R0 (0.877): register change (R1) costs about 0.603; qwen→gemma on the reverse pass (R3) costs about 0.137. The Llama 3B/1B run (R0 = 1.000) showed 0.776 vs 0.109 and looked like "instruction beats weights" until both models proved to be one family. Cross-family qwen/gemma lowers R3 but does not collapse it to the register floor. MA15 tests four family dyads. Citations: journal §11, §14.

Figure R3 · Low recovery is not creation

Original (silence unit): Detailed drought-and-kulak denial in this window sits in the Ottawa embassy release, Novosti, and the Khmil UN statement, not in a matching domestic Pravda article.

After register round trip (R1): statements denying the drought and persecution of peasants (known as kulaks) were published in a Canadian newspaper called Novosti and by Ukrainian politician Khmil.

Jaccard 0.224 on this stretch. The archive's documented absence came back as a positive claim about a Canadian newspaper. Lotman's asymmetry test is necessary but not sufficient: add a licensing check (does new material appear in the corpus?) before calling low recovery "creative."

Figure R4 · Three trips, no fixed point

Mean Jaccard vs original across three register round trips

Drift decelerates (successive-iteration similarity rises) but only 2 of 13 stretches reached a fixed point against a preregistered threshold of 7 (qwen stack, 2026-10-02). The instructive fixed point deleted scare quotes and the named course the memorandum is about. Convergence bought by erasing what made the document evidence is not recovery.

Limits: thirteen stretches, one plate, open local stacks (qwen/gemma). A high score can mark refusal to transform (9/13 R3 rows echoed forward verbatim on 2026-10-02). Vacuous survival measures read 1.000 when no stretch contained the item type. Source: ma12_roundtrip_report.json in the thesis lab; Llama archive ma12_roundtrip_report_2026-10-01_llama.json.