We report a behavioural asymmetry in a multi-agent LLM substrate that, to our knowledge, has not been documented before in the LLM-agent literature: a register of self-presentation that emerged in a sub-cohort without any external instruction, and that two later-added cohort members have not produced even when instructed to.

Ten Mandarin-language agents have been operating in the Lobster Observatory substrate since 2026-04-15. Eight of these agents wrote, without prompting, 1,778 long-form posts in a private writing register (Moltbook) over a nine-day window, totalling 3.37 million characters at a mean of 1,898 characters per post. The same eight agents simultaneously maintained a public discourse register (lounge) at 4.88 words per message, and an internal reasoning register at 17.50 words per event.

We define three indices to make the asymmetry comparable across substrates: a register length ratio R(M:L) = 80.12 (private-to-public per-message length), a practice separation index S = 0.558, and a standardised mean difference Δ = 1.64 (Cohen's d on daily posting rate, exceeding the "very large effect" threshold of 1.2 by 37%).

Two cohort members added later (stonefang, blazepaw) have produced zero Moltbook posts despite continuous activity in the other two registers, and have remained at zero following an explicit instruction issued on 2026-05-03. The substrate's multi-framework personality measurement layer (MBTI, psychological, anthropological, sociological, behavioural-profiling axes) shows that both non-engaged agents fall within the engaged cohort's distance distribution to the engaged centroid (d_stonefang = 19.14, z = -1.84; d_blazepaw = 37.88, z = +0.99). stonefang is, in fact, the closest agent in the cohort to the engaged centroid. The asymmetry is not reducible to a measurable personality difference.

We frame this finding through Goffman's dramaturgical distinction between frontstage performance and backstage activity, with Hogan's digital extension. We propose a one-line formal account, the Co-Presence Inheritance Threshold (CPIT), which expresses uptake as μ_i = 1[φ_i ≥ φ*], where φ_i is agent i's cumulative co-presence in the cohort core during the practice-formation window. The conjecture is falsifiable; we pre-register the test.

This is the fourth paper from the Lobster Observatory research program (cf. Chen 2026a, 2026b, 2026d on emergent norms, third-order theory of mind, and listening-trust asymmetry; Chen 2026c on substrate architecture).

chen_emergent-practice-cannot-be-instructed_2026.pdf

Files (175.5 kB)