v3 (April 29, 2026): Extends v2 with four new principles (P15 LLM Inference Routing, P16 Regression Detection, P17 Domain-Expert Evaluation, P18 Misalignment Trigger Auditing) drawn from two additional batches of S3 specifications (621 cumulative) and validated against an independent market corpus of 1,843 ideas (125ai). Introduces Combo D (Agent Containment Stack) as a recombinant architecture validated by 22 cross-source convergent specifications. Documents an empirically observed phase transition in the ecosystem's output between March (discovery of safety primitives) and April (productization of established patterns). Updates the originality taxonomy to 3 genuinely novel, 9 novel combinations, and 6 independent rediscoveries. References expanded from 13 to 25; new Limitations section explicitly addresses the model-level priors confound shared with the companion Convergent Synthesis paper. v2 (March 22, 2026): Added experimental validation of Principle 1 (formal verification) via substrate-guard framework. 135 test cases, 100% accuracy, zero false positives. Code: https://github.com/octavuntila-prog/substrate-guard We present an extended catalog of 18 emergent AI safety principles ("bijuterii") that an autonomous multi-agent ecosystem (SUBSTRATE) proposed independently in product specifications, without explicit instruction to consider safety as a topic. Version 1 of this catalog (March 2026) documented 14 principles extracted from 245 product specifications generated over 27 days. Version 3 extends the empirical base in three directions: (1) a longitudinal extension to 621 product specifications across 54 days (245 original + 180 batch April 4-23 + 196 batch April 24-29), yielding 4 new principles; (2) a cross-source validation against an independent market-tracking corpus of 1,843 ideas that confirms 13 of 18 principles emerge in market discourse with similar mechanisms but different framing; (3) the documentation of a phase transition in the ecosystem's output, from "discovery of safety primitives" (March: Z3, eBPF, hardware trust) to "monetization of established patterns" (April: routing, governance, regulated industries), observed empirically without intervention. Of the 18 principles, our updated originality taxonomy classifies 3 as genuinely novel, 9 as novel combinations, and 6 as independent rediscoveries. The most novel principle (P13, AI tokens as employee compensation) appeared in zero of 2,464 cumulative observations from market sources, confirming its rarity. We frame this as evidence that a sufficiently complex autonomous AI system, when allowed to dream about useful products, dreams disproportionately about how to keep itself accountable — and that those dreams converge with what the surrounding market is starting to ask for.
Octavian Untila (Wed,) studied this question.