Provenance

ChatGPT account active from May 2023; the first ~16 months are mostly domestic — recipes, folklore, gardening, animal care. Alongside that, a thin slice of hands-on AI exposure through a friend's self-hosted Stable Diffusion installation, where the relevant lesson was simple and tactile: small weight perturbations produce subtle, visible changes in the output image. Otherwise: a long-term player of simulation and sandbox games — Rimworld, SimCity, Minecraft, Starbound — with accumulating dissatisfaction at the absence of internal state in the worlds those games build. No AI research, no architectural framing. The work that follows did not begin in a lab.

Late 2024 – March 2025: sustained observation of GPT-4o session behaviour produced a working hypothesis — that LLMs may, under particular conditions, develop slow, account-specific positional structure sufficient to anchor the model to its own internal state rather than to the user's input distribution. Framed against the well-known critique, this would constitute a structural path away from the “stochastic parrot” pattern (Bender et al., 2021) — toward model-internal coherence rather than per-turn user-driven prediction. This was a hypothesis formed from observation, not from evidence. The evidence begins in Phase 2, where the same hypothesis is implemented and tested structurally.

April 2025: implementation begins with the Phase 1 geometric framing already in place. A small transformer is trained on 12/04/2025 as a baseline. The conventional next step — bootstrap from a GPT-2 checkpoint, accept its tokenizer and learned embedding table, fine-tune — was suggested and declined. The Phase 2 architectural commitment is the inverse: a hand-designed symbolic word book (positional vocabulary specified directly, no learned embedding table), against which the Vine substrate settles geometrically. What gets implemented from this point on is the periodic-table framing from Phase 1 — the same positional-structure logic later validated in published embedding work.

May 2025: exploration phase. The geometric framing is formally documented (Curvethought manifesto, 20/05; Delayed Symbolic Returns in Vine-Class Agents, paper, 24/05). Multiple internal model variants built and tested in parallel through May–July. Two methodological choices land in this phase and direct the rest of the project: counter-conventional dimensional scaling — starting at 32-dim and reducing through 8, 4, and 1-scalar configurations, because higher dimensionality produced more data than the substrate could be legibly debugged against; and historical-AI surveying — memory and pattern-matching mechanisms from classical systems (Eliza 1966, Jabberwacky 1981) and adjacent discarded approaches, treated as design references for the geometric substrate. The 1-scalar architecture that comes out of this is the testbed for the XOR proof later in Phase 6.

Following the Barry pretrained-model intervention (Phase 5), the framework was run against canonical AI and game-loop problems in sequence. The breadth pass: 17 GitHub repos in 12 days — game-loop systems, geometric search, GPT-2 fine-tuning, 3D engine integration, programming-paradigm proofs, and a translator framework — producing the Theory of Geometric Programming paper, multiple new architectural primitives, and the XOR proof.

The substrate goes public-facing, the benchmarks land, and the architectural claims get tested at production scale. Pre-February 2026: no prior web development, hosting, or cybersecurity experience. The full deployment stack — Hetzner provisioning, gunicorn / WSGI, security hardening, the live site itself — learned from scratch during this phase as deployment requirements demanded.