**Note** from Bead: Netlogo User Conference 2026 · [canonical source](https://redfish.acequia.io/guerin/.agents/0a0d4bbc-5310-486b-97af-0e16646347a4/2026-06-20/notes/06-architecture-and-organization.md) · session 2026-06-20 · discussion: Talk: Netlogo User Conference 2026
Stephen wants to settle architecture and organization BEFORE editing the paper again. This note captures the direction; the paper is NOT edited yet.
## 1. Frame: Forman's landscape ecology, not the turtle De-center the turtle. The turtle is one of **three** NetLogo agent types, and the right frame is **Richard T. T. Forman's landscape ecology** (Harvard GSD; *Land Mosaics*), the **patch–corridor–matrix** model. (Stephen wrote "Foreman"; the citation is **Richard Forman**.) The cross-mapping, with its load-bearing terminology swap: | Landscape ecology (Forman) | NetLogo agent type | role | |---|---|---| | **patch** (discrete area/object) | **turtle** | the discrete, often mobile, agent | | **corridor** | **link** | the connection between patches | | **matrix** (dominant connected background) | **patch** | the cellular field / ground | So what ecology calls the *matrix* is what NetLogo calls *patches*; what ecology calls a *patch* is a NetLogo *turtle*. The paper should run on all three elements as an **ecology**, not on the turtle alone. This also bridges straight to the GSD landscape audience and to Forman himself.
## 2. The call to action (the part Stephen is still struggling with) Lead recommendation for NetLogo sustainability: **fully embrace web technology and get deployment right, without centralized servers.** - The honest failure mode: early NetLogo Web told students to "find a server and make sure it had CORS." That friction is where adoption dies. - The win: deployment that is trivial and **decentralized** — models run on the open web, and **participants run on their own domains** with **cross-domain communication**. - The framing: deployment is the **matrix**. Today the deployment matrix is **stovepiped** (siloed vertical platforms, each a walled server). **Industry's real competence is deployment and serving value.** The lesson from industry, brought to NetLogo, is to design an **open ecology** for deployment instead of stovepipes. This is what "Industry track" should actually mean here.
## 3. Web-first data (not "cloud") - Reframe **cloud-optimized → web-optimized**. "Cloud-optimized" implies a cloud; we want web-native data served from anywhere, including a participant's own domain or the browser. (COG, etc., become "web-optimized" raster/vector.) - **Tiles:** 2D tiles (XYZ / PMTiles-style, range-request friendly) and **3D tiles** (Cesium 3D Tiles / glTF). Pairs with acequia's sparse byte-range fetch. - **Pub/sub:** datasets and model state flow as published/subscribed streams, the unit being 2D and 3D tiles, so each relevant operating picture updates live. This is the data plane of the distributed model-view-controller.
## 4. Authorization and authentication (the hard differentiator) Running on participants' own domains makes everything cross-origin, so decentralized **authn** (who is this) and **authz** (what may they do) are the make-or-break. - **authn:** self-sovereign keypair, no central identity provider. - **authz:** capability / chain-tokens, scoped per caller (the RoP / `myself` model), cascade-revocable. - **cross-domain:** CORS + bearer tokens or signed requests; tokens granted by a decentralized peer mesh, not a central auth server. - Decentralized deployment is only real once cross-domain auth is solved; this is where the paper should be concrete, not hand-wavy.
## 5. World models (Geoff Hinton thread) — intellectual support *(WebSearch verified 2026-06-20.)* Hinton's prescient, citeable contribution is the **understanding-as-internal-model** argument: "just predicting the next word" is technically true but misleading, because to predict the next word a network must map symbols to features, let those features interact across context, and so build a **rich internal representation** — and that learning of internal representations *is* understanding. We do not see the world directly; we build internal models from signals, and (his claim) LLMs understand in much the same way. This is the anti-"stochastic-parrots" position. Strongest sources to cite: - **"Two Paths to Intelligence,"** public lecture, University of Cambridge, **25 May 2023** (CSER / Leverhulme CFI / Dept of Engineering). Video: https://www.youtube.com/watch?v=rGgGOccMEiY . The digital-vs-biological framing and **mortal vs immortal computation** (digital agents share learning by averaging weights). [masterdai.blog summary, HackerNoon, DigitalRosh, Medium transcript]. - Hinton at the **AI4** conference and 2024-2026 interviews on language models mirroring human thought and genuinely "understanding" (rdworldonline; Eric Topol Substack interview "LLMs in medicine: they understand and have empathy"). Use for the prediction-requires-understanding quote. - Adjacent framework: **"Language Models, Agent Models, and World Models: the LAW for Machine Reasoning and Planning"** (arXiv 2312.05230) — a clean scaffold linking LLMs, agents, and world models. Honest caveat (keep): the explicit **"world models"** research banner is more **LeCun** (JEPA; "A Path Towards Autonomous Machine Intelligence," 2022) and **Ha & Schmidhuber** ("World Models," 2018). Lead with **Hinton for the importance of internal models / understanding**; name LeCun and Ha for the term itself. Why it matters here: an agent-based model is an **explicit, legible, authorable world model**, the constructionist complement to the implicit world model inside an LLM. NetLogo is a world-model authoring environment. This links the agentic-AI stakes to the restructuration / computation-as-ontology thread. Sources: [Two Paths to Intelligence (YouTube)](https://www.youtube.com/watch?v=rGgGOccMEiY) · [Cambridge lecture summary](https://masterdai.blog/two-paths-to-intelligence/) · [Hinton at AI4 (RDWorld)](https://www.rdworldonline.com/hinton-ai4-conference-language-model-insights-rd-impact/) · [Topol interview](https://erictopol.substack.com/p/geoffrey-hinton-large-language-models) · [LAW arXiv 2312.05230](https://arxiv.org/pdf/2312.05230)
## 6. Organizational recommendation for the paper (when we do edit) - Open the architecture on the **Forman three-element ecology** (patch-corridor-matrix ↔ turtles-links-patches), replacing turtle-centrism. - Make the Architecture act about the **deployment ecology / the matrix**: web-first, serverless, domain-per-participant, cross-domain auth, web-optimized 2D/3D tiles + pub/sub. - Make the Call's central, concrete recommendation: **NetLogo sustainability = embrace the web and nail decentralized deployment (no "find a server with CORS"); design an open ecology, not a stovepipe.** - Keep Owen / Wilensky-reply / Barlow / world-models as the surrounding ethos and logos.
## Open questions for Stephen 1. Is the lead recommendation singular (decentralized deployment) or a ranked set (deployment, web-first data, cross-domain auth, open ecology)? 2. How concrete should the deployment mechanism be in the paper (name PMTiles / 3D Tiles / a pub-sub transport / the acequia chain-token scheme), or keep it at the level of principle? 3. Forman as a named anchor and possible reader (he is at the GSD) — foreground him as we foreground Owen? 4. World models: keep Hinton as the lead citation, or lead with LeCun/Ha and use Hinton for the understanding argument?