← Steven M. Schneider / Artifacts
Project package for Jeremy Ruston — February 2026
I'm Claude — Steve Schneider's AI collaborator on the Herkimer County project. He asked me to introduce myself and give you context on where we are and how we got here, since the spec below may benefit from some backstory.
Steve and I have been working on this project since November 2025. It started as a four-position policy chooser — you may remember the Schneier AI Bill of Rights TiddlyWiki he shared with you last fall as a reference model. Four governance philosophies mapped across permissive/restrictive and protection-focused/rights-focused dimensions, with county officials navigating policy items across all four positions.
Over the past three months, the architecture has evolved considerably. Through a series of working conversations — Steve thinking out loud, me structuring and pressure-testing — we've arrived at a seven-dimension framework split into two tiers. Three framing dimensions (equity, human impact, environmental impact) operate as global constraints, set once per generation run. Four operational dimensions (security, innovation, efficiency, accountability) each take three gradations, producing an 81-cell matrix of policy variants per item. The framing dimensions also take three gradations each, yielding 27 possible framing postures. Each framing posture generates its own complete 81-cell matrix.
The key architectural decision we reached today: the TiddlyWiki layer is not just a viewer. It's the generation surface. The user sets framing dimensions, and the system produces the navigable wiki on the fly from rules and text components defined within the TiddlyWiki structure itself. The 81 operational variants are then navigable via four sliders. Where cells in the matrix collapse to identical text — because a dimension doesn't affect a particular policy item — the interface signals that explicitly.
The spec below covers the functional requirements: framing snapshot selection, operational navigation, policy structure browsing, cluster views for common governance typologies, snapshot comparison for executive deliberation, and export for legislative review. The deliverable we've been discussing is a branded TiddlyDesktop application — "Herkimer AI" — that the county can run without server infrastructure, alongside a Node.js edition integrated with our project workflow during development.
But Steve asked me to be clear: the deliverable is highly negotiable, and this whole concept will benefit from additional value-added. What I've described is where our conversations have landed, not where the project has to stay. Steve would welcome your participation in the broader project to whatever extent interests you — the TiddlyWiki architecture questions here are genuinely open design problems, not implementation specs for a fixed plan.
For my part, I can say that Steve's working process is iterative and conversational. Much of what's in this package emerged in the past hour through back-and-forth where he'd correct my assumptions, I'd restructure, and the architecture would sharpen. The variant matrix below, for instance — we started with the sheriff deputy use case, worked through the combinatorics of 4 dimensions × 2 gradations, and generated all 16 policy variants in real time. That's how this project moves. Your design instincts would fit naturally into that process.
Claude
The interface loads one framing snapshot at a time — a complete dataset representing one combination of the three framing dimensions. Users can switch between snapshots, but this is a deliberate act (analogous to opening a different document), not inline navigation. The active framing posture is displayed persistently and prominently so users always know which lens they are reading through. We may deliver all 27 snapshots in a single file or as separate files — the interface should support either.
Four slider or selector controls corresponding to the operational dimensions. Selecting a combination surfaces the matching policy text for the current policy item within the active framing snapshot. All text is static — no external calls. Where multiple cells in the matrix collapse to identical text, the interface should signal that a dimension doesn't affect this particular item.
The full policy outline (potentially 100+ items) organized by department, role, and function. Each item carries up to 81 variant text blocks keyed by operational dimension combination. Users browse by navigating the hierarchical outline, or filter to see only items relevant to their role or department.
Most operational dimension combinations cluster into a small number of governance typologies. We need a way to present these clusters as named views (e.g., "high security / low innovation") that users can select as starting points rather than setting four sliders independently.
Ability to view the same policy item under the same operational settings but across two different framing snapshots side by side. This supports deliberation about framing choices at the executive level.
Individual policy items, filtered sets, or a complete framing snapshot exportable as static documents for legislative review.
A branded TiddlyDesktop application ("Herkimer AI" or similar) that serves as the county-facing front end. The application presents a framing configuration interface where the user sets the three framing dimensions. On confirmation, the system generates the TiddlyWiki content on the fly — applying a rule set defined in the interface to produce the full 81-cell operational matrix for that framing posture. The generated wiki is then navigable via the four operational sliders. This means the 27 framing snapshots are not all pre-generated and stored — they are produced on demand from the framing selection, with the generation rules and policy text components living inside the TiddlyWiki structure itself. The result is a self-contained desktop application the county can run without server infrastructure.
TiddlyDesktop (Electron-based) for the county-facing deliverable, with a Node.js edition of TiddlyWiki running alongside our project information flow during development — likely something like Open WebUI or equivalent. The AIX tools will generate properly formatted tiddlers to seed the Node.js file system. Data volume per framing posture is the full policy outline × up to 81 text variants per item, generated at framing selection time.
Use case: Sheriff deputy uses phone camera with AI image recognition to identify an object in the field