# Portability & Model-Agnostic Design

The core analysis tools are designed to run in a modern browser without an AI model at runtime. This note states the tested boundary and the known exceptions; it is not a universal compatibility or offline guarantee.

## The core guarantee

The query and search tools are **deterministic engines, not AI models**. When you ask "what is reinforcement learning?" the answer is computed by plain JavaScript scoring over the term data, not generated by a language model. That means:

- No API key, no account, no sign-in.
- The fixed analysis engines do not call a model API.
- Most analysis remains local after the page and its declared assets have loaded; the 3D map has an optional library dependency described below.
- For a pinned data snapshot, engine version, and supported browser semantics, the scored output is reproducible. Presentation details and unpinned assets are outside that claim.

This is the model-agnostic boundary: the fixed engines do not depend on Claude, GPT, or another model to produce their results. Some pages contain embedded snapshots; others load adjacent local assets, so portability must be checked per row rather than assumed for every file.

## What each file needs

| File | Runs offline? | Needs a model? | Notes |
|---|---|---|---|
| Agentic-AI-Governance-Glossary.md | Yes | No | Plain text/markdown, opens anywhere. |
| terms.enriched.json | Yes | No | Declared candidate canonical input. Current embedded consumers are not yet byte-equivalent; see `governance/contracts/atlas-data-sync-baseline.md`. |
| Agentic-AI-Governance-Reflections.html | Yes | No | Self-contained 2D view. Double-click to open. |
| Agentic-AI-Governance-Query.html | Yes | No | Embedded recorded snapshot; deterministic retrieval, but currently not synchronized to `terms.enriched.json`. |
| Agentic-AI-Governance-Chat.html | Yes | No | Embedded recorded snapshot; deterministic retrieval, not an LLM, but currently not synchronized to `terms.enriched.json`. |
| Agentic-AI-Governance-Map.html | Partly | No | 2D layered view works fully offline. The 3D view needs the Three.js library (see below). |

## The one optional dependency: the 3D map

The 3D map uses the Three.js graphics library. The file looks for it in this order:

1. A local file named `three.min.js` sitting in the same folder as the map. (Best for offline.)
2. Public CDNs (needs internet).
3. If neither is available, it automatically shows the **same map as a flat 2D layered list**, so the file is never broken.

To make the 3D map independent of that CDN, download the declared `three.min.js` version and verify its provenance before saving it beside `Agentic-AI-Governance-Map.html`. That removes this library fetch; it does not by itself prove every project route is offline-complete.

## How to run anything here

Most self-contained pages can be opened directly. Pages that load adjacent JSON or JavaScript can be restricted by `file://` browser rules; serve the folder from a local static server for the closest match to the deployed site. The service-worker offline promise is separately checked through the runtime contract.

## If you ever want to add a model on top

The intended architecture is model-agnostic: a future model may consume a typed, versioned projection while the deterministic engine remains separately testable. Today, `terms.enriched.json` is only the declared candidate canonical input; several pages still embed older projections. The sync baseline makes that drift explicit, and the candidate materializer demonstrates a reproducible path without silently rewriting consumer policy. A model, if added, should sit above an accepted projection and must not become evidence, authority, or the source of meaning by itself.