# Model Observatory — full agent brief (llms-full.txt)

Canonical: https://viz.kyanitelabs.tech/
Updated: 2026-08-27

## One-sentence pitch

Model Observatory is a free, MIT-licensed **3D LLM benchmark instrument** that maps models on **intelligence × cost (blended $/M) × speed (tok/s)** and supports **Decide mode**: intelligence floor → cost/speed shortlist with shareable URLs.

## Product surface (human UI)

1. **3D stage** — WebGL points; open = wire shape; lab brand colors; Pareto ridge; multi-effort trails.
2. **2D projections** — linked speed/intel, speed/cost, cost/intel views.
3. **Table** — membership list of the visible set.
4. **Filter shelf** — age, multi-effort, providers, families, openness, VRAM tiers, non-reasoning exclusion.
5. **Decide panel** — floor slider, cost/speed bias, eligible shortlist.
6. **Atlas dock** — natural language / voice over **catalog tools** (offline grammar by default; optional OpenAI- or Anthropic-compatible endpoint, including NUCBox Unsloth BYOK via local Vite proxy).
7. **Cinema mode** — stage-first presentation chrome.

## URL state (shareable)

Query params (subset): `decide=1`, `floor=`, `bias=`, `anchor=`, `providers=`, `families=`, `age=`, `me=`, `ax=`, `w=`, `catalog=`. Session-only: hover, pin, cinema.

## Static model pages, embeds, and frontier log

Generated from the catalog on every build (zero JS, self-contained):

- `/m/<slug>/` — one static card per catalog model: measured tok/s, blended $/M, intelligence index (+ coding/agentic indexes where measured), percentile context across the catalog, per-field provenance labels, catalog-date stamp. All-models index: `/m/`.
- `/embed/<slug>` — iframe-embeddable per-model cards, regenerated with every catalog refresh.
- `/frontier-watch.md` — dated movement log of the measured Pareto frontier (blended price ↓ × intelligence index ↑); `/frontier-watch.json` carries the machine-readable diff.
- `/blog/2026-08-24-model-observatory/` — launch post (SEO + AIGEO structured).
- `/sitemap.xml` — `/`, `/m/`, every `/m/<slug>/`, `/frontier-watch.md`, `/blog/2026-08-24-model-observatory/`, `/about.md`, and self-listings for `/llms.txt` and `/llms-full.txt`.

Per-model numbers on these pages are catalog-sourced and labeled as such; lab-measured lanes stay separate by design.

## Data policy

- Catalog is **curated static/refresh pipeline**, not live free-form scraping in the browser.
- **Nulls are honest** — unmeasured Index/cost/speed never fabricated.
- Sources disclosed in UI footer + /sources/: Arena (CC BY 4.0), vendor-published figures, first-party pricing. Third-party measurement services are not republished here (license).

## Atlas agent tools (in-app, not public HTTP API yet)

Host-side pure tools (ground truth from catalog):

- get_catalog_meta
- search_models / get_model
- list_eligible / rank_eligible
- propose_floor
- compare_models
- finish_turn → structured UI proposal (floor, filters, shortlist, pin)

High-impact UI writes use confirm/apply. Optional LLM only drives tool calls; metrics still come from tools.

## MCP / CLI (status)

**Local (in-repo) live:**
- CLI: `npm run atlas:cli -- meta|search|eligible|rank|get|compare|floor`
- MCP stdio: `npm run atlas:mcp` (JSON-RPC tools/list + tools/call over catalog)
- Snapshot: `npm run catalog:snapshot` → `data/atlas-catalog-snapshot.json`
- Full-app Atlas in browser: cinema/pin/filters/decide/axes + Apply/Undo

**Public origin:** MCP HTTP endpoint **not** deployed yet. Requirements: `docs/agents/mcp-cli-api-requirements.md`.
Do not invent a public MCP URL until M5 ships.

## What not to claim

- Not an LLM API marketplace checkout.
- Not a substitute for Artificial Analysis official rankings without attribution.
- Not offline-capable for full WebGL without a modern browser.
- Public site does not host private NUCBox / Unsloth endpoints.

## Repos

| Role | URL |
|------|-----|
| Product SoT | https://git.kyanitelabs.tech/simon/llm-3d-viz |
| Public MIT | https://github.com/KyaniteLabs/llm-3d-viz |
| Live | https://viz.kyanitelabs.tech/ |

## License

MIT — see repository LICENSE.