# DO-002 — Propose §15.3-bis (Browser-Loadable Model Tier)

**Date:** 2026-06-29
**Status:** Locally binding via OSCR- (proposed upstream to FVW v8 §15.3-bis)

---

## Decision

Oscar's brain manifest declares `Qwen/Qwen2.5-0.5B-Instruct-GGUF` Q4_K_M (500M parameters). Per FVW v8 §15.3 ratified, micro-models MUST be 100k-2M parameters. Oscar's brain exceeds the constitutional ceiling by **250×**.

We propose a new tier: **§15.3-bis Browser-Loadable Model** (1M-1B parameters).

## Why?

The 2M ceiling was set in v6 (2026-06-13) before:
- **Qwen 2.5** series (released 2024-2026)
- **Llama 3.2** series (released 2024)
- **Phi-3** series (released 2024)
- **Modern WebGPU + llama.cpp + GGUF Q4_K_M** ecosystem (mature 2025-2026)

Models in the 100M-1B range now run efficiently in browsers:
- Qwen 0.5B Q4_K_M: ~380 MB RAM footprint, runs on WebGPU/WASM
- Llama 3.2 1B: similar footprint, similar runtime
- Phi-3 mini: comparable

A 2M ceiling is anachronistic. We need a new tier.

## Proposed §15.3-bis text

```markdown
## 15.3-bis — Browser-Loadable Model tier (constitutional, NEW 2026-06-29)

A "browser-loadable model" is an AI model sized 1M-1B parameters, embedded in
the app, runnable on standard browser hardware via WebGPU + WASM + llama.cpp + 
GGUF Q4_K_M + Transformers.js. Constitutional properties:

| Size class | Parameters | Typical model | RAM footprint |
|---|---|---|---|
| Browser-loadable (small) | 1M-100M | Phi-3-mini, Llama-3.2-1B | 100-300 MB |
| Browser-loadable (medium) | 100M-500M | Qwen-0.5B | 300-500 MB |
| Browser-loadable (large) | 500M-1B | Qwen-1.5B, Llama-3.2-3B (quantized) | 500-1000 MB |

Above 1B parameters, the model requires a Cloud LLM tier (§15.4 — non-micro-models).

A Product Smart App at level 5+ may embed a browser-loadable model.

## 15.3-ter — Updated Smart App Level Mapping

Per §16.2 the 6 Smart App levels are now:
1. Plain App
2. Smart Behaviours
3. Smart Engines
4. Smart Modules
5. Browser-Loadable Model (was: Embedded Micro-Model)
6. Superpowered Model (was: Superpowered Micro-Model)
```

## Effect on Oscar

After §15.3-bis ratified, Oscar's brain manifest becomes Level 5 (Browser-Loadable Model). The 250× violation is resolved doctrinally.

## Until ratified

Oscar's brain is **STUB**:
- Brain chunks at `/brain/qwen2.5-0.5b-q4km/chunk-NN.gguf` return PLACEHOLDER_REPLACE_BEFORE_DEPLOY SHA-256s
- `LocalAdapter.complete()` returns PLACEHOLDER error
- `brain-downloaded` localStorage flag remains `false`
- Smart adapter rules engine is the operational fallback

Per the no-invention rule (§08 Rule 3), Oscar does not pretend the brain is available.

## Cross-references

- `pact/platform/app-pact.md` §3 (Micro-Model Status)
- `pact/recipes/smart/micro-model.json` (status: STUB)
- `pact/dna/app.dna.json` `doctrinal_extensions.micro_model_status`
- `pact/codex/intelligence.codex.md` C5 (Failure Modes: §15.3 ceiling)