UPLOAD_STANDARDS.md

# Agent Q3 — Models Bucket Standards
**Bucket:** `MADdegens/Models`
**Owner:** Nick McLeod (MADdegen)
**Last updated:** 2026-05-27

---

## What this bucket is

The GGUF model file store for the Agent Q3 model stack. These files are the source for the Ollama Cloud bootstrap process — `start.sh` with `BOOTSTRAP_MODE=true` downloads GGUFs from here, creates local Ollama models, then pushes them to `nicholasjmcleod/*` namespace on Ollama Cloud.

**This bucket holds large binary files (total ~246GB). Do not sync locally without intent.**

---

## Active model inventory

| GGUF filename | Ollama Cloud tag | Role | Size |
|---------------|-----------------|------|------|
| `Kimi-Linear-48B-A3B-Instruct.i1-Q6_K.gguf` | `nicholasjmcleod/kimi-linear:q6_k` | Reasoner / multimodal | ~38GB |
| `Harmonic-Hermes-9B.i1-Q6_K.gguf` | `nicholasjmcleod/harmonic-hermes:q6_k` | Tandem + Monitor | ~7GB |
| `Qwen3-Coder-53B-...-TOTAL-RECALL-...i1-Q6_K.gguf` | `nicholasjmcleod/qwen3-coder-53b:q6_k` | Coder + Remediate | ~45GB |
| `Qwen3.6-35B-A3B-Kimi-K2.6-Reasoning-Distilled.IQ4_XS.gguf` | `nicholasjmcleod/kimi-distilled:iq4_xs` | Structured reasoning | ~18GB |
| `Genstruct-7B.i1-Q6_K.gguf` | `nicholasjmcleod/genstruct:q6_k` | Generator middleware | ~6GB |
| `Infinity-Parser2-Pro` | `nicholasjmcleod/infinity-parser:latest` | Parser middleware | TBD |

---

## Deprecated — do NOT re-upload

| Filename | Reason deprecated |
|----------|------------------|
| `List-3.0-Ultra-Coder-Brain.i1-Q4_K_M.gguf` | Replaced by Qwen3-Coder-53B TOTAL-RECALL |
| `Qwopus3.6-27B-v1-preview-Q8_0.gguf` | Fallback model — role removed from stack |

If these files are found in the bucket, flag for owner review before deleting.

---

## Naming convention

- GGUF files: use the exact upstream filename — do not rename
- Ollama Cloud tags always use `nicholasjmcleod/` namespace — never `maddegens/`
- Tag format: `nicholasjmcleod/<model-shortname>:<quant>` e.g. `kimi-linear:q6_k`

---

## Bootstrap process (`start.sh BOOTSTRAP_MODE=true`)

```bash
# Downloads all GGUFs from this bucket → creates Ollama models → pushes to nicholasjmcleod/*
BOOTSTRAP_MODE=true ./scripts/start.sh
```

Requires:
1. `ollama signin` with nicholasjmcleod credentials before running
2. Sufficient disk space (~246GB) in `$MODEL_CACHE_DIR` (default: `/models`)
3. `hf` CLI authenticated as madDegen

---

## Upload procedure

```bash
# Upload a single new GGUF
hf buckets cp ./ModelName.gguf hf://buckets/MADdegens/Models/ModelName.gguf

# Never sync the full bucket locally — 246GB
```

---

## Post-upload audit checklist

```bash
hf buckets ls MADdegens/Models
```

- [ ] 6 active GGUFs present (listed in inventory above)
- [ ] `Infinity-Parser2-Pro` present (or flagged as missing for bootstrap skip)
- [ ] No deprecated GGUFs re-introduced
- [ ] `LICENSE` and `README.md` present
- [ ] `UPLOAD_STANDARDS.md` present

---

## Change log

| Date | Action | File | By |
|------|--------|------|----|
| 2026-05-27 | Added UPLOAD_STANDARDS.md | `UPLOAD_STANDARDS.md` | Claude Code / MADdegen |