Update README.md

README.md

@@ -9,14 +9,16 @@ thumbnail: >-
   https://cdn-uploads.huggingface.co/production/uploads/6905c926d19f5673e5754b17/FPra8KU5Dq95iVPBgIAfe.jpeg
 ---
 
- <div align="center">
+<div align="center">
+
+<img src="https://faultward-failure.zocomputer.io/cards/banner.svg" alt="Faultward — non-profit open-source model training" width="100%"/>
 
-<img src="https://faultward-failure.zocomputer.io/cards/fleet.svg" alt="Faultward — fine-tuned OSS models, verified & audited, giant-context" width="720"/>
+# FAULTWARD
 
-# ⚡ FAULTWARD
+**A non-profit open-source model-training lab.**
+*We take strong OSS bases, make them faster and smarter, and give every quant away. No website. No API. No customers. Just weights on Hugging Face, with receipts.*
 
-**We fine-tune open-source models and give them away.**
-*No website. No API to sell. No customers. Just cooked weights on Hugging Face — fast, smart, and free.*
+**Status: pre-launch — no releases yet. Numbers get published only once measured on our own hardware.**
 
 `⢕⣿⣏⣿⢳⡕⣆⢺⣋⢟`
 
@@ -24,33 +26,27 @@ thumbnail: >-
 
 ---
 
-## 🩸 Who we are
+## Who we are
 
-A one-person model lab. We take strong open-source bases — starting with **DavidAU's Qwen3.8-27B Cold Fusion** — and make them **as fast and as smart as possible** with new speed tactics, then release every quant, free, Apache-2.0.
+A one-person, non-profit model lab. We start from strong open-source bases — first up, **DavidAU's Qwen3.8-27B Cold Fusion** — push them **as fast and as smart as possible** with current speed tactics, then release every quant, free, under Apache-2.0.
 
-That's the whole company. We don't run a website, sell an API, or take clients. We ship models with receipts.
+That's the whole thing. Nothing is for sale.
 
-## 🎯 Mission
+## Mission
 
 > Same brains. More tokens per second. Proven, not vibed.
 
-- **Fast** — speculative MTP, NEO-imatrix int4, flash-attention everywhere, reasoning-effort compression. 2.5–4× the tok/s, identical outputs.
-- **Smart** — every release ships a tok/s table **and** a quality A/B against the base. No "smarter than base" claim without a measured A/B.
-- **Open** — all genres, all use cases, all quants, one gloriously cursed card at a time.
+- **Fast** — the toolkit we're aiming at: speculative decoding (draft + EAGLE-3), NEO-imatrix int4, flash-attention, reasoning-effort compression. Goal is a real multiple on tok/s with outputs unchanged — measured, not assumed.
+- **Smart** — every release will ship a tok/s table **and** a quality A/B vs the base. No "smarter than base" without a measured A/B.
+- **Open** — all quants, all use cases, one gloriously cursed card at a time.
 
-## 🔥 Flagship
+## Flagship
 
-<div align="center">
+**First target: `DavidAU/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1`** (Apache-2.0) — a 27B dense, vision-language model with 64 layers, 16 full-attention layers (4 KV heads), MTP self-speculation, and 262K native context (→1M extensible). The plan is to re-quantise it (NEO-imatrix) and speed it up (draft/EAGLE-3 speculative), then publish an honest A/B.
 
-<img src="https://faultward-failure.zocomputer.io/cards/qwen3.8-27b-coldfusion.svg" alt="Qwen3.8-27B Cold Fusion — 91 tok/s, 62% MTP acceptance, 262K context" width="720"/>
-
-</div>
+*No benchmarks here yet — nothing has been trained or measured. Speed and quality numbers will land on each release card once they come from our own runs. The full recipe and blank result tables live in [`FINETUNE-PLAN.md`](./FINETUNE-PLAN.md).*
 
-**Qwen3.8-27B Cold Fusion GAIN V1.1 · NEO-MAX · MTP** — DavidAU's six-stage Cold-Fusion + GAIN base, re-quantised NEO-imatrix and sped up. Regular + MTP GGUFs, every quant. Vision-language, three reasoning modes (`xhigh`/`medium`/`low`), 262K context native (→1M extensible).
-
-*Live numbers on the banner above update straight from the bench — the card animates inside any Hugging Face or GitHub README with zero JavaScript.*
-
-## 🖥️ The Fleet
+## The Fleet
 
 | Node | VRAM | Job |
 |---|---|---|
@@ -58,41 +54,40 @@ That's the whole company. We don't run a website, sell an API, or take clients.
 | **2× Titan RTX** | 24 GB each (NVLink 48 GB) | Quant + imatrix workhorse, parallel human eval |
 | **RTX 4060 Laptop** | 8 GB | Kitchen-table testing (not a fleet node) |
 
-*Stack:* Ubuntu 22.04 · CUDA 12.x (never 13 — Turing + GA100 are dropped there) · 570+ open-kernel driver (for the 170HX unlock) · llama.cpp (MTP + imatrix) · vLLM on Ampere. 170HX lives on its own pinned host — snapshot + `apt-mark hold`, never shares a driver stack.
+*Rig is being assembled — the 170HX host isn't in hand yet.*
 
-## 🛠️ How we cook
+*Stack:* Ubuntu 22.04 · CUDA 12.x (never 13 — Turing + GA100 are dropped there) · 570+ open-kernel driver (for the 170HX unlock) · llama.cpp (MTP + imatrix) · vLLM on Ampere. The 170HX lives on its own pinned host — snapshot + `apt-mark hold`, never shares a driver stack.
+
+## How we cook
 
 1. **Baseline** the base — fixed prompts, tok/s + a small quality set. Every later number is measured against this.
 2. **Free-lunch config** — flash-attn, `-ctk/-ctv q8_0`, NEO-imatrix int4, tuned low/medium `reasoning_effort`.
 3. **Draft-model speculative** in llama.cpp — pair a Qwen3-0.6B draft with the 27B. Lossless, headline speed-up.
 4. **Stretch: EAGLE-3 head** on vLLM/SGLang — bigger acceptance, same output distribution. The signature move nobody else trains.
 
-## 📌 Operating principles
+## Operating principles
 
 1. **Proof before panic.** Measure, don't vibe.
 2. **Prove it's lossless** — or say what the speed cost.
-3. **HF is discovery; our cards do the cooking.**
-4. **Boring-reliable beats clever-fragile.** The 170HX stays pinned.
-5. **Ship the series, not the single.** One legendary model beats twenty random merges.
+3. **Boring-reliable beats clever-fragile.** The 170HX stays pinned.
+4. **Ship the series, not the single.** One legendary model beats twenty random merges.
 
-## 🔮 Goals
+## Goals
 
-- [x] v1 Cold Fusion pipeline (COLD-FUSION + GAIN stages) on the 170HX
-- [x] NEO-imatrix + MTP quant chain
-- [x] Animated benchmark cards (live, self-updating, embed anywhere)
-- [ ] **EAGLE-3 speculative draft head** — same outputs, ~3× tok/s
+- [ ] Acquire + unlock the 170HX host, pin the CUDA 12.x / driver stack
+- [ ] Baseline the target model on our own hardware (tok/s + quality set)
+- [ ] NEO-imatrix quant chain
+- [ ] Draft-model speculative decoding (llama.cpp)
+- [ ] **EAGLE-3 speculative head** — same outputs, faster (target TBD by measurement)
 - [ ] Benchmark harness with honest A/B vs base on every release
-- [ ] `faultward.ai` DNS (nice-to-have; the cards don't need it)
-- [ ] **10 public releases**, receipts on every card
+- [ ] **First public release**, receipts on the card
 
 ---
 
 <div align="center">
 
-<img src="https://faultward-failure.zocomputer.io/cards/fleet.svg" alt="Faultward fleet" width="720"/>
-
-<sub><b>Faultward</b> — fine-tuned OSS models · verified · audited · giant-context</sub>
+<sub><b>Faultward</b> — non-profit · open-source · model training</sub>
 <br/>
-<sub>⢕⣿⣏⣿⢳⡕⣆⢺⣋⢟ · built on <a href="https://zocomputer.com">Zo</a>, a noble steed for the mind</sub>
+<sub>⢕⣿⣏⣿⢳⡕⣆⢺⣋⢟ · built on <a href="https://zocomputer.com">Zo</a></sub>
 
 </div>