Honesty: supporting occupancy + MiaAI TP=2 credit (do not merge)

FEEL_PASS.md

@@ -1,9 +1,9 @@
 ## Feel-pass (Dream stack · 2026-08-19)
 
-1. Cold open: occupancy table and **measured** teb-69 scores before any flag diary. Hero is the scoreboard (Baton vs 0731 vs Qwen). Not a blank recipe dump.
+1. Cold open: occupancy table and **measured** teb-69 scores before any flag diary. This repo is **supporting** occupancy, not the public hero ([miaai35-tune](https://github.com/Coinupbtc/miaai35-tune)). Screenshot is the scoreboard (Baton vs 0731 vs Qwen).
 2. Primary path: `./setup.sh` copies `env.example`, syntax-checks, prints laptop-vs-cluster next steps. `./setup.sh --up` is the two-Spark path and refuses if `:8888` is not 0731.
 3. Break / empty: no `.env` → copy + “edit N2_IP / SPARK2 / GGUF”. 0731 down → `up.sh` exits 2 with Mia start knobs. Never pretends a laptop can load 0731.
 4. Loading: roommate + baton take minutes; `scripts/status.sh` / `smoke.sh` are the success surface (`UP 0731 / Qwen / baton`).
-5. Would a stranger keep this? **yes if they own two Sparks.** Everyone else still sees proof (scores + flags) and a one-command no-GPU setup. Mia credit is scoped.
+5. Would a stranger keep this? **yes if they already run MiaAI 0731 TP=2 on two Sparks.** Everyone else still sees proof (scores + flags) and a one-command no-GPU setup. MiaAI/Anemll credit for TP=2 dual-node serving is above the fold.
 
 Evidence: `docs/screenshots/hero.png` · `RESULTS.md` · `./setup.sh` (no `--up`).

README.md

@@ -4,17 +4,19 @@
 
 ![CI](https://github.com/Coinupbtc/dream-stack/actions/workflows/ci.yml/badge.svg)
 
-Occupancy recipe for **two NVIDIA DGX Spark** boxes (GB10, CX7).
+Supporting occupancy for **two NVIDIA DGX Spark** boxes (GB10, CX7) after 0731 TP=2 is already up. Not the public hero — that is [miaai35-tune](https://github.com/Coinupbtc/miaai35-tune).
+
+**Credit — TP=2 / dual-node serving:** [MiaAI-Lab / Anemll DeepSeek-V4-Flash DSpark TP=2](https://github.com/MiaAI-Lab/DeepSeek-v4-Flash-DSpark-2x-DGX-Spark) is the recipe that runs 0731 across both Sparks. This repo adds leftover-UMA occupancy (Qwen GGUF roommate + baton) on that pair.
 
 ## At a glance
 
 | | |
 |---|---|
-| **What it is** | Keep Mia 0731 TP2 as the chat brain, room a Qwen 3.8 GGUF on leftover UMA, and talk to **one URL** (`dream-baton` `:8877`). |
-| **What it’s for** | Spark owners who want **measured occupancy**, not “one model per box.” Laptop clones still get the scores, flags, and a no-GPU `./setup.sh`. |
+| **What it is** | Supporting occupancy: keep MiaAI 0731 TP2 as the chat brain, room a Qwen 3.8 GGUF on leftover UMA, talk to **one URL** (`dream-baton` `:8877`). |
+| **What it’s for** | Spark owners who already run the MiaAI TP=2 pair and want leftover-UMA occupancy, not a hire-lead or Spark flagship. Laptop clones still get the scores, flags, and a no-GPU `./setup.sh`. |
 | **How to use it** | `./setup.sh` (orients). Two Sparks + 0731 already at 348k / 0.74: edit `.env`, then `./setup.sh --up`. |
 
-**GitHub description:** What: 2× Spark occupancy — 0731 TP2 @ 348k + Qwen roommate + baton. For: clone, copy `.env`, bring the stack up. How: `./setup.sh` then `./setup.sh --up`.
+**GitHub description:** What: supporting occupancy on MiaAI 0731 TP2 — leftover Qwen + baton (hero is miaai35-tune). For: clone, copy `.env`, bring the stack up. How: `./setup.sh` then `./setup.sh --up`.
 
 This is **not** a new engine. You already run the public [MiaAI / Anemll DeepSeek-V4-Flash-0731 DSpark TP=2](https://github.com/MiaAI-Lab/DeepSeek-v4-Flash-DSpark-2x-DGX-Spark) pair. Dream keeps that chat brain and rooms two more services in leftover UMA:
 
@@ -27,7 +29,7 @@ This is **not** a new engine. You already run the public [MiaAI / Anemll DeepSee
 
 NVFP4 Qwen **does not fit** next to 0731 TP2. The roommate has to be the GGUF. This is **not** “one model per Spark.” 0731 still owns both UMAs.
 
-Measured 2026-08-16 on this occupancy (tool-eval-bench 69, seed 42, thinking off, 69/69, error 0):
+Measured 2026-08-16 on this occupancy ([tool-eval-bench](https://github.com/SeraphimSerapis/tool-eval-bench) suite of 69, seed 42, thinking off). All 69 scenarios **completed** (0 infra errors). That is run completion, not a 69/69 pass score. Numbers below are the harness `score` and pass / partial / fail. CI in this repo is syntax + `./setup.sh` only — it does not run teb. Reproduce with `bash scripts/run-teb.sh 69 baton` (and `0731` / `qwen`):
 
 | Brain | Score | Pass / partial / fail | Wall |
 |---|---:|---|---:|
@@ -232,4 +234,4 @@ House-only A/B (parks engines on **this** cluster): `run-prime-then-solo-qwen-69
 - 0731 TP2 engine: [MiaAI-Lab / Anemll DSpark](https://github.com/MiaAI-Lab/DeepSeek-v4-Flash-DSpark-2x-DGX-Spark)
 - Qwen 3.8 weights: [Qwen](https://huggingface.co/Qwen) / [Unsloth GGUF](https://huggingface.co/unsloth/Qwen3.8-27B-GGUF)
 - Public tool-call eval: [tool-eval-bench](https://github.com/SeraphimSerapis/tool-eval-bench) (`v2.5.1.dev29+g573a3ec70` for the numbers in RESULTS.md)
-- Occupancy, roommate flags, vision pin, and baton: this repo
+- Occupancy notes, roommate flags, vision pin, and baton: this repo (supporting; public hero is [miaai35-tune](https://github.com/Coinupbtc/miaai35-tune))

RESULTS.md

@@ -1,13 +1,13 @@
 # Measured scores (2026-08-16)
 
 Instrument: [tool-eval-bench](https://github.com/SeraphimSerapis/tool-eval-bench) `v2.5.1.dev29+g573a3ec70`
-Suite: **69** scenarios · seed **42** · thinking **off** · completion **69/69** · error **0**
+Suite: **69** scenarios · seed **42** · thinking **off** · all 69 **completed** (0 infra errors). That is run completion, not a 69/69 pass score — see Pass / partial / fail. CI in this repo does not run teb; these rows are live `scripts/run-teb.sh 69 …`.
 
 Live occupancy while scoring Dream rows: 0731 TP2 `:8888` `max_model_len=347392` + Qwen GGUF n2 `:8100` `n_ctx=88064` + 4B vision n1.
 
 **Live roommate (2026-08-17 night):** Qwen `n_ctx=116224`, **MTP4** (unique-256 **20.1** tok/s vs MTP3 **18.4**). 124k OOM-killed n2. First baton teb (91) was 88k occupancy.
 
-**2026-08-17 rematch** (same seed 42, thinking off, 69/69): live 116k + n2 traffic cop. MTP5 unique-256 **20.2** (wash vs MTP4 **20.1**). 0731 `MAX_NUM_SEQS=4` / `MAX_NUM_BATCHED_TOKENS=16384` **did not boot** while Qwen held n2 (free 89.71G < 90.05G; later KV 3.91G < 8.14G). Reverted to **6 / 8192**. Do not retry those 0731 knobs without parking Qwen first.
+**2026-08-17 rematch** (same seed 42, thinking off, all 69 completed): live 116k + n2 traffic cop. MTP5 unique-256 **20.2** (wash vs MTP4 **20.1**). 0731 `MAX_NUM_SEQS=4` / `MAX_NUM_BATCHED_TOKENS=16384` **did not boot** while Qwen held n2 (free 89.71G < 90.05G; later KV 3.91G < 8.14G). Reverted to **6 / 8192**. Do not retry those 0731 knobs without parking Qwen first.
 
 | Occupancy | Score | Pass / partial / fail | Wall | Unique tok/s |
 |-----------|------:|----------------------:|-----:|-------------:|

TWEET.md

@@ -1,22 +1,22 @@
-# Ready to post — this is YOUR occupancy recipe
+# Draft tweets — supporting occupancy, not the public hero (miaai35-tune)
 
-Do not quote-tweet Mia as the frame. One credit line is enough (0731 TP2 engine).
-The recipe is: leftover Qwen + n1 vision + baton on top of a live 0731 pair.
+0731 TP=2 serving is [MiaAI-Lab / Anemll DSpark](https://github.com/MiaAI-Lab/DeepSeek-v4-Flash-DSpark-2x-DGX-Spark). This repo is leftover Qwen + n1 vision + baton on a live pair.
 
 Images: tweet-cards/01-teb69-scoreboard.png + 02-speed-tax-fails.png
 
 ---
 
 ## Tweet 1 (card 01)
 
-I run two DGX Sparks as one occupancy I call Dream.
+I run two DGX Sparks as leftover occupancy I call Dream. Public hero is miaai35-tune.
 
+0731 TP=2 serving is MiaAI/Anemll DSpark.
 0731 TP2 @ 348k across the pair.
 Qwen 3.8 27B GGUF on node2 leftover (88k, MTP3).
 4B vision stay-up on node1.
 One URL in front: baton :8877. No tags. It picks the brain.
 
-tool-eval-bench 69 · seed 42 · thinking off · 69/69 · error 0
+tool-eval-bench 69 · seed 42 · thinking off · 69 completed · 0 infra errors (not a 69/69 pass)
 
 0731 Dream  87  56/8/5  8.1 min  39.6 tok/s
 Qwen Dream  90  58/8/3  18.0 min 18.4 tok/s
@@ -39,8 +39,8 @@ Baton flipped required (TC-45). Live 91. Still miss 03 and 61.
 
 Keep-warm: idle TTFT 16.1s → 0.38s.
 
-0731 engine is the public DSpark TP2 stack. Occupancy, roommate, vision pin, and baton are mine.
+0731 TP=2 serving is MiaAI/Anemll DSpark. This tweet is leftover occupancy (roommate + baton), not the lab hero.
 
 ---
 
-No @MiaAI_lab on tweet 1 unless you want a courtesy tag. Optional last line only.
+Credit line for tweet 1: MiaAI-Lab / Anemll DSpark TP=2. Optional @MiaAI_lab.

make-tweet-cards.py

@@ -4,7 +4,8 @@
 
 from PIL import Image, ImageDraw, ImageFont
 
-OUT = Path("/home/coinupbtc/Documents/projects/dream-stack/tweet-cards")
+# Write next to this file so a clone can regenerate cards (no house path).
+OUT = Path(__file__).resolve().parent / "tweet-cards"
 OUT.mkdir(parents=True, exist_ok=True)
 
 NAVY = (11, 18, 32)
@@ -40,7 +41,7 @@ def card1() -> Path:
     im = Image.new("RGB", (W, H), NAVY)
     d = ImageDraw.Draw(im)
     d.text((40, 28), "DREAM STACK  ·  2× DGX Spark GB10", font=font(28, True), fill=TEAL)
-    d.text((40, 72), "tool-eval-bench 69   seed 42   thinking OFF   69/69   infra error 0", font=font(20), fill=MUTED)
+    d.text((40, 72), "tool-eval-bench 69   seed 42   thinking OFF   69 completed   infra error 0", font=font(20), fill=MUTED)
     d.text((40, 104), "v2.5.1.dev29+g573a3ec70   occupancy + baton by coinupbtc", font=font(18), fill=MUTED)
 
     headers = ["Occupancy", "Score", "P / part / F", "Points", "Wall", "Unique t/s", "Med turn"]

setup.sh

@@ -4,8 +4,8 @@ set -euo pipefail
 ROOT="$(cd "$(dirname "$0")" && pwd)"
 cd "$ROOT"
 
-echo "==> Dream stack — occupancy for 2× NVIDIA DGX Spark"
-echo "    0731 TP2 @ 348k + Qwen GGUF roommate + baton :8877"
+echo "==> Dream stack — supporting occupancy on MiaAI 0731 TP2"
+echo "    leftover Qwen GGUF + baton :8877. Public hero: miaai35-tune"
 echo
 
 ok=0