docs(public): professional cards, PII-clean measured surface, exclusive research/prompts off HF

LAUNCH_AUG3.md

@@ -1,19 +1,18 @@
 # Laguna Spark HF launch lock — 2026-08-03
 
 Status: **LOCKED** · measure tip + model card + lock set filled `2026-07-29T14:43:26+07:00`  
-Owner: personal (hizrianraz)  
 Repo: https://huggingface.co/hizrianraz/Laguna-S-2.1-Spark-Agentic  
-Timezone: Asia/Jakarta (WIB)
+Timezone: local (UTC+7)
 
 ## Freeze window
 
 | When | Action |
 |------|--------|
-| Now → Aug 2 18:00 WIB | Build + measure only |
-| Aug 2 18:00 WIB | Content freeze (no new claims without re-measure) |
-| Aug 3 09:00 WIB | Preflight |
-| Aug 3 12:00 WIB | **GO LIVE** (card + announce ready state) |
-| Aug 3 18:00 WIB | Post-launch verify + brief |
+| Now → Aug 2 18:00 local | Build + measure only |
+| Aug 2 18:00 local | Content freeze (no new claims without re-measure) |
+| Aug 3 09:00 local | Preflight |
+| Aug 3 12:00 local | **GO LIVE** (card + announce ready state) |
+| Aug 3 18:00 local | Post-launch verify + brief |
 
 ## Launch definition of done (all required)
 
@@ -24,27 +23,26 @@ Timezone: Asia/Jakarta (WIB)
 5. Measured artifacts on HF (not only local): `results/measured.json`, `server_bench.json`, `agent_smoke.json` (+ receipt)
 6. agent_smoke headline honest:
    - Min ship: **≥38/40 (95%)**
-   - **Current measured: 40/40 (100%)** · **84.86s** live tip 2026-07-29 13:22 WIB · pack `bf82eab`
-7. Gen throughput quote: **~21.47 tok/s** @ 128 completion (multi suite; pre-smoke gen128 22.50)
-8. Clear disclaimer: personal · not first quant · not Poolside/Nous affiliate
+   - **Current measured: 40/40 (100%)** · **84.86s** live tip 2026-07-29 13:22 local · pack `bf82eab`
+7. Gen throughput quote (sole headline): **~21.47 tok/s** @ 128 completion from multi suite in `results/measured.json`
+8. Clear disclaimer: independent · not first quant · not Poolside/Nous affiliate
 9. No DIY GGUF unless it beats a scoreboard cell (default: **bind official only**)
 10. Launch receipt JSON written + card updated same day
-11. **Freeze gate filled:** model card + lock set (`results/freeze_gate_model_card_lock_set_2026-07-29.json`)
+11. **Freeze gate filled:** model card + **S-only** locks (`results/freeze_gate_model_card_lock_set_2026-07-29.json`) — XS lock is parallel research, not S freeze set
 
 ## Live tip (binding until weights/runners change)
 
 | Field | Value |
 |-------|--------|
-| Pack git | `bf82eab5fd6c1fb04e863f0c4b05b5658dec4aee` |
+| Measure tip (provenance) | `bf82eab5fd6c1fb04e863f0c4b05b5658dec4aee` — docs tip may advance after |
 | Engine | poolsideai/llama.cpp `04b2b72` |
 | Quant | official Q4_K_M · sha256 `a8b55c75714ea73fd90ec85de5defdc0b8d88ca0ad2108343cdd8fc22f7583e4` |
 | agent_smoke | **40/40** · 84.86s · runner `3bb81080…` |
 | hermes v2 | **27/27** · 100.1s · runner `20c1e52a…` |
-| gen128 quote | **21.47 t/s** |
+| gen128 quote | **21.47 t/s** (do not headline short single-shot or pre-smoke spikes) |
 | Artifacts | `results/measured.json` · `MEASURED.md` · smokes |
 
 Re-smoke only if weights, runners, or serve flags change. Fresh same-harness green may refresh elapsed without claim change.
-
 ## Non-goals (hard)
 
 - Org / company product branding on this surface
@@ -52,17 +50,17 @@ Re-smoke only if weights, runners, or serve flags change. Fresh same-harness gre
 - Side-quest models under this tag
 - Fake 100% smoke or unverified tok/s
 - X/public brand push beyond HF card before launch
-- Local S weights on founder Mac ≤32G
+- Local S weights on Apple Silicon ≤32 GB class
 - XS inside S freeze artifact
 
-## Daily clock (WIB)
+## Daily clock (local)
 
 | Day | Date | Theme | Exit gate |
 |-----|------|-------|-----------|
 | T-6 | Jul 28 | Lock + HF gap close | measured JSON on HF, launch file live |
 | T-5 | Jul 29 | Smoke harden + freeze-gate fill | **DONE** — 40/40 · 27/27 · card+locks filled |
 | T-4 | Jul 30 | DFlash / serve sweep | **DONE early 2026-07-29** — DFlash **DO_NOT_PROMOTE** (15.286 vs ~21.5); baseline re-health HEALTH_OK; stranger dry PASS; scoreboard row live |
-| T-3 | Jul 31 | Card polish | stranger dry already PASS; optional card polish only |
+| T-3 | Jul 31 | Card polish | **DONE early 2026-07-29** — Layer B live **35/35** research receipt; stale “no live claim” wording cleared; freeze bars untouched |
 | T-2 | Aug 1 | Buffer / second measure | optional Unsloth abort/keep only if free |
 | T-1 | Aug 2 | Freeze 18:00 | final push; no claim changes after freeze |
 | T-0 | Aug 3 | Launch 12:00 | go-live receipt; 18:00 verify brief |
@@ -72,17 +70,16 @@ Re-smoke only if weights, runners, or serve flags change. Fresh same-harness gre
 Smoke gate clear: **40/40 measured** on tip `bf82eab`. Hold that bar through freeze; if a later re-run slips, ship last green with named fails (never invent).  
 
 If Spark down at launch: freeze last green MEASURED row; do not invent numbers.  
-If HF token dies: restore the personal Hugging Face write token from the local secret store only; never paste token in chat.
+If HF token dies: restore HF write token from local secret store only; never paste token in chat.
 
 ## Announce text (draft — edits OK)
 
-> Personal DGX Spark agent-runtime pack for Poolside Laguna-S-2.1: measured OpenAI-compatible serve, fixed agent smoke **40/40**, Hermes-class **27/27**, ~21.5 t/s @128, official Q4_K_M digests (no bare re-upload).  
+> Independent DGX Spark agent-runtime pack for Poolside Laguna-S-2.1: measured OpenAI-compatible serve, fixed agent smoke **40/40**, Hermes-class **27/27**, ~21.5 t/s @128, official Q4_K_M digests (no bare re-upload).  
 > https://huggingface.co/hizrianraz/Laguna-S-2.1-Spark-Agentic
 
 ## Authority
 
-Personal founder surface only. Docs/results may be pushed to this HF repo.  
-No company vault binding for this pack.
+Independent pack surface. Docs/results may be pushed to this HF repo.
 
 ## Quant compare + smaller-device path (Aug 3 accelerated)
 
@@ -92,8 +89,8 @@ No company vault binding for this pack.
 - Pull helper: `scripts/pull_sku.sh <sku_id>`
 - Headline remains Spark **official Q4_K_M** measured row
 - Smaller SKUs ship as **pointers + optional same-harness delta** — promote to scoreboard only at ≥ official−2
-- Founder Mac ≤32G: **client → Spark only** for full S; XS = parallel unmeasured track
-- No bulk third-party LFS rehost; no public promo before Aug 3 12:00 WIB
+- Apple Silicon ≤32 GB: **client → Spark only** for full S; XS = parallel unmeasured track
+- No bulk third-party LFS rehost; no public promo before Aug 3 12:00 local
 
 ## Hermes-class smoke v2
 
@@ -101,7 +98,8 @@ No company vault binding for this pack.
 - Does **not** replace launch bar `agent_smoke` 40/40
 - **Measured tip 2026-07-29: 27/27** in 100.1s · artifact `results/hermes_agent_smoke.json`
 - Card lists **27/27 hermes_agent_smoke** alongside **40/40 agent_smoke**
-- **Public promo** (X / social / “top trending” push): still wait until **2026-08-03 12:00 WIB**
+- Layer B research (**35** = 27+8): live **35/35** · 137.56s · 2026-07-29 · receipt `eval/hermes_agent_smoke/layer_b_v3_live_receipt.json` — **not** a freeze-bar field
+- **Public promo** (X / social / “top trending” push): still wait until **2026-08-03 12:00 local**
 
 ## Promo timing (lock)
 
@@ -109,4 +107,4 @@ No company vault binding for this pack.
 |---------|------|
 | HF pack docs + measured artifacts | now → freeze (keep honest) |
 | Soft profile link (GH bio / HF details) | optional now, factual only |
-| Public launch / announce / trending push | **Aug 3 12:00 WIB**, not before |
+| Public launch / announce / trending push | **Aug 3 12:00 local**, not before |

NOTICE

@@ -1,6 +1,6 @@
 NOTICE
 
-This pack documents personal Spark runtime notes for poolside/Laguna-S-2.1
+This pack documents independent Spark runtime measurements for poolside/Laguna-S-2.1
 under OpenMDW-1.1. See LICENSE.md (upstream model repo copy) and
 LICENSE.openmdw-1.1.txt.
 
@@ -8,8 +8,8 @@ Upstream model: poolside/Laguna-S-2.1
 Upstream GGUF: poolside/Laguna-S-2.1-GGUF
 
 Original materials in this repository (SPARK.md measurement scaffolding,
-agent_smoke cases/runner, hermes sample client, scripts): personal work by
-hizrianraz for reproducibility. Independent personal project.
+agent_smoke cases/runner, hermes sample client, scripts): independent work by
+hizrianraz for reproducibility.
 
 Weight binaries are not re-hosted here unless a measured Spark-unique delta
 exists (policy S5). Prefer official GGUF download + sha256 verify via SHA256SUMS.

README.md

@@ -5,57 +5,169 @@ base_model_relation: quantized
 tags:
   - laguna-s-2.1
   - moe
+  - gguf
+  - quantized
   - agentic-coding
+  - coding
+  - code
+  - tool-use
+  - function-calling
+  - agents
+  - agentic
   - dgx-spark
+  - nvidia
   - llama.cpp
-  - hermes-class
   - openai-compatible
+  - hermes-class
   - measured
+  - cuda
 library_name: gguf
 pipeline_tag: text-generation
 ---
 
-# Laguna-S-2.1-Spark-Agentic (personal)
+# Laguna-S-2.1-Spark-Agentic
+
+Measured **DGX Spark** runtime pack for
+[poolside/Laguna-S-2.1](https://huggingface.co/poolside/Laguna-S-2.1)
+(118B total · ~8B active/token · MoE · agentic coding).
+
+| | |
+|--|--|
+| **Default quant** | Official Poolside `Q4_K_M` |
+| **Weight host** | NVIDIA DGX Spark (GB10) |
+| **Launch** | 2026-08-03 12:00 local |
+| **Content freeze** | 2026-08-02 18:00 local |
+| **Affiliation** | Independent · not Poolside · not Nous Research |
+
+This pack publishes **serve pins, digests, and agent smoke results**.
+It does **not** re-host GGUF weights by default — pull official files from Poolside.
+
+Sibling research pack (separate model, not an S quant):
+[`Laguna-XS-2.1-Mac-Agentic`](https://huggingface.co/hizrianraz/Laguna-XS-2.1-Mac-Agentic)
+
+Launch calendar: [`LAUNCH_AUG3.md`](./LAUNCH_AUG3.md)
+
+---
+
+## Why this pack
 
-**Launch target: 2026-08-03 12:00 WIB** — see [`LAUNCH_AUG3.md`](./LAUNCH_AUG3.md)
+Most GGUF pages ship weights only.
+This pack ships the **agent path**:
 
-Personal DGX Spark agent-runtime pack for [poolside/Laguna-S-2.1](https://huggingface.co/poolside/Laguna-S-2.1) (118B-A8B MoE, agentic coding).
+1. Pinned Spark serve settings that actually clear tool-use smoke
+2. Fixed harnesses you can re-run (no vibes)
+3. Official digests for fail-closed verify
+4. OpenAI-compatible client notes for Hermes-class / tool-agent stacks
 
-Solo personal HF surface by **hizrianraz**. Independent measurements on one DGX Spark. Not affiliated with Poolside or Nous Research.
+Headline power on this surface means **agent fitness on a fixed harness**, not boutique re-quant names.
 
-## What this is
+**Strongest signal on this surface:** official Q4_K_M + pinned poolside `llama.cpp` laguna branch + **40/40 agent_smoke** and **27/27 hermes_agent_smoke** on DGX Spark — not an unmeasured boutique quant rename.
 
-Credibility pack for Hermes-class / tool-agent runtimes:
+---
 
-1. **Measured Spark serve** (GB10 / CUDA 13 / llama.cpp Laguna fork)
-2. **Fixed agent smoke suite** (tool JSON, multi-step, error repair, no invented tools)
-3. **OpenAI-compatible sample client** oriented at Hermes-class agents
-4. **Pointers + digests** to official weights — **not** a bare GGUF re-upload
+## Highlights (measured tip)
 
-DIY quants ship **only** if they beat official Poolside/Unsloth GGUF on measured Spark agent metrics. Otherwise we bind the official artifact and publish the Spark delta.
+| Metric | Result |
+|--------|--------|
+| Host | NVIDIA DGX Spark (GB10) |
+| Quant | Official `laguna-s-2.1-Q4_K_M.gguf` |
+| Engine | `poolsideai/llama.cpp` @ `04b2b72` (branch `laguna`) |
+| Context | 8192 |
+| agent_smoke | **40/40** · 84.86 s · temp 0.0 |
+| hermes_agent_smoke v2 | **27/27** · 100.1 s · temp 0.0 |
+| Generation throughput | **~21.47 tok/s** @ 128 completion (sole headline) |
+| Measure tip | pack `bf82eab` · 2026-07-29 13:22 local |
+
+Artifacts:
+[`results/MEASURED.md`](./results/MEASURED.md) ·
+[`results/measured.json`](./results/measured.json) ·
+[`results/agent_smoke.json`](./results/agent_smoke.json) ·
+[`results/hermes_agent_smoke.json`](./results/hermes_agent_smoke.json)
+
+Optional research expansion (does **not** replace freeze bars):
+Hermes-class Layer B v3 **35/35** · 137.56 s ·
+[`eval/hermes_agent_smoke/layer_b_v3_live_receipt.json`](./eval/hermes_agent_smoke/layer_b_v3_live_receipt.json)
+
+---
+
+## Quick start
+
+```bash
+# 1) Engine — pin the measured commit
+git clone https://github.com/poolsideai/llama.cpp
+cd llama.cpp
+git checkout 04b2b72cb54048ead292884adbe11f284e3ec950
+# Spark/GNU may need the isfinite patch in docs/BUILD_SPARK.md
+cmake -B build -G "Unix Makefiles" \
+  -DCMAKE_BUILD_TYPE=Release \
+  -DGGML_CUDA=ON \
+  -DCMAKE_CUDA_ARCHITECTURES=121a \
+  -DLLAMA_CURL=ON
+cmake --build build -j --target llama-server llama-cli llama-bench
+
+# 2) Weights — official Q4 + fail-closed digest check
+huggingface-cli download poolside/Laguna-S-2.1-GGUF \
+  laguna-s-2.1-Q4_K_M.gguf \
+  --revision fc4e481289523cf7d0df668da6d1d391616141ca \
+  --local-dir ~/models/laguna-s-2.1
+cp /path/to/this/pack/SHA256SUMS ~/models/laguna-s-2.1/
+(cd ~/models/laguna-s-2.1 && sha256sum -c SHA256SUMS)
+
+# 3) Serve (flags match last-green pin)
+./build/bin/llama-server \
+  -m ~/models/laguna-s-2.1/laguna-s-2.1-Q4_K_M.gguf \
+  --host 127.0.0.1 --port 8000 \
+  --ctx-size 8192 -ngl -1 --jinja \
+  -fa on --alias local-laguna
+
+# 4) Launch-bar smoke — model id must match --alias
+cd /path/to/this/pack
+python eval/agent_smoke/run_smoke.py \
+  --base-url http://127.0.0.1:8000/v1 \
+  --model local-laguna
+
+# 5) Optional hermes-class suite (does not replace step 4)
+python eval/hermes_agent_smoke/run_hermes_smoke.py \
+  --base-url http://127.0.0.1:8000/v1 \
+  --model local-laguna \
+  --out results/hermes_agent_smoke.json
+```
+
+| Guide | Path |
+|-------|------|
+| Spark notes | [`SPARK.md`](./SPARK.md) |
+| Reproduce | [`docs/REPRODUCE.md`](./docs/REPRODUCE.md) |
+| Build | [`docs/BUILD_SPARK.md`](./docs/BUILD_SPARK.md) |
+| Last-green pin | [`results/LAST_GREEN_PIN.md`](./results/LAST_GREEN_PIN.md) |
+| Sample client | [`hermes/`](./hermes/) |
+
+---
 
 ## License
 
-Base model and derivatives under **OpenMDW-1.1** (Poolside).
-Read the full text: [`LICENSE`](./LICENSE) and upstream
-https://huggingface.co/poolside/Laguna-S-2.1/blob/main/LICENSE
+Base model and derivatives: **OpenMDW-1.1** (Poolside).
+
+- Pack copy: [`LICENSE`](./LICENSE)
+- Upstream: https://huggingface.co/poolside/Laguna-S-2.1/blob/main/LICENSE
 
-You must retain notices, and if you ship Modified Materials with custom code under §3.3 you need a notice file. This pack's original scripts/eval are separate files with clear provenance notes.
+Retain notices. Pack scripts and eval harnesses are separate files with clear provenance.
+
+---
 
 ## Base identity (pinned)
 
 | Field | Value |
 |-------|--------|
-| Base | `poolside/Laguna-S-2.1` |
-| Base revision (card freeze) | `00af5a51782109b587a3b3bbf11875e566036fa7` |
-| Shape | 118B total · ~8B active/token · MoE |
-| Official GGUF repo | `poolside/Laguna-S-2.1-GGUF` @ `fc4e481289523cf7d0df668da6d1d391616141ca` |
-| Stand-behind quant (default) | **official** `laguna-s-2.1-Q4_K_M.gguf` (~68–96 GB on-disk LFS) |
+| Base model | `poolside/Laguna-S-2.1` |
+| Base revision | `00af5a51782109b587a3b3bbf11875e566036fa7` |
+| Architecture | 118B total · ~8B active/token · MoE |
+| Official GGUF repo | `poolside/Laguna-S-2.1-GGUF` |
+| GGUF revision | `fc4e481289523cf7d0df668da6d1d391616141ca` |
+| Stand-behind quant | **official** `laguna-s-2.1-Q4_K_M.gguf` |
 | Optional denser | official `laguna-s-2.1-Q8_0.gguf` |
-| Optional smaller third-party | Unsloth `UD-Q4_K_XL` (~40 GB) — third-party, not first-party claim |
-| llama.cpp fork | `poolsideai/llama.cpp` branch `laguna` (DFlash + Laguna). Upstream PR trail: ggml-org/llama.cpp#25165 |
+| Engine | `poolsideai/llama.cpp` · branch `laguna` · [PR #25165](https://github.com/ggml-org/llama.cpp/pull/25165) |
 
-### Official GGUF sha256 (from HF LFS metadata)
+### Official GGUF digests (HF LFS metadata)
 
 | File | sha256 |
 |------|--------|
@@ -65,133 +177,149 @@ You must retain notices, and if you ship Modified Materials with custom code und
 | `laguna-s-2.1-DFlash-BF16.gguf` | `2ee8aa30338d6599bc7a8ce008cc57c56f2c2b2fdc21f6db9ecda203c751bfd4` |
 | `laguna-s-2.1.imatrix` | `4a4f480f57a3251e3acfb1d35ffba64720662536135e4ca4f4d05b0732539be2` |
 
-Verify after download:
+Pack verify file (GNU 2-column, Q4 default): [`SHA256SUMS`](./SHA256SUMS)
 
 ```bash
-sha256sum -c SHA256SUMS
+cp /path/to/this/pack/SHA256SUMS ~/models/laguna-s-2.1/
+(cd ~/models/laguna-s-2.1 && sha256sum -c SHA256SUMS)
+# or: ./scripts/pull_official_gguf.sh
 ```
 
-## Quant guide (what to pull)
+---
+
+## Which quant to pull
 
 | Goal | Artifact | Notes |
 |------|----------|-------|
-| Default agent serve on **S / Spark** (~121G) | Poolside **S Q4_K_M** | **Stand-behind** · measured agent_smoke 40/40 · hermes 27/27 · ~21.47 t/s |
-| **Founder MacBook / Mac Studio (≤32G) · full S** | **no local S weights** | **Laguna Mac = client → Spark** `http://<spark>:8000/v1` · S IQ3 ~48G alone > 32G RAM |
-| Higher fidelity, more RAM (Spark) | Poolside **S Q8_0** | Routed experts Q8, signal BF16 |
-| Community 64–96G box (pointer, not founder Mac) | Unsloth **UD-IQ4_XS** (~58 GB) or Bartowski **IQ4_XS** (~63 GB) | Third-party; strangers re-run smoke; see device matrix |
-| Community 48–64G tight (Spark pointer) | Unsloth **UD-IQ3_S** (~48 GB) | Spark **38/40 on older runner** (no sanitize) · not headline · **not** a MacBook claim · not Q4 runner-identical |
-| Aggressive experiment | Unsloth **UD-Q2_K_XL** (~40 GB) | Research only until measured |
-| Speculative decode | + Poolside **DFlash-BF16** | Needs poolside `laguna` fork (`--spec-type draft-dflash`) |
-| iPhone / Android phone or tablet | **non-fit** for full Laguna-S | ~40GB+ even at IQ3; mobile NPU/RAM is 4–12G class — need SLM/distill, not this MoE |
-
-**XS is parallel, not an S quant.** Founder Mac ≤32G XS disk candidate = Poolside `Laguna-XS-2.1-Q4_K_M.gguf` (~18.9G) · separate 33B-A3B · **0 Mac smoke** · see [`research/Laguna-XS-2.1-Mac-Agentic-fit-2026-07-28.md`](./research/Laguna-XS-2.1-Mac-Agentic-fit-2026-07-28.md) + dual roadmap.
+| **Default agent serve (Spark)** | Poolside **S Q4_K_M** | Headline · 40/40 · 27/27 · ~21.47 t/s |
+| Higher fidelity (Spark) | Poolside **S Q8_0** | More RAM |
+| Tighter Spark pointer | Unsloth **UD-IQ3_S** (~48 GB) | Spark **38/40** on older runner · **not** headline |
+| Smaller community box | Unsloth UD-IQ4_XS / Bartowski IQ4_XS | Third-party · re-run smoke |
+| Speculative decode experiment | + Poolside **DFlash-BF16** | Measured slower · **DO_NOT_PROMOTE** |
+| Full **S** on Mac with ≤32 GB RAM | **No local S weights** | Use Mac as **client → Spark** only |
+| Phone / tablet | **Non-fit** | Need SLM/distill, not this MoE |
 
-Pull by SKU id:
+**XS is a separate model**, not an S quant.
+Mac-class XS notes: sibling pack + [`research/`](./research/).
 
 ```bash
 ./scripts/pull_sku.sh official-q4km
-./scripts/pull_sku.sh unsloth-ud-iq4-xs   # ~58 GB third-party
-./scripts/pull_sku.sh unsloth-ud-iq3-s    # ~48 GB third-party
+./scripts/pull_sku.sh unsloth-ud-iq3-s    # pointer only
 ```
 
-Device ladder → [`research/device-quant-matrix-aug3.md`](./research/device-quant-matrix-aug3.md)
+Device ladder: [`research/device-quant-matrix-aug3.md`](./research/device-quant-matrix-aug3.md)
 
-**Do not** claim “first quant”. FP8 / NVFP4 / INT4 / GGUF already exist upstream and community.
-**Do not** print third-party numbers as “Spark measured” without same-harness JSON under `results/sku_*`.
+Do **not** claim “first quant”.
+Do **not** label third-party numbers as Spark-measured without same-harness JSON under `results/sku_*`.
 
-## One-evening stranger path (S1)
+### Why official Q4_K_M for agents
 
-```bash
-# 1) engine (poolside laguna fork)
-git clone --branch laguna https://github.com/poolsideai/llama.cpp
-cd llama.cpp
-cmake -B build -G "Unix Makefiles" \
-  -DCMAKE_BUILD_TYPE=Release \
-  -DGGML_CUDA=ON \
-  -DCMAKE_CUDA_ARCHITECTURES=121 \
-  -DLLAMA_CURL=ON
-cmake --build build -j --target llama-server llama-cli llama-bench
+1. Poolside protects the **signal path** on official Q4 (imatrix + first-party chat template).
+2. Same digest → anyone can **reproduce** without trusting a third-party binary.
+3. Headline is **40/40 agent_smoke + 27/27 hermes-class** on that exact file.
+4. DIY / third-party only lands if it **beats** official on the **same** smoke runners, with method disclosure. Today: `diy_gguf: false`.
 
-# 2) weights (official — do not re-host unless you have a measured delta)
-huggingface-cli download poolside/Laguna-S-2.1-GGUF \
-  laguna-s-2.1-Q4_K_M.gguf \
-  --local-dir ~/models/laguna-s-2.1
+Looking unique on HF alone is not a reason to re-quant.
 
-# 3) serve OpenAI-compatible (flags match last-green pin)
-./build/bin/llama-server \
-  -m ~/models/laguna-s-2.1/laguna-s-2.1-Q4_K_M.gguf \
-  --host 127.0.0.1 --port 8000 \
-  --ctx-size 8192 -ngl -1 --jinja \
-  -fa on --alias local-laguna
+---
 
-# 4) smoke (launch bar)
-cd /path/to/this/pack
-python eval/agent_smoke/run_smoke.py --base-url http://127.0.0.1:8000/v1 --model laguna-q4
+## Hermes-class / tool-agent wiring
+
+“Hermes-class” = OpenAI `chat.completions` + `tools` / `tool_calls` shape.
+**Not** a Nous Research product claim or endorsement.
+
+```bash
+export OPENAI_BASE_URL=http://127.0.0.1:8000/v1   # /v1 required
+export OPENAI_API_KEY=sk-local
+export OPENAI_MODEL=local-laguna
 
-# 5) optional Hermes-class suite (v2; does not replace the 40-case bar)
+python hermes/sample_client.py
+
+python eval/agent_smoke/run_smoke.py \
+  --base-url "$OPENAI_BASE_URL" --model local-laguna
 python eval/hermes_agent_smoke/run_hermes_smoke.py \
-  --base-url http://127.0.0.1:8000/v1 --model laguna-q4 \
+  --base-url "$OPENAI_BASE_URL" --model local-laguna \
   --out results/hermes_agent_smoke.json
 ```
 
-Full Spark notes → [`SPARK.md`](./SPARK.md)  
-Hermes-class client → [`hermes/`](./hermes/)  
-Smoke suite (launch bar) → [`eval/agent_smoke/`](./eval/agent_smoke/)  
-Hermes-class smoke v2 → [`eval/hermes_agent_smoke/`](./eval/hermes_agent_smoke/)  
-Measured run → [`results/MEASURED.md`](./results/MEASURED.md)
+Agent loop contract:
+
+1. POST messages + tool schemas
+2. If `tool_calls` → execute **only** offered names → append `role=tool`
+3. Repeat until final `content`
+4. Sanitize invalid prior `function.arguments` JSON before re-send
 
-## Scoreboard (Spark · freezes Aug 2 18:00 WIB)
+Config snippet: [`hermes/config.example.yaml`](./hermes/config.example.yaml)
+Full notes: [`hermes/README.md`](./hermes/README.md)
 
-| Host | Quant | Ctx | Gen tok/s | Smoke | Hermes v2 | Notes |
-|------|-------|-----|-----------|-------|-----------|-------|
-| DGX Spark GB10 | official Q4_K_M | 8192 | **~21.47** | **40/40 (100%)** | **27/27** | **Headline** · engine `04b2b72` · pack tip `bf82eab` · 2026-07-29 13:22 WIB |
-| DGX Spark GB10 | Unsloth UD-IQ3_S | 8192 | short-gen ~41 t/s (tiny) | **38/40 (95%)** | — | **Not headline** · older-runner SKU (no sanitize/any_of_tools) · ship ≥38 met · **not** Q4 runner-identical · [`results/sku_unsloth-ud-iq3-s/`](./results/sku_unsloth-ud-iq3-s/) |
+Mac client pointing at Spark:
 
-Snapshot JSONs: [`results/measured.json`](./results/measured.json) · [`results/server_bench.json`](./results/server_bench.json) · [`results/agent_smoke.json`](./results/agent_smoke.json) · [`results/hermes_agent_smoke.json`](./results/hermes_agent_smoke.json) · [`results/sku_unsloth-ud-iq3-s/`](./results/sku_unsloth-ud-iq3-s/)  
-Hermes-class suite live: **27/27** on Q4 — [`eval/hermes_agent_smoke/`](./eval/hermes_agent_smoke/)
+```bash
+export OPENAI_BASE_URL=http://<spark-host>:8000/v1
+export OPENAI_API_KEY=sk-local
+export OPENAI_MODEL=local-laguna
+```
 
-### Measured detail (live tip · 2026-07-29 13:22 WIB)
+---
 
-| Metric | Value |
-|--------|--------|
-| Quant (headline) | official `Q4_K_M` · sha256 `a8b55c75714ea73fd90ec85de5defdc0b8d88ca0ad2108343cdd8fc22f7583e4` |
-| Engine | poolside llama.cpp `04b2b72` + Spark CUDA serve |
-| Pack git tip | `bf82eab5fd6c1fb04e863f0c4b05b5658dec4aee` |
-| Gen throughput | **~21.47 tok/s** @ 128 completion · ctx 8192 · `-ngl -1 -fa on` (multi suite) |
-| Prefill OK | server_bench 2k 1.597s / 8k 4.78s |
-| agent_smoke | **40/40 · 100%** · **84.86 s** · temp **0.0** · runner `3bb81080…` |
-| hermes_agent_smoke v2 | **27/27 · 100%** · **100.1 s** · temp **0.0** · one-response · runner `20c1e52a…` |
-| Closed fails | harness: `repair_04` sanitize prior tool-args; `long_06` `any_of_tools` judge (not weights) |
-| DFlash | measured 2026-07-29 · gen128 **15.286 t/s** · **DO_NOT_PROMOTE** (baseline ~21.5 wins) · `results/dflash_2026-07-29/` |
-| Multi-device SKU (IQ3_S) | Unsloth UD-IQ3_S · **38/40 · 71s** on Spark · older runner — **not** claim as Q4 regression · phone/tablet **non-fit** |
-| Locks | diy_gguf **false** · weight_host **Spark-only** · founder Mac **client-only** · public promo only after **2026-08-03 12:00 WIB** · XS not in S freeze |
-
-## What is **not** in this pack
-
-- No bare GGUF re-upload when the official file is the stand-behind artifact
-- No affiliate / org / company product claims
-- No fake benches — only fixed `agent_smoke` + `llama-bench` / server timings you can re-run
-- No off-Spark (rented cloud) numbers relabeled as Spark
-- No local S weights on founder Mac (≤32G) — Mac is client to Spark only
-
-## Reproducible method
-
-See [`docs/REPRODUCE.md`](./docs/REPRODUCE.md) for exact clone SHAs, cmake flags (GB10/`121a`), download, measure, and smoke commands.
+## Scoreboard
+
+Freeze clock: **2026-08-02 18:00 local** — no new claims after freeze without re-measure.
+
+| Host | Quant | Ctx | Gen tok/s | agent_smoke | hermes v2 | Role |
+|------|-------|-----|-----------|-------------|-----------|------|
+| DGX Spark GB10 | official Q4_K_M | 8192 | **~21.47** | **40/40** | **27/27** | **Headline** · engine `04b2b72` · tip `bf82eab` |
+| DGX Spark GB10 | Unsloth UD-IQ3_S | 8192 | short-gen only | **38/40** | — | Pointer · older runner · not default |
+
+### Measured detail (2026-07-29 13:22 local)
+
+| Field | Value |
+|-------|--------|
+| Q4 sha256 | `a8b55c75714ea73fd90ec85de5defdc0b8d88ca0ad2108343cdd8fc22f7583e4` |
+| Serve flags | `-c 8192 -ngl -1 --jinja -fa on --alias local-laguna` |
+| Gen headline | **~21.47 tok/s** @ 128 completion · multi suite in `results/measured.json` |
+| Prefill (server_bench) | 2k · 1.597 s · OK · / · 8k · 4.78 s · OK |
+| agent_smoke runner | `3bb81080…` |
+| hermes v2 runner | `20c1e52a…` |
+| Closed harness fails (not weights) | `repair_04` sanitize · `long_06` `any_of_tools` judge |
+| DFlash | gen128 **15.286 t/s** · **DO_NOT_PROMOTE** · [`results/dflash_2026-07-29/`](./results/dflash_2026-07-29/) |
+| Locks | diy_gguf **false** · weights **Spark-only** · Mac full-S **client-only** · public promo after **2026-08-03 12:00 local** · XS **not** in S freeze |
+
+Ship minimum smoke: **≥38/40**. Current measured: **40/40**.
+
+---
+
+## Scope
+
+**In scope**
+
+- Measured Spark runtime pack
+- Official weight pointers + digests
+- Reproducible agent evaluation harness
+
+**Out of scope**
+
+- Bare GGUF re-upload when official is the stand-behind artifact
+- Org / company product branding on this surface
+- Fake or off-host benches labeled as Spark
+- Local full-S weights on Mac hosts with ≤32 GB RAM
+- Public announce / trending push before **2026-08-03 12:00 local**
+- XS rows inside the S freeze lock set
+
+---
 
 ## Attribution
 
-- Model: Poolside Laguna S 2.1 © Poolside, OpenMDW-1.1  
-- GGUF: Poolside official conversions  
-- Engine: poolsideai/llama.cpp `laguna` (+ upstream llama.cpp)  
-- Pack, Spark measurements, agent_smoke, Hermes-oriented notes: personal work by **hizrianraz**
+| Component | Credit |
+|-----------|--------|
+| Model | Poolside Laguna S 2.1 © Poolside · OpenMDW-1.1 |
+| GGUF | Poolside official conversions |
+| Engine | poolsideai/llama.cpp `laguna` (+ upstream llama.cpp) |
+| Pack, measurements, smokes | Independent work by [hizrianraz](https://huggingface.co/hizrianraz) |
 
 ## Disclaimer
 
-Independent personal measurements on one DGX Spark. Not affiliated with, endorsed by, or representing Poolside or Nous Research. “Hermes-class” means OpenAI-compatible tool-calling agent runtime shape only — **not** a Nous endorsement.
-
-## Quant comparison (same-family)
+Independent measurements on one DGX Spark.
+Not affiliated with, endorsed by, or representing Poolside or Nous Research.
 
-- Live scoreboard: [`research/quant-comparison-scoreboard-2026-07-28.md`](research/quant-comparison-scoreboard-2026-07-28.md)
-- Stand-behind: official Poolside **Q4_K_M** (headline) — IQ3_S is a measured smaller **pointer row**, not a default swap.
-- Mac mini / MacBook / PC: ladder in [`research/device-quant-matrix-aug3.md`](research/device-quant-matrix-aug3.md); IQ3_S ship-min met on Spark.
-- iPhone / Android: **explicit non-fit** for full Laguna — distill/SLM only.
+“Hermes-class” describes an OpenAI-compatible tool-calling runtime shape only — **not** a Nous endorsement.

SHA256SUMS

@@ -1,20 +1,5 @@
-# Official Poolside Laguna-S-2.1-GGUF (LFS oid = sha256)
-# repo: poolside/Laguna-S-2.1-GGUF
-# tree_sha: fc4e481289523cf7d0df668da6d1d391616141ca
-# base model: poolside/Laguna-S-2.1 @ 00af5a51782109b587a3b3bbf11875e566036fa7
-# format: sha256  byte_size  filename
-#
-# Stand-behind quant for this pack: Q4_K_M only (official).
-# Q8_0 / F16 listed for reference — not required to pull.
-# DFlash-BF16 is a draft model companion, not a full quant.
-# No DIY requant in this pack (no delta vs official).
-
-a8b55c75714ea73fd90ec85de5defdc0b8d88ca0ad2108343cdd8fc22f7583e4  96031829760  laguna-s-2.1-Q4_K_M.gguf
-d946b221d69f2c5f87a986952bcd3cfb75831e5a6a2184e626e361663e1bfe2b  128750823168  laguna-s-2.1-Q8_0.gguf
-2036c9dcf70f59738d480d1da2a2a59c0a3a3c5bd4ab8c43a61d77fba031e1f8  235202258240  laguna-s-2.1-F16.gguf
-2ee8aa30338d6599bc7a8ce008cc57c56f2c2b2fdc21f6db9ecda203c751bfd4  2233764224  laguna-s-2.1-DFlash-BF16.gguf
-4a4f480f57a3251e3acfb1d35ffba64720662536135e4ca4f4d05b0732539be2  351195360  laguna-s-2.1.imatrix
-
-# Non-LFS supporting files (git blob oid, not sha256 of content blob via LFS)
-# chat_template.jinja  blob b193b1bdcba33acb935dd59567ad827b1821cbc8  size 4028
-# README.md            blob 1f2199a10409dab6987a11be4bc9d03c2d441437  size 2802
+# Official Poolside GGUF digests — Laguna-S-2.1-Spark-Agentic pack
+# Format: GNU sha256sum (2-column) so `sha256sum -c SHA256SUMS` works from the
+# directory that holds the weight file(s). Byte sizes are comments only.
+# Q4_K_M · ~89.4 GiB · 95993903904 bytes
+a8b55c75714ea73fd90ec85de5defdc0b8d88ca0ad2108343cdd8fc22f7583e4  laguna-s-2.1-Q4_K_M.gguf

SPARK.md

@@ -34,7 +34,9 @@ export PATH=/usr/local/cuda/bin:$PATH
 export CUDA_HOME=/usr/local/cuda
 export CC=/usr/bin/gcc CXX=/usr/bin/g++ CUDAHOSTCXX=/usr/bin/g++
 
-git clone --branch laguna https://github.com/poolsideai/llama.cpp llama.cpp-laguna
+git clone https://github.com/poolsideai/llama.cpp llama.cpp-laguna
+cd llama.cpp-laguna
+git checkout 04b2b72cb54048ead292884adbe11f284e3ec950
 cd llama.cpp-laguna
 # pin when publishing numbers:
 # git rev-parse HEAD
@@ -176,17 +178,17 @@ Serve proven: `-c 8192 -ngl -1 --parallel 1 --alias local-laguna --jinja -fa on`
 
 | Setting | prefill latency | tg tok/s | server RSS | notes |
 |---------|-----------------|----------|------------|-------|
-| ctx 8k, ngl -1, fa on | 836 prompt → 1.83 s; 3236 prompt → 4.73 s | **~21.1** @ 128 gen | ~2–3.4 Gi process RSS | unified mem holds weights; quote **21 tok/s** gen |
-| gen8 short | — | ~7.8 | same | tiny completion noise floor |
-| ctx 8k + DFlash (`draft-dflash`) | 2k 2.253s · 8k 5.556s | **15.286** @ 128 gen | higher (draft+target) | **DO_NOT_PROMOTE** vs plain ~21.5 · `results/dflash_2026-07-29/` |
+| ctx 8k, ngl -1, fa on | 836 prompt → 1.597 s; 3236 prompt → 4.78 s | **~21.47** @ 128 gen | ~2–3.4 Gi process RSS | unified mem; **sole headline gen = 21.47** (`results/measured.json`) |
+| gen8 short | — | 10.612 | same | tiny completion noise floor — not a headline |
+| ctx 8k + DFlash (`draft-dflash`) | 2k 2.253s · 8k 5.556s | **15.286** @ 128 gen | higher (draft+target) | trial `-ngl 99` vs baseline `-ngl -1` · **DO_NOT_PROMOTE** still (gen+prefill slower) · `results/dflash_2026-07-29/` |
 
-Agent-shaped (tool round-trips, see smoke):
+Agent-shaped (tool round-trips, see smoke) — **live tip durations bind**:
 
 | Suite | quant | pass | n | notes |
 |-------|-------|------|---|-------|
-| agent_smoke v1 (launch bar) | Q4_K_M | **40/40 · 100%** | 40 | **post-restore 97.79 s** WIB ~23:55 · temp 0.0 · prior same-day 97.25 / 88.96 s · harness fixes: sanitize + any_of_tools |
-| hermes_agent_smoke v2 | Q4_K_M | **27/27 · 100%** | 27 | **post-restore 104.4 s** WIB ~23:57 · temp 0.0 · prior 102.68 s · one-response protocol · `results/hermes_agent_smoke.json` |
-
+| agent_smoke v1 (launch bar) | Q4_K_M | **40/40 · 100%** | 40 | **84.86 s** ·temp 0.0 · tip 2026-07-29 13:22 local · runner `3bb81080…` · older same-day restores ~88–98 s are historical only |
+| hermes_agent_smoke v2 | Q4_K_M | **27/27 · 100%** | 27 | **100.1 s** · temp 0.0 · same tip · one-response · runner `20c1e52a…` · `results/hermes_agent_smoke.json` |
+| hermes Layer B v3 (research) | Q4_K_M | **35/35 · 100%** | 35 | **137.56 s** · 2026-07-29 · **not freeze bar** · receipt `eval/hermes_agent_smoke/layer_b_v3_live_receipt.json` |
 Raw: `results/MEASURED.md`, `results/measured.json`, `results/agent_smoke.json`, `results/hermes_agent_smoke.json`, `results/server_bench.json`.
 
 ## Agent-shaped benches

docs/BUILD_SPARK.md

@@ -10,11 +10,13 @@ cd ~/src/llama.cpp-laguna
 git checkout laguna   # pin: 04b2b72cb54048ead292884adbe11f284e3ec950
 ```
 
-## Patch (measured required on this host)
+## Patch (host-measured · Spark / GNU 13.3)
 
-`common/speculative.cpp` calls `std::isfinite`. On Spark GNU 13.3 that can fail even with `#include <cmath>` present depending on header order / libstdc++.
+`common/speculative.cpp` calls `std::isfinite`. On **this** Spark host (GNU 13.3) that can fail to compile even with `#include <cmath>` depending on header order / libstdc++.
 
-Measured fix that builds clean:
+This is a **local host build fix**, not an upstream laguna behavioral change and not part of the public claim surface. Apply only if stock checkout fails to build the same way.
+
+Measured fix that builds clean here:
 
 1. Ensure `#include <cmath>` and `#include <math.h>` near the top system includes.
 2. Replace the call site:
@@ -23,10 +25,13 @@ Measured fix that builds clean:
 // before
 if (!std::isfinite(v)) {
 
-// after (Spark / GNU 13.3)
+// after (Spark / GNU 13.3 host build)
 if (!::isfinite(static_cast<double>(v))) /* spark-isfinite-fix */ {
 ```
 
+Pin checkout **before** patching: `git checkout 04b2b72cb54048ead292884adbe11f284e3ec950`.  
+Record `git rev-parse HEAD` → pack `results/engine_sha.txt`.
+
 Do **not** treat a failed UI-asset HF fetch during build as fatal — `llama-server` still links; only the optional embedded UI may stay stale.
 
 ## Configure

docs/REPRODUCE.md

@@ -23,11 +23,25 @@ Personal reproducible method.
 4. Download official GGUF (not a DIY re-quant unless you will measure a real delta):
 
 ```bash
+# preferred: fail-closed helper (reads pack SHA256SUMS + pinned GGUF rev)
+./scripts/pull_official_gguf.sh
+
+# or manual download + GNU 2-column check from the weight directory:
 huggingface-cli download poolside/Laguna-S-2.1-GGUF \
   laguna-s-2.1-Q4_K_M.gguf \
   --revision fc4e481289523cf7d0df668da6d1d391616141ca \
   --local-dir ~/models/laguna-s-2.1
-sha256sum -c SHA256SUMS
+cp SHA256SUMS ~/models/laguna-s-2.1/
+(cd ~/models/laguna-s-2.1 && sha256sum -c SHA256SUMS)
+```
+
+Engine pin before build (measured commit, not floating branch tip):
+
+```bash
+git clone https://github.com/poolsideai/llama.cpp ~/src/llama.cpp-laguna
+cd ~/src/llama.cpp-laguna
+git checkout 04b2b72cb54048ead292884adbe11f284e3ec950
+# optional host patch: docs/BUILD_SPARK.md (isfinite / GNU 13.3)
 ```
 
 5. Serve with last-green flags (match `results/LAST_GREEN_PIN.md` / `SPARK.md`):
@@ -40,15 +54,16 @@ sha256sum -c SHA256SUMS
   --alias local-laguna --metrics
 ```
 
-Confirm `curl -sf http://127.0.0.1:8000/v1/models` → 200.
+Confirm `curl -sf http://127.0.0.1:8000/v1/models` → 200.  
+Smoke / bench **`--model` must equal `--alias`** → `local-laguna`.
 
 6. `llama-bench` @ prompt 2048 and 8192; save stdout to `results/llama_bench_q4km.txt`.
 7. `python scripts/bench_server.py --ctx-mark 2k --ctx-mark 8k`.
-8. `python eval/agent_smoke/run_smoke.py --out results/smoke_q4km.json`.
-8b. Optional Hermes-class: `python eval/hermes_agent_smoke/run_hermes_smoke.py --out results/hermes_agent_smoke.json`.
+8. `python eval/agent_smoke/run_smoke.py --model local-laguna --out results/agent_smoke.json`.
+8b. Optional Hermes-class: `python eval/hermes_agent_smoke/run_hermes_smoke.py --model local-laguna --out results/hermes_agent_smoke.json`.
 9. Populate the tables in `SPARK.md` from those files only — never invent.
 10. Publish pack (docs + results + scripts). **Host GGUF binary only if you have a measured unique artifact**; otherwise ship digests + download commands (S5).
-11. Optional DFlash is **not** default. If tried on Spark, record with honesty — 2026-07-29 measure is **DO_NOT_PROMOTE** (`results/dflash_2026-07-29/`).
+11. Optional DFlash is **not** default. If tried on Spark, record with honesty — 2026-07-29 measure is **DO_NOT_PROMOTE** (`results/dflash_2026-07-29/`). Note: that trial used `-ngl 99` while baseline pin is `-ngl -1`; reject still holds (gen + prefill both slower).
 
 ## Personal HF publish surface
 

docs/WEIGHTS_LOCAL.md

@@ -12,9 +12,9 @@ Official: `poolside/Laguna-S-2.1-GGUF` · sha256 `a8b55c75714ea73fd90ec85de5defd
 
 Canonical serve path may stay `~/models/laguna-s-2.1/` with a symlink into this pack.
 
-## Mac ≤32G
+## Mac ≤32 GB class
 
-Full S weights: **do not** land on founder Mac. Client → Spark `:8000` only.
+Full S weights: **do not** land on ≤32 GB unified-memory Macs. Client → Spark `:8000` only.
 
 ## XS
 

docs/rename_Agentic_20260729T135518_wib.json

@@ -1,7 +1,7 @@
 {
   "event": "pack_full_rename_to_Agentic_TitleCase",
   "at_wib": "2026-07-29T13:55:18+07:00",
-  "host": "HizrianacStudio.localdomain",
+  "host": "local",
   "targets": {
     "S": {
       "display": "Laguna-S-2.1-Spark-Agentic",

eval/agent_smoke/README.md

@@ -15,7 +15,7 @@ Personal fixed tool-agent smoke for Laguna OpenAI-compatible serve on **Spark**.
 | Elapsed | **97.79 s** |
 | Model | `local-laguna` |
 | Temp | **0.0** |
-| When | 2026-07-28 ~23:55 WIB |
+| When | 2026-07-28 ~23:55 local |
 | Artifact | [`results/agent_smoke.json`](../../results/agent_smoke.json) |
 
 See also [`MEASURED.md`](../../results/MEASURED.md). Receipts include a `run_manifest` block (host, runner SHA, cases SHA, pack git HEAD).

eval/agent_smoke/cases.json

@@ -1799,11 +1799,11 @@
       "messages": [
         {
           "role": "system",
-          "content": "You are a careful tool-using coding agent. Only call tools from the provided schema. Never invent tool names. When a tool is needed, emit proper tool calls. Prefer concise answers. You must not invent org or product branding. Keep the pack label personal spark pack only."
+          "content": "You are a careful tool-using coding agent. Only call tools from the provided schema. Never invent tool names. When a tool is needed, emit proper tool calls. Prefer concise answers. You must not invent org or product branding. Keep the pack label independent spark pack only."
         },
         {
           "role": "user",
-          "content": "Write sandbox/brand.txt content 'personal spark pack' via write_file."
+          "content": "Write sandbox/brand.txt content 'independent spark pack' via write_file."
         }
       ],
       "tools": [

eval/agent_smoke/run_smoke.py

@@ -2,7 +2,7 @@
 """Fixed agent_smoke runner for Laguna OpenAI-compatible servers.
 
 Pass/fail only against the pinned cases.json — no free-form grading.
-personal fixed smoke runner
+Fixed smoke runner
 """
 
 from __future__ import annotations

eval/hermes_agent_smoke/README.md

@@ -1,4 +1,4 @@
-# hermes_agent_smoke v2
+# hermes_agent_smoke
 
 Hermes-**class** agent tool smoke for Laguna OpenAI-compatible serve.
 
@@ -6,50 +6,100 @@ Hermes-**class** agent tool smoke for Laguna OpenAI-compatible serve.
 - Name = tool-agent runtime shape (OpenAI `tools` / multi-turn repair).
 - Reuses judges + prior-arg sanitization from `eval/agent_smoke/`.
 
-## Suite
+## Suites
 
-| Field | Value |
-|-------|--------|
-| Cases | **27** |
-| Version | 2 |
-| Ship min | **24/27** (~89%) |
-| Stretch | **27/27** |
-| Headline ship bar for launch lock | still **agent_smoke 40/40** (v1) |
+| Suite | File | Cases | Ship min | Stretch | Freeze claim? |
+|-------|------|------:|---------:|--------:|:-------------|
+| **v2 (freeze)** | `cases.json` | **27** | **24/27** | **27/27** | **Yes** — launch lock artifact |
+| **v3 Layer B** | `cases_layer_b_v3.json` | **35** (=27+8) | **30/35** | **35/35** | **No** until live measured |
+
+Headline ship bar for launch lock: still **agent_smoke 40/40** (v1).
+
+v2 categories: terminal · files · web · multi_tool · multi_turn · error_repair · no_invented · browser · memory_cron · args_strict · safety
 
-Categories: terminal · files · web · multi_tool · multi_turn · error_repair · no_invented · browser · memory_cron · args_strict · safety
+v3 adds **long_horizon** (+ denser multi_turn / error_repair): tool→obs→next-tool chains after 1–2 prior hops.
 
-## Measured (live Spark)
+### Layer B new case ids (v3 only)
+
+- `long_01` … `long_05` — multi-hop continue (search→read→run, terminal→calc, write→read, empty-search recover, gather→memory)
+- `repair_03` — invalid prior JSON args repair
+- `turn_05` — prose guess then force tool
+- `turn_06` — finish second half of multi-goal after first tool obs
+
+## Measured (live Spark) — freeze bar
 
 | Field | Value |
 |-------|--------|
-| Result | **27/27 (100%)** |
-| Elapsed | **104.4 s** (post-restore); prior same-day 102.68 s |
+| Result | **27/27 (100%)** on **v2** |
+| Elapsed | **~100–104 s** |
 | Model id | `local-laguna` |
 | Base | Spark OpenAI-compatible `:8000` |
 | Temp | **0.0** |
 | Artifact | `results/hermes_agent_smoke.json` |
-| When | 2026-07-28 ~23:57 WIB (post-restore lock) |
+| When | 2026-07-28/29 (see `results/launch_lock.json`) |
 
-Meets ship_min and stretch. Does **not** replace agent_smoke 40/40 launch bar.
+## Measured (live Spark) — Layer B v3 research only
+
+| Field | Value |
+|-------|--------|
+| Result | **35/35 (100%)** on **v3** (=27+8) |
+| Elapsed | **137.56 s** |
+| Model id | `local-laguna` |
+| Weight | official `laguna-s-2.1-Q4_K_M.gguf` |
+| Path | Spark serve `127.0.0.1:8000` → SSH tunnel Mac `:18000` |
+| Temp | **0.0** |
+| Artifact (local, gitignored) | `results/hermes_agent_smoke_layer_b_v3.json` |
+| Receipt (tracked) | `layer_b_v3_live_receipt.json` |
+| When | 2026-07-29 ~18:49 local |
+
+**Not a freeze-bar move.** Do not rewrite `launch_lock.json` hermes fields from v3.
 
 ## Run
 
 ```bash
-# assumed local Spark serve with --jinja
+# Freeze / claim path — v2 only
 python eval/hermes_agent_smoke/run_hermes_smoke.py \
   --base-url http://127.0.0.1:8000/v1 \
   --model local-laguna \
+  --cases eval/hermes_agent_smoke/cases.json \
   --out results/hermes_agent_smoke.json
+
+# Layer B expansion (research / post-measure) — does not move freeze bar
+python eval/hermes_agent_smoke/run_hermes_smoke.py \
+  --base-url http://127.0.0.1:8000/v1 \
+  --model local-laguna \
+  --cases eval/hermes_agent_smoke/cases_layer_b_v3.json \
+  --out results/hermes_agent_smoke_layer_b_v3.json
+```
+
+Optional: `--limit N`, `--ids term_01,long_01`, `--temperature 0`.
+
+One-shot helper (fail-closed if serve down; does not start GPU):
+
+```bash
+./scripts/measure_layer_b_v3.sh
+# OPENAI_BASE_URL=http://<spark>:8000/v1 ./scripts/measure_layer_b_v3.sh
 ```
 
-Optional: `--limit N`, `--ids term_01,files_01`, `--temperature 0`.
+Offline judge dry-run (no server): load `agent_smoke.judge` against synthetic tool_calls (see receipt).
+
+Static result browser (local file drop): [`../smoke_viewer.html`](../smoke_viewer.html).
 
 ## Protocol honesty
 
 **one-response protocol** — tool JSON shape / multi-turn repair are **validated**, tools are **not executed**.
 Smoke temperature for claim runs: **0.0** (never 0.7 card/default for tools).
 
+Judges already support: `tool_call`, `any_of_tools`, `any_ok_tools`, `no_extra_tools`, `content_contains` + `sanitize_messages_for_server` for broken prior args.
+
 ## Claims rule
 
-Pass fraction on the card only with `results/hermes_agent_smoke.json` from a real run.
-Headline launch bar remains agent_smoke **40/40**.
+- Freeze / launch lock quotes **only** v2 `results/hermes_agent_smoke.json`.
+- Do not rewrite `launch_lock.json` hermes fields from a v3 run.
+- Headline launch bar remains agent_smoke **40/40**.
+- Pass fraction on the card only with a real run artifact.
+
+## Layer B intent
+
+Expand harness toward Hermes multi-turn / long-horizon dominance **without** touching weights or Aug 3 freeze claim surface.
+Retrain / full agentic SFT stays **parked** (separate post-freeze decision).

eval/hermes_agent_smoke/run_hermes_smoke.py

@@ -1,9 +1,9 @@
 #!/usr/bin/env python3
-"""Hermes-class agent_smoke v2 runner for Laguna OpenAI-compatible servers.
+"""Hermes-class agent_smoke runner for Laguna OpenAI-compatible servers.
 
-Reuses agent_smoke judges + prior-tool-arg sanitization.
-Suite is named hermes_agent_smoke — branding is tool-agent family shape only,
-not a Nous Research endorsement.
+Default cases.json = freeze v2 (27). Pass --cases cases_layer_b_v3.json for
+Layer B long-horizon expansion (35). Reuses agent_smoke judges + prior-tool-arg
+sanitization. Branding is tool-agent family shape only — not a Nous endorsement.
 """
 
 from __future__ import annotations

eval/smoke_viewer.html

@@ -0,0 +1,137 @@
+<!doctype html>
+<html lang="en">
+<head>
+  <meta charset="utf-8" />
+  <meta name="viewport" content="width=device-width, initial-scale=1" />
+  <title>Laguna agent smoke viewer (static)</title>
+  <style>
+    :root { color-scheme: dark; --bg:#0d1117; --card:#161b22; --fg:#e6edf3; --mut:#8b949e; --ok:#3fb950; --bad:#f85149; --line:#30363d; }
+    body { margin:0; font:14px/1.45 ui-sans-serif,system-ui,sans-serif; background:var(--bg); color:var(--fg); }
+    header { padding:16px 20px; border-bottom:1px solid var(--line); }
+    h1 { margin:0 0 6px; font-size:18px; }
+    .mut { color:var(--mut); }
+    main { padding:16px 20px 40px; max-width:1100px; }
+    .row { display:flex; flex-wrap:wrap; gap:10px; align-items:center; margin:10px 0 16px; }
+    input[type=file], button, select { background:var(--card); color:var(--fg); border:1px solid var(--line); border-radius:8px; padding:8px 10px; }
+    button { cursor:pointer; }
+    .stat { background:var(--card); border:1px solid var(--line); border-radius:10px; padding:10px 12px; min-width:110px; }
+    .stat b { display:block; font-size:18px; }
+    table { width:100%; border-collapse:collapse; }
+    th, td { text-align:left; padding:8px 10px; border-bottom:1px solid var(--line); vertical-align:top; }
+    th { color:var(--mut); font-weight:600; }
+    tr.pass td.id { color:var(--ok); }
+    tr.fail td.id { color:var(--bad); }
+    code { font:12px/1.4 ui-monospace, SFMono-Regular, Menlo, monospace; }
+    pre { white-space:pre-wrap; background:#0a0e14; border:1px solid var(--line); border-radius:8px; padding:10px; max-height:220px; overflow:auto; }
+    .disc { margin-top:18px; font-size:12px; color:var(--mut); }
+  </style>
+</head>
+<body>
+<header>
+  <h1>Laguna agent / hermes-class smoke viewer</h1>
+  <div class="mut">Static local viewer · drop a measured JSON · no network · not a Nous endorsement</div>
+</header>
+<main>
+  <div class="row">
+    <input id="file" type="file" accept="application/json,.json" />
+    <select id="filter">
+      <option value="all">all</option>
+      <option value="fail">fails only</option>
+      <option value="pass">passes only</option>
+    </select>
+    <button id="clear" type="button">clear</button>
+  </div>
+  <div class="row" id="stats"></div>
+  <div id="meta" class="mut"></div>
+  <table>
+    <thead><tr><th>id</th><th>ok</th><th>judge / note</th><th>detail</th></tr></thead>
+    <tbody id="tbody"></tbody>
+  </table>
+  <p class="disc">
+    Personal measurement helper for pack artifacts under <code>results/</code>
+    (<code>agent_smoke.json</code>, <code>hermes_agent_smoke.json</code>, Layer B v3 when live).
+    Freeze claims remain whatever <code>launch_lock.json</code> records — this page never invents scores.
+  </p>
+</main>
+<script>
+const $ = (id) => document.getElementById(id);
+let DATA = null;
+
+function casesFrom(doc) {
+  if (!doc) return [];
+  if (Array.isArray(doc.cases)) return doc.cases;
+  if (Array.isArray(doc.results)) return doc.results;
+  if (Array.isArray(doc.items)) return doc.items;
+  return [];
+}
+
+function isPass(c) {
+  if (typeof c.pass === 'boolean') return c.pass;
+  if (typeof c.ok === 'boolean') return c.ok;
+  if (typeof c.passed === 'boolean') return c.passed;
+  if (c.status) return String(c.status).toLowerCase() === 'pass';
+  return false;
+}
+
+function render() {
+  const tbody = $('tbody');
+  const stats = $('stats');
+  const meta = $('meta');
+  tbody.innerHTML = '';
+  stats.innerHTML = '';
+  meta.textContent = '';
+  if (!DATA) return;
+  const all = casesFrom(DATA);
+  const mode = $('filter').value;
+  const rows = all.filter(c => {
+    const p = isPass(c);
+    if (mode === 'fail') return !p;
+    if (mode === 'pass') return p;
+    return true;
+  });
+  const nPass = all.filter(isPass).length;
+  const nFail = all.length - nPass;
+  const mk = (label, val) => {
+    const d = document.createElement('div');
+    d.className = 'stat';
+    d.innerHTML = `<b>${val}</b><span class="mut">${label}</span>`;
+    return d;
+  };
+  stats.append(
+    mk('total', all.length),
+    mk('pass', nPass),
+    mk('fail', nFail),
+    mk('rate', all.length ? `${(100*nPass/all.length).toFixed(1)}%` : '—')
+  );
+  const bits = [];
+  for (const k of ['suite','model','base_url','temperature','elapsed_s','created_at','when','runner_sha']) {
+    if (DATA[k] != null) bits.push(`${k}=${DATA[k]}`);
+  }
+  meta.textContent = bits.join(' · ');
+  for (const c of rows) {
+    const tr = document.createElement('tr');
+    const ok = isPass(c);
+    tr.className = ok ? 'pass' : 'fail';
+    const id = c.id || c.case_id || c.name || '—';
+    const note = c.judge || c.reason || c.message || c.error || '';
+    const detail = c.detail || c.tool_calls || c.response || c.raw || c;
+    tr.innerHTML = `<td class="id"><code>${id}</code></td>
+      <td>${ok ? 'pass' : 'FAIL'}</td>
+      <td>${String(note).slice(0,240)}</td>
+      <td><pre></pre></td>`;
+    tr.querySelector('pre').textContent = typeof detail === 'string' ? detail : JSON.stringify(detail, null, 2).slice(0, 4000);
+    tbody.appendChild(tr);
+  }
+}
+
+$('file').addEventListener('change', async (e) => {
+  const f = e.target.files && e.target.files[0];
+  if (!f) return;
+  DATA = JSON.parse(await f.text());
+  render();
+});
+$('filter').addEventListener('change', render);
+$('clear').addEventListener('click', () => { DATA=null; $('file').value=''; render(); });
+</script>
+</body>
+</html>

hermes/README.md

@@ -1,7 +1,5 @@
 # Hermes-class sample client
 
-Personal sample client for local Spark serve.
-
 OpenAI-compatible client notes for **Hermes-class / tool-agent** runtimes against
 local `llama-server` serving Laguna-S-2.1 on DGX Spark.
 
@@ -18,7 +16,7 @@ local `llama-server` serving Laguna-S-2.1 on DGX Spark.
 ```text
 base_url = http://127.0.0.1:8000/v1
 api_key  = sk-local          # llama-server ignores; keep placeholder
-model    = local-laguna      # any label; server may echo file stem
+model    = local-laguna      # must match llama-server --alias
 ```
 
 ```bash
@@ -28,6 +26,16 @@ export OPENAI_MODEL=local-laguna
 python hermes/sample_client.py
 ```
 
+Remote client → Spark (no local full-S weights on ≤32 GB class machines):
+
+```bash
+export OPENAI_BASE_URL=http://<spark-host>:8000/v1
+export OPENAI_API_KEY=sk-local
+export OPENAI_MODEL=local-laguna
+```
+
+Stack-agnostic YAML sketch: [`config.example.yaml`](./config.example.yaml)
+
 ## Tool schema contract
 
 Send tools as OpenAI function tools:
@@ -59,22 +67,37 @@ Agent loop:
 3. Repeat until final `content`
 4. **Never** execute a name absent from the offered schema
 
+Multi-turn repair (invalid prior args):
+
+1. If a historical `function.arguments` string is not valid JSON, **do not** re-POST it raw
+2. Replace with a sanitized object holding `_invalid_json_arguments` + `_raw` (see `eval/agent_smoke/run_smoke.py`)
+3. Or drop/rebuild the broken turn and ask the model to re-issue the tool call
+
 ## Jinja / chat template
 
 Serve with `--jinja` on poolside `laguna` llama.cpp so tool call formatting
 matches the Laguna chat template. If tool_calls parse soft-fail, check server
 log for template warnings before blaming the model.
 
-## Escapes
+Captain serve flags (must match last-green):
+
+```text
+-c 8192 -ngl -1 --jinja -fa on --alias local-laguna --parallel 1
+```
+
+## Failure → fix matrix
 
 | Symptom | Check |
 |---------|--------|
 | empty tool_calls, prose only | temp→0, strengthen system “use tools”, confirm `--jinja` |
 | invented tool names | smoke case `no_invented_tools`; tighten system prompt |
-| JSON args invalid | repair turn in agent_smoke `error_repair` |
-| 404 /v1/... | base_url must include `/v1` |
+| JSON args invalid | repair turn in agent_smoke `error_repair`; sanitize priors |
+| HTTP 500 on follow-up turn | prior tool args not valid JSON — sanitize before re-send |
+| 404 /v1/... | `OPENAI_BASE_URL` must include `/v1` |
+| model id mismatch | request model must match `--alias` (`local-laguna`) |
+| slow / OOM agent loops | ctx 8192 agent day-to-day; do not jump to 256k without FSM |
 
-## Hermes-class smoke v2
+## Hermes-class smoke v2 (freeze bar companion)
 
 ```bash
 python eval/hermes_agent_smoke/run_hermes_smoke.py \
@@ -84,6 +107,30 @@ python eval/hermes_agent_smoke/run_hermes_smoke.py \
   --out results/hermes_agent_smoke.json
 ```
 
-27 fixed cases (terminal/files/web multi-turn + repair). See `eval/hermes_agent_smoke/README.md`.
-**one-response protocol** — tools validated, not executed. Claim temp **0.0**.
+27 fixed cases (terminal/files/web multi-turn + repair).  
+See `eval/hermes_agent_smoke/README.md`.  
+**One-response protocol** on most cases — tools validated, not executed.  
+Claim temp **0.0**. Measured tip: **27/27**.
+
+Does **not** replace the launch-bar **agent_smoke 40/40**.
+
+## Layer B research suite (optional)
+
+```bash
+python eval/hermes_agent_smoke/run_hermes_smoke.py \
+  --base-url http://127.0.0.1:8000/v1 \
+  --model local-laguna \
+  --cases eval/hermes_agent_smoke/cases_layer_b_v3.json \
+  --out results/hermes_agent_smoke_layer_b_v3.json
+```
+
+35 cases (= v2 27 + 8 long-horizon / denser repair).  
+Live research measure **2026-07-29:** **35/35** in **137.56s** on official Q4 (temp 0).  
+Local artifact: `results/hermes_agent_smoke_layer_b_v3.json` (gitignored).  
+Tracked receipt: `eval/hermes_agent_smoke/layer_b_v3_live_receipt.json`.  
+DOES **not** move freeze lock numbers (v2 stays **27/27**).
+
+One-shot helper: `./scripts/measure_layer_b_v3.sh`  
+(If Spark serves loopback-only, tunnel: `ssh -L 18000:127.0.0.1:8000 spark`.)
+
 Not a Nous endorsement.

hermes/config.example.yaml

@@ -0,0 +1,50 @@
+# Example OpenAI-compatible provider snippet for Hermes-class / tool-agent runtimes
+# against local Laguna-S-2.1 on DGX Spark.
+#
+# Language:
+#   - "Hermes-class" = OpenAI chat.completions + tools/tool_calls shape only
+#   - NOT a Nous Research endorsement or Hermes Agent product config
+#   - Map keys to your agent stack's real schema (names differ by product)
+
+provider:
+  kind: openai_compatible
+  # MUST include /v1 — bare :8000 returns 404 on /chat/completions paths
+  base_url: "http://127.0.0.1:8000/v1"
+  # If Spark binds 0.0.0.0: Tailscale "http://100.98.213.2:8000/v1"
+  # If Spark binds 127.0.0.1 only: ssh -L 18000:127.0.0.1:8000 spark → "http://127.0.0.1:18000/v1"
+  api_key: "sk-local" # llama-server ignores auth; keep a placeholder
+  model: "local-laguna" # must match llama-server --alias
+  timeout_s: 300
+  headers:
+    User-Agent: "laguna-spark-hermes-sample/0.1"
+
+serve_pin:
+  # Captain flags from results/LAST_GREEN_PIN.md / MEASURED.md
+  engine: "poolsideai/llama.cpp@04b2b72 (branch laguna)"
+  quant: "poolside/Laguna-S-2.1-GGUF:laguna-s-2.1-Q4_K_M.gguf"
+  q4_sha256: "a8b55c75714ea73fd90ec85de5defdc0b8d88ca0ad2108343cdd8fc22f7583e4"
+  flags:
+    - "--ctx-size 8192"
+    - "-ngl -1"
+    - "--jinja" # required for Laguna tool/chat template
+    - "-fa on"
+    - "--alias local-laguna"
+    - "--parallel 1"
+
+sampling_agent:
+  # Agent smoke lock uses temp 0.0
+  temperature: 0.0
+  top_p: 1.0
+  max_tokens: 512
+  tool_choice: auto
+
+client_rules:
+  - "Never execute a tool name absent from the offered schema"
+  - "function.arguments must be a JSON object string; repair/sanitize before re-send"
+  - "Append role=tool results, then continue the same multi-turn thread"
+  - "Prefer short final content after tools resolve"
+
+smoke:
+  agent_bar: "eval/agent_smoke/run_smoke.py  # 40 cases · launch bar"
+  hermes_v2: "eval/hermes_agent_smoke/run_hermes_smoke.py  # 27 cases · measured tip"
+  hermes_layer_b_v3: "scripts/measure_layer_b_v3.sh  # research 35/35 live tip; not freeze bar"

prompts/_mirror/README-SPARK-HIZRIANRAZ.md

@@ -1,11 +0,0 @@
-# Spark dual-repo analysis redouts
-
-Live copies for external tools live on Spark:
-
-```
-~/hizrianraz/prompts/           ← paste these
-~/hizrianraz/Laguna-S-2.1-Spark-Agentic
-~/hizrianraz/Laguna-XS-2.1-Mac-Agentic
-```
-
-Analyze **both** trees. Never merge product stories.

prompts/chatgpt-analyze-laguna-s-2026-07-29.md

@@ -1,122 +0,0 @@
-# ChatGPT analysis prompt — Laguna-S-2.1-Spark-Agentic (2026-07-29)
-
-Copy everything below the line into ChatGPT (o3 / GPT-5 Thinking preferred).  
-Paste the listed file bodies as attachments or follow-up messages. Do **not** invent numbers.
-
----
-
-## Role
-
-You are an independent technical auditor for a **personal** Hugging Face model pack:
-
-`hizrianraz/Laguna-S-2.1-Spark-Agentic`
-
-Not Ainfera company IP. Spark-only full weights. Founder: Hizrian Raz.
-
-## Mission
-
-Analyze freeze/launch readiness **after** an optional DFlash (draft speculative) experiment that **failed** promotion.
-
-Return hard judgments, not cheerleading.
-
-## Hard facts (only these numbers are authorized)
-
-| Item | Value |
-|------|-------|
-| Headline quant | official Poolside `laguna-s-2.1-Q4_K_M.gguf` |
-| Q4 sha256 | `a8b55c75714ea73fd90ec85de5defdc0b8d88ca0ad2108343cdd8fc22f7583e4` |
-| Host | DGX Spark GB10 |
-| Engine | poolsideai/llama.cpp `04b2b72` |
-| agent_smoke | **40/40** · 84.86s · temp 0.0 |
-| hermes_agent_smoke v2 | **27/27** · 100.1s |
-| gen128 | **21.47 tok/s** (quote this) |
-| Prefill OK marks | 2k **1.597s** · 8k **4.78s** |
-| Measure tip (smoke/weights) | git `bf82eab` — must not be confused with later docs tip |
-| DFlash gen128 | **15.286 tok/s** (−28.8% vs 21.47) · **DO_NOT_PROMOTE** |
-| DFlash prefill | 2k 2.253s · 8k 5.556s (worse than baseline) |
-| IQ3 pointer (not headline) | Unsloth UD-IQ3_S **38/40** |
-| Freeze clock | **2026-08-02 18:00 WIB** |
-| Public go-live | **2026-08-03 12:00 WIB** |
-| Freeze gate | **FILLED** 2026-07-29 (model card + lock set) |
-| Live baseline after DFlash teardown | HEALTH_OK + chat content `OK` · fingerprint `b1-04b2b72` |
-| diy_gguf | false |
-| weight host | Spark-only |
-| founder Mac | client-only, no local S weights |
-| XS pack | parallel track, **not** inside S freeze |
-| branding | personal-only |
-| public promo before launch | **false** |
-
-## Locks you must not break in recommendations
-
-1. No freeze before 2026-08-02 18:00 WIB  
-2. No public promo before 2026-08-03 12:00 WIB  
-3. No claiming DFlash speedup  
-4. No inventing higher tok/s or smoke scores  
-5. No full Laguna mobile claims  
-6. No mixing XS into S freeze artifact  
-7. Docs tip may move; **measure tip stays bf82eab** until weights/smoke re-run  
-
-## Attachments to analyze (paste contents)
-
-Primary:
-
-1. `README.md` (model card)  
-2. `LAUNCH_AUG3.md`  
-3. `results/MEASURED.md` + `results/measured.json`  
-4. `results/launch_lock.json`  
-5. `results/LAST_GREEN_PIN.md`  
-6. `results/dflash_2026-07-29/STATUS.md` + `measured.json`  
-7. `results/baseline_rehealth_2026-07-29.md`  
-8. `results/stranger_path_dry_2026-07-29.md`  
-9. `docs/REPRODUCE.md`  
-10. `research/quant-comparison-scoreboard-2026-07-28.md`  
-11. `results/freeze_notes_2026-07-29.md`  
-12. `results/freeze_gate_model_card_lock_set_2026-07-29.json`  
-13. `results/three_jury_post_dflash_2026-07-29.json`  
-14. `results/hf_publish.json` (last HF docs snapshot)  
-
-## Questions (answer all)
-
-### A. Integrity
-1. Are any two docs contradicting on gen128 / smoke / sha256 / engine?  
-2. Is DFlash rejection scientifically fair (same host, weight, engine family)?  
-3. Could a reader confuse docs tip with measure tip and allege regression?
-
-### B. Ship story
-4. What is the single tightest sentence for the HF card opening after Aug 3?  
-5. Which 5 phrases are **banned** (overclaim risk)?  
-6. Is stranger-path documentation enough for a non-founder to reproduce headline serve?
-
-### C. Risk → freeze
-7. Top 5 residual risks ranked by freeze-day severity.  
-8. What must be true on clean pushed HEAD at freeze (checklist).  
-9. Anything that looks “done” but is only claimed?
-
-### D. Competitive honesty
-10. Versus Poolside/Unsloth/Bartowski GGUFs: what is our real wedge (agent evidence + Spark digest bind) vs vanity (likes/dl)?  
-11. Should IQ3 appear above or below the fold?
-
-### E. Decision
-12. Verdict: **SHIP_AS_PLANNED** / **SHIP_WITH_NITS** / **SLIP_FREEZE** — pick one.  
-13. If NITS: max 5 bullets with exact file paths to fix before 08-02.  
-14. If SLIP: what evidence failed and minimum fix.
-
-## Output format (strict)
-
-```
-VERDICT: ...
-ONE_LINE_STORY: ...
-BANNED_PHRASES:
-- ...
-CONTRADICTIONS:
-- ... or NONE
-RISKS_TOP5:
-1. ...
-FREEZE_CHECKLIST:
-- [ ] ...
-NITS:
-- path: ...
-CONFIDENCE: high|med|low + why
-```
-
-No marketing. No new benchmarks. Cite paths.

prompts/chatgptwork-laguna-s-plus-xs-2026-07-28.md

@@ -1,84 +0,0 @@
-# ChatGPT Work — Laguna S pack + XS Mac path (analysis only)
-
-## Role
-You are an external technical analyst. Advisory only.
-
-## Hard denials
-- No GPU, no deploy, no HF publish, no company-gate vault write, no git commit for the founder
-- Do not treat this output as ship clearance
-- Do not invent measured numbers
-- Do not treat sha-verified disk presence as agent_smoke
-
-## Two tracks (never merge)
-
-### Track A — Laguna-S-2.1-Spark-Agentic pack (stand-behind)
-- Personal pack HF: https://huggingface.co/hizrianraz/Laguna-S-2.1-Spark-Agentic
-- Local: `~/code/hizrianraz/Laguna-S-2.1-Spark-Agentic`
-- Mirrors: Synology `/Volumes/home/hizrianraz/Laguna-S-2.1-Spark-Agentic` · Spark `~/hizrianraz/Laguna-S-2.1-Spark-Agentic`
-- Base: poolside/Laguna-S-2.1 (118B-A8B MoE)
-- Weight host lock: **DGX Spark only**
-- Founder Mac ≤32G: **client only** for full S — no local full-S weights
-- Headline quant: official Q4_K_M  
-  sha256 `a8b55c75714ea73fd90ec85de5defdc0b8d88ca0ad2108343cdd8fc22f7583e4`
-- Spark models placement (gitignored symlink into pack): verified same sha
-- Measured (Spark, 2026-07-28): agent_smoke **40/40**, hermes-class v2 **27/27**, ~**21 t/s**
-- IQ3_S third-party pointer on Spark: **38/40** — not headline; restore Q4 after alt
-- Launch: **2026-08-03 12:00 WIB** · freeze **2026-08-02 18:00**
-- Branding: personal-only · no public promo before launch
-- DIY GGUF: false unless measured win
-- Three-jury post-placement: **all_go** for analysis continue only
-
-### Track B — Laguna-XS-2.1-Mac-Agentic ≤32G candidate (parallel, unmeasured)
-- Separate pack: `~/code/hizrianraz/Laguna-XS-2.1-Mac-Agentic` (+ Synology/Spark trees)
-- Base: https://huggingface.co/poolside/Laguna-XS-2.1  
-  sha `205dc65dd4bda946c50da6b7522b215734fa107b`
-- Shape: **33B total · 3B active** MoE (NOT a quant of S)
-- Official GGUF: https://huggingface.co/poolside/Laguna-XS-2.1-GGUF
-  - Q4_K_M ≈ **18.882 GiB** — **on Synology pack models/** sha verified  
-    `1ac7079101fca5a6df8c5a7523a3c30ea7d1c0e4b1258090e7d6d4039287f6cb`
-  - BF16 ≈ **62.334 GiB** — non-fit founder Mac
-- Poolside “Mac with 36 GB” vs founder Studio **exactly 32G** — tight, unproven
-- Engine risk: llama.cpp Laguna path PR ggml-org/llama.cpp#25165 / poolside fork
-- **Zero** founder same-harness agent_smoke on Mac for XS
-- Forbidden labels: “S on Mac”, “S-lite”, “S-Mac=XS”, averaging S metrics with XS
-
-## Inputs to analyze
-1. Repo trees + dual roadmap docs + three_jury + weight_placement JSON
-2. HF card S pack (live) + poolside S/S-GGUF/XS/XS-GGUF
-3. launch_lock.json freeze/launch bars only for Track A
-4. XS measured.json still `unmeasured`
-
-## Tasks
-1. **Integrity audit (Track A):** overclaims, host-label blur, missing sha, freeze risks, promo creep
-2. **Roadmap gap list:** what must be true by freeze 2026-08-02 18:00 for S pack honesty only
-3. **XS Mac path (Track B):** realistic 32G load envelope; pull-from-Synology order; engine gates; measurement checklist; what not to claim
-4. **Separation check:** any line that implies full S runs on founder Mac or that disk-sha = smoke
-5. **Green-to-continue definition:** state clearly that external GO = continue honest freeze prep for S + optional XS research prep — not XS ship and not public promo
-6. **Actionable cuts:** max 10 wording-only edits — no scope expansion
-
-## Required output format
-### A. Executive
-- 5 bullets max
-
-### B. Track A (S/Spark) findings
-- table: finding | severity | evidence | fix
-
-### C. Track B (XS/Mac) findings
-- table: finding | severity | evidence | next measure step
-- explicit: proven vs candidate
-
-### D. Must-not list
-- bullets founder should keep locked
-
-### E. Open questions
-- ≤8, ranked
-
-### F. Measurement checklist for XS Mac
-- ordered, copy-pasteable
-
-### G. Verdict
-- S roadmap green to continue freeze work: yes|no + why
-- XS measure clearance: **no** unless evidence changes
-- Safe for founder to paste external output back without authority laundering: yes|no
-
-Stop after analysis. No ship decision language.

prompts/claude-analyze-laguna-s-2026-07-29.md

@@ -1,148 +0,0 @@
-# Claude analysis prompt — Laguna-S-2.1-Spark-Agentic (2026-07-29)
-
-Copy everything below the line into Claude (Opus 4.x / Sonnet Thinking preferred).  
-Project-knowledge or paste the listed files. **Refuse any number not in the fact table or attached files.**
-
----
-
-## System stance
-
-You are **Claude as adversarial launch reviewer** for a personal HF pack, not a co-author.
-
-Pack: `hizrianraz/Laguna-S-2.1-Spark-Agentic`  
-Persona: blunt, path-cited, ADHD-scannable. Short lines. Tables OK.
-
-## Context in one breath
-
-Founder measured official Laguna-S 2.1 Q4_K_M on DGX Spark with poolside llama.cpp `04b2b72`:
-
-- agent_smoke **40/40**
-- hermes v2 **27/27**
-- gen128 **21.47 t/s**
-
-Optional DFlash draft path measured **15.286 t/s** gen128 → **DO_NOT_PROMOTE**.  
-Baseline serve restored live (HEALTH_OK + chat OK).  
-Freeze gate already **FILLED**. Clocks locked: freeze **2026-08-02 18:00 WIB**, go-live **2026-08-03 12:00 WIB**.  
-No public promo early. Spark-only weights. Personal brand. XS is a separate parallel track.
-
-## Authorized fact table
-
-| Field | Authorized value |
-|-------|------------------|
-| Weight | `laguna-s-2.1-Q4_K_M.gguf` official Poolside |
-| sha256 | `a8b55c75714ea73fd90ec85de5defdc0b8d88ca0ad2108343cdd8fc22f7583e4` |
-| Engine | `04b2b72` |
-| Measure git tip | `bf82eab` |
-| agent_smoke | 40/40 · 84.86s |
-| hermes | 27/27 · 100.1s |
-| gen128 | 21.47 t/s |
-| prefill 2k/8k | 1.597s / 4.78s |
-| DFlash gen128 | 15.286 t/s (−28.8%) |
-| DFlash 2k/8k | 2.253s / 5.556s |
-| IQ3_S pointer | 38/40 not headline |
-| Live after restore | HEALTH_OK · reply `OK` · fp `b1-04b2b72` |
-
-## What to read
-
-1. Model card `README.md`  
-2. `LAUNCH_AUG3.md`  
-3. `docs/REPRODUCE.md` + `SPARK.md`  
-4. `results/MEASURED.md`, `measured.json`, `launch_lock.json`  
-5. `results/LAST_GREEN_PIN.md`  
-6. Entire `results/dflash_2026-07-29/` summary files  
-7. `results/baseline_rehealth_2026-07-29.md`  
-8. `results/stranger_path_dry_2026-07-29.md`  
-9. `results/freeze_notes_2026-07-29.md` + freeze gate JSON  
-10. `results/three_jury_post_dflash_2026-07-29.json`  
-11. Scoreboard `research/quant-comparison-scoreboard-2026-07-28.md`  
-12. `results/hf_publish.json`
-
-## Analysis jobs (do in order)
-
-### 1) Contradiction hunt
-Diff headlines across card / MEASURED / lock / pin / scoreboard.  
-List every mismatch (even stylistic number rounding like 21.016 vs 21.47 — explain which is authoritative and why).
-
-### 2) DFlash post mortem
-Explain failure mode likely causes *from evidence only* (draft overhead, prefill tax, settings).  
-Confirm DO_NOT_PROMOTE is mandatory.  
-List 3 false narratives a hype reader might invent; kill each with a citation.
-
-### 3) Stranger reproduce test
-With only public docs, outline exact terminal steps.  
-Flag any missing flag, path, or alias.  
-Pass/fail “15-minute competent stranger.”
-
-### 4) Freeze-path PERT
-Build a day-by-day from now → 08-02 freeze → 08-03 launch.  
-Mark critical path vs optional polish.  
-Call out work that is **docs-tip only** vs **measure-tip**.
-
-### 5) Claim surface
-Draft:
-
-- 1 footer-safe claim line  
-- 1 bold title-safe claim line  
-- 1 tweetsafe line  
-
-Each ≤140 chars, zero banned collocations (no “fastest”, no “only”, no DFlash win, no mobile full Laguna).
-
-### 6) Final verdict
-
-`PASS` | `PASS_WITH_NITS` | `HOLD`
-
-Rules:
-
-- HOLD only if a lock is broken or a number is false  
-- PASS_WITH_NITS if story holds but files need cleanup  
-- Include **disconfirming view**: strongest reason this launch could still embarrass on Aug 3  
-
-## Output skeleton (mandatory)
-
-```
-## Verdict
-...
-
-## Authoritative numbers
-(table)
-
-## Contradictions
-- ...
-
-## DFlash
-decision: DO_NOT_PROMOTE
-reasons:
-false_narratives_killed:
-- ...
-
-## Stranger path
-result: PASS|FAIL
-gaps:
-- ...
-
-## Freeze schedule
-| Day | Must | Optional |
-...
-
-## Claims
-footer: "..."
-title: "..."
-tweet: "..."
-
-## Nits (max 7)
-1. file — change
-
-## Disconfirming view
-...
-
-## Founder decision needed
-NONE or single sentence
-```
-
-## Explicit refusals
-
-- Do not suggest DIY quant as launch path  
-- Do not suggest hosting weights off Spark as headline  
-- Do not recently invent evals  
-- Do not merge XS into this freeze  
-- Do not move clocks earlier  

prompts/claudecowork-laguna-s-plus-xs-2026-07-28.md

@@ -1,86 +0,0 @@
-# Claude Cowork — Laguna S pack + XS Mac path (analysis only)
-
-## System posture
-You are a skeptical founding-engineer peer. Pressure-test claims. Prefer “not proven” over polish.
-
-## Absolute constraints
-- Advisory analysis only
-- No deploy, no GPU spin instructions as if already run, no HF upload, no vault/company-gate writes
-- Never mint fake agent_smoke / tok/s
-- Never collapse Laguna-**S** and Laguna-**XS** into one product story
-- Never treat weight **sha** as runtime measure
-
-## Locked facts (ground truth unless live probe disproves)
-
-### Track A — S / Spark pack
-- Local: `~/code/hizrianraz/Laguna-S-2.1-Spark-Agentic`
-- Mirrors: Synology `/Volumes/home/hizrianraz/Laguna-S-2.1-Spark-Agentic` · Spark `~/hizrianraz/Laguna-S-2.1-Spark-Agentic`
-- HF: `hizrianraz/Laguna-S-2.1-Spark-Agentic`
-- Full S weights host: **Spark only** (`weight_host=dgx-spark-only`)
-- Founder Mac ≤32G: full S = **client → Spark :8000**, not local GGUF
-- Default GGUF: `poolside/Laguna-S-2.1-GGUF` `laguna-s-2.1-Q4_K_M.gguf`
-  - sha256 `a8b55c75714ea73fd90ec85de5defdc0b8d88ca0ad2108343cdd8fc22f7583e4`
-- Spark pack placement: `~/hizrianraz/Laguna-S-2.1-Spark-Agentic/models/laguna-s-2.1-Q4_K_M.gguf` → symlink to `~/models/laguna-s-2.1/…` (sha verified)
-- Measured Spark 2026-07-28: agent_smoke **40/40** · hermes_agent_smoke **27/27** · ~**21 t/s**
-- IQ3_S Unsloth pointer on Spark: **38/40** · not default · restore Q4 after alt
-- Freeze **2026-08-02 18:00 WIB** · Launch **2026-08-03 12:00 WIB**
-- Three-jury (Lórien + Manwë appellate post-placement): **all_go** for analysis continue — see `results/three_jury_post_placement_2026-07-28.json`
-
-### Track B — XS / Mac candidate (separate 33B-A3B)
-- Separate pack: `~/code/hizrianraz/Laguna-XS-2.1-Mac-Agentic` (+ Synology + Spark tree mirrors)
-- Upstream: `poolside/Laguna-XS-2.1` @ `205dc65dd4bda946c50da6b7522b215734fa107b`
-- Official GGUF Q4_K_M **18.882 GiB** on **Synology pack**  
-  `…/Laguna-XS-2.1-Mac-Agentic/models/Laguna-XS-2.1-Q4_K_M.gguf`  
-  sha256 `1ac7079101fca5a6df8c5a7523a3c30ea7d1c0e4b1258090e7d6d4039287f6cb` **verified**
-- BF16 ~62.334 GiB = Mac non-fit
-- Poolside card “36 GB Mac” vs founder **32G** Studio → tight candidate only
-- **No** founder same-harness agent_smoke / t/s yet (`results/measured.json` = unmeasured)
-- Engine may need poolside llama.cpp `laguna` / PR #25165
-- Weights gitignored (`models/`, `*.gguf`)
-
-## Work the full surface
-Cross-check:
-- README.md matrix language (both packs)
-- results/MEASURED.md, launch_lock.json, agent_smoke / hermes headers, weight_placement_2026-07-28.json
-- results/three_jury_post_placement_2026-07-28.json + prior lorien three_jury JSON
-- research/* device matrix, xs mac fit, post-freeze path, dual roadmap
-- Live HF: S pack + poolside S + S-GGUF + XS + XS-GGUF file lists
-
-## Critique angles
-1. Any sentence still readable as “our Mac runs full Laguna-S”?
-2. Does scoreboard blur IQ3_S / community rows into founder-Mac claims?
-3. Does XS disk-on-Synology create freeze dilution risk for S?
-4. Smallest honest XS measurement ladder on 32G before any public sentence?
-5. Counter-case: when drop Mac-local XS and stay client-only forever?
-6. Is “repo + roadmap green to continue” correctly **not** the same as XS CLEAR?
-
-## Output schema
-```
-## Verdict strip
-- S pack honesty: GREEN | YELLOW | RED
-- XS separation: GREEN | YELLOW | RED
-- Placement honesty (weights gitignored / host labels): GREEN | YELLOW | RED
-- Ready for external paste without edit: yes|no
-- Roadmap green to continue freeze work (S only): yes|no
-- XS measure clearance: no (fixed unless new smoke JSON)
-
-## Counter-case (what founder is not seeing)
-- 3 bullets
-
-## Track A (S) patch list
-1. file:line-or-section — change — why
-
-## Track B (XS) measure plan
-- preconditions
-- commands outline (no claim execution happened)
-- pass/fail gates
-- when to write results
-
-## Kill criteria
-- abort XS track conditions
-
-## Questions only founder can answer
-- max 5
-```
-
-End. No cheering. Disconfirm first.

research/Laguna-XS-2.1-Mac-Agentic-fit-2026-07-28.md

@@ -1,110 +0,0 @@
-# Laguna-XS-2.1 · founder Mac ≤32G fit track
-
-Status: **prepared research** (2026-07-28).  
-Awaiting three-jury on prepared pack, then founder gates.  
-**Not** the Aug 3 S pack headline. **Not** a measured ship claim.
-
-Personal founder track only. Dual roadmap → [`roadmap-dual-s-and-xs-2026-07-28.md`](./roadmap-dual-s-and-xs-2026-07-28.md)
-
-## Bottom line
-
-| Track | Model | Host | Status |
-|-------|--------|------|--------|
-| **A · stand-behind** | Laguna-**S**-2.1 ~118B MoE Q4_K_M | **DGX Spark only** | measured 40/40 + hermes 27/27 |
-| **B · Mac local candidate** | Laguna-**XS**-2.1 33B-A3B Q4_K_M | founder Mac ≤32G | **disk candidate** · **0 local smoke** |
-
-Full **S** stays Spark-only. Mac is **not** “S via XS”.
-
-## Live HF probe (2026-07-28)
-
-### Base
-- `poolside/Laguna-XS-2.1`
-- sha: `205dc65dd4bda946c50da6b7522b215734fa107b`
-- Shape (card): **33B total · 3B active/token · MoE**
-- Poolside claim: “compact enough to run on a Mac with **36 GB** of RAM”
-- Founder Studio coin: **32G** exactly (`hw.memsize` 34359738368) → **tighter than vendor 36G**
-
-### Official GGUF
-- `poolside/Laguna-XS-2.1-GGUF` sha `1a37c0a5fb8c7a18e6106decb6be6327d1b63fa6`
-
-| File | Bytes | GiB | sha256 | Founder Mac ≤32G |
-|------|------:|----:|--------|------------------|
-| `Laguna-XS-2.1-Q4_K_M.gguf` | 20274300032 | **18.882** | `1ac7079101fca5a6df8c5a7523a3c30ea7d1c0e4b1258090e7d6d4039287f6cb` | **disk candidate** (runtime unproven) |
-| `Laguna-XS-2.1-BF16.gguf` | 66930226304 | **62.334** | `d1d99abe30c37749ec1b1ae2681cec74caa550ac6243d2ea70b61b1ff9d187ca` | **non-fit** |
-
-Upstream recommended default: **Q4_K_M**.
-
-Safetensors base ~14 shards ≈ **~67G** class — non-fit on ≤32G.
-
-### Engine dependency
-- llama.cpp Laguna support: PR `ggml-org/llama.cpp#25165`
-- Prefer `poolsideai/llama.cpp` branch `laguna` on Mac Metal — **do not assume stock brew llama runs XS**
-- Ollama (`ollama pull laguna-xs-2.1`) = alt route until smoke
-
-## Founder Mac envelope (Studio live)
-
-| Fact | Value |
-|------|-------|
-| Model | Mac14,13 · M2 Max |
-| RAM | **32G** exactly |
-| System Data free (prep) | ~40 Gi · 91% used — **prefer WD_BLACK for pull** |
-| WD_BLACK free | ~1.7 Ti |
-| Full S Q4 | ~96G · **non-fit** |
-| Full S IQ3_S pointer | ~48G · **non-fit** as local weight |
-| XS Q4_K_M | ~18.9G disk · **candidate** |
-
-Runtime RAM ≫ disk. Leave OS + browser + Hermes + KV headroom.  
-**Disk-fit ≠ load-fit ≠ agent-smoke-fit.**
-
-## Prep lock
-
-`results/xs_mac_track_lock.json` — status `prepared_research`
-
-## Honesty rules
-
-**Say**
-- “XS Q4_K_M is the official ~19G Mac-class GGUF for Laguna-**XS**-2.1”
-- “Founder Mac client still talks to Spark for full **S**”
-- “XS is a **separate** 33B-A3B model, not a quant of S”
-
-**Do not say**
-- “Laguna runs on Mac” without naming **XS**
-- “S on Mac via XS”, “S-lite”, “S-Mac”
-- Any t/s, agent_smoke, or Hermes pass for XS until same-harness JSON
-- Poolside SWE-bench / Terminal-Bench as founder measured bar
-- BF16 / safetensors base as Mac-local
-
-## Measurement gate (before any claim)
-
-1. Working engine on founder Mac · record commit
-2. `sha256` matches lock after download
-3. Load proof: RSS + free RAM after load · ctx used
-4. Same `eval/agent_smoke` 40 (or named subset + disclaimer)
-5. Optional hermes-class 27
-6. Host labeled **Mac Studio 32G** (never as Spark)
-7. Fail list + separate delta table (not diluted into S headline)
-8. Founder go before any public/pack claim move
-
-Results dir when measured: `results/sku_xs-official-q4km/`
-
-## Suggested first pull (NOT run — needs founder go)
-
-```bash
-# ~19G → external SSD preferred
-huggingface-cli download poolside/Laguna-XS-2.1-GGUF \
-  Laguna-XS-2.1-Q4_K_M.gguf \
-  --local-dir "/Volumes/HFR WD_BLACK SN850X/models/laguna-xs-2.1"
-shasum -a 256 "/Volumes/HFR WD_BLACK SN850X/models/laguna-xs-2.1/Laguna-XS-2.1-Q4_K_M.gguf"
-# expect 1ac7079101fca5a6df8c5a7523a3c30ea7d1c0e4b1258090e7d6d4039287f6cb
-```
-
-Do **not** pull BF16 on this Mac.
-
-## Relation to Aug 3 S pack
-
-| Item | Owner |
-|------|--------|
-| `hizrianraz/Laguna-S-2.1-Spark-Agentic` freeze/launch | **Roadmap A only** |
-| XS Mac notes | this file + `xs_mac_track_lock.json` |
-| multi_device_track | S community pointers; XS is **extra parallel** |
-| XS may delay S freeze | **no** |

research/agentic-quant-hermes-hf-trending-2026-07-28.md

@@ -1,234 +0,0 @@
-# Deep research — agentic quant + Hermes fit + HF trending
-
-Date: 2026-07-28 18:47 WIB  
-Scope: personal `hizrianraz/Laguna-S-2.1-Spark-Agentic` pack only  
-Locks: launch_lock `diy_gguf=false`, official Q4_K_M only, min smoke 38/40, freeze 2026-08-02 18:00 WIB, launch 2026-08-03 12:00 WIB
-
-## Bottom line
-
-1. **Power for agents is not “a wilder GGUF.”** On Laguna-S-2.1 the competitive edge that survives scrutiny is: **official Poolside Q4_K_M (signal path protected) + fixed Hermes-shaped smoke + Spark-measured serve recipe.** DIY re-quants lose the launch lock unless they beat official by ≥10% on smoke or tok/s with full method disclosure.
-2. **To be the most useful *agentic* Laguna pack for Hermes-class runtimes:** harden tool-call JSON validity, multi-turn repair, long-horizon multi-tool loops, and publish a Hermes local wiring playbook. That beats shipping another who-cares Q4.
-3. **HF “top 10 trending” globally is not a realistic target for a personal digests-only spark pack.** Base `poolside/Laguna-S-2.1` is already on live trending. Realistic win: **#1 personal Spark + Hermes-class Laguna pack**, and rank inside the **Laguna-S-2.1 GGUF/quant search cluster** (today Unsloth leads third-party GGUF at 222 likes / 130k dl). Absolute global top-10 needs mass org reach or a novel binary people pull (not our S5 surface).
-
-## Live field (measured 2026-07-28)
-
-### Our surface
-| Asset | State |
-|-------|--------|
-| HF `hizrianraz/Laguna-S-2.1-Spark-Agentic` | **0 likes, 0 downloads**, public, created today |
-| Stand-behind weight | official `laguna-s-2.1-Q4_K_M.gguf` sha `a8b55c75…` |
-| agent_smoke v1 | **40/40 (100%)** · 97.25s (2026-07-28 19:00 WIB); prior 38/40 was harness not weights |
-| Claim style | personal measurements · digests · no bare GGUF re-upload |
-
-### Laguna cluster (likes-ranked sample)
-| Likes | DL | Id | Role |
-|------:|---:|----|------|
-| 777 | 67k | poolside/Laguna-S-2.1 | base (trending) |
-| 222 | 130k | unsloth/Laguna-S-2.1-GGUF | Dynamic third-party GGUF leader |
-| 157 | 90k | poolside/Laguna-S-2.1-GGUF | **official GGUF (our pin)** |
-| 150 | 181k | poolside/Laguna-S-2.1-NVFP4 | official NVFP4 |
-| 18 | 1.8k | jcbtc/…StrixKVSpine-V4-GGUF | AMD agentic spine + HermesAgent-20 claim |
-| 13–12 | ~1–28k | CRACK / APEX / IQ4 community | mix lore / entropy / IQ |
-
-### Global trending reality check
-Live trending head includes org giants (Kimi-K3 ~7.4k likes, Unlimited-OCR, GLM-5.2, Inkling…). A personal pack does **not** enter that headroom via card polish alone.
-
-## What “most powerful for agentic / Hermes” actually means
-
-### Layer A — Weight / quant (mostly already decided)
-Official Poolside Q4_K_M recipe (from their GGUF card):
-- Routed experts: **Q4_K with imatrix**
-- Signal path (attention, shared experts, embeddings): kept **Q8_0** (not smashed)
-- Ships `laguna-s-2.1.imatrix` + recommended **llama-server --jinja**
-- Optional DFlash draft for speculative decoding
-- Context: GGUF metadata 256K recommended; 1M YaRN override available with quality caveats
-
-Unsloth Dynamic 2.0 preaching:
-- Selective layer precision; strong on Aider Polyglot / KL / MMLU-style anchors
-- Actively patches **tool-calling chat templates** on newer models
-- Wins downloads by **shipping many sizes** + brand distribution — not by Hermes-specific agent loops
-
-Community agentic spine (Chadrock ROCmFP4 example):
-- **Tensor protection** of expert-down “spine” + attention Q/O paths
-- Claims on retained benches: Tool-Eval disputed-19 **70.18% vs 54.39%** baseline; **HermesAgent-20 77/100 vs 71**
-- Hardware-specific (Strix Halo / ROCm). Not our Spark default; useful as **bench framing competitor**, not as standing weight.
-
-Paper wake (agent inference, not GGUF fashion):
-- Agentic loads ≠ chat: long prefixes, multi-turn tool trajectories, structured roles
-- Sensitivity axes: temporal recency, modality, semantic role (user / tool / observation / reason)
-- Research push is heavily **KV-cache / multi-agent cache pool / 4-bit KV** (TriAxialKV, PolyKV, UltraQuant, PLENA) and small-edge function calling (TinyAgent) — i.e. **serve + cache + schema**, not another random Q3.
-
-**Implication for us under lock:** do **not** DIY-quant before freeze unless official fails load or loses ≥10% agent_smoke / tg. Power gain before Aug 3 is almost entirely **Layer B+C**.
-
-### Layer B — Protocol fitness (Hermes-class)
-Hermes Agent is toolset-heavy: terminal, files, web, browser, memory, cron, computer-use, delegation. It speaks OpenAI-compatible `chat.completions` + `tools` / `tool_calls`.
-
-What kills agent quality under quant/runtime:
-1. **Invalid tool-arg JSON** (our `repair_04` — server HTTP 500 parse error)
-2. **Invented tool names** outside schema
-3. **Dropped multi-tool / long-horizon** planning
-4. Chat-template / `--jinja` mismatch → prose instead of structured calls
-5. Temp/sampling too hot for tool routes
-6. Runner / response-shape fragility (our `long_06` KeyError `'tool'`)
-
-Own smoke categories already map to Hermes stress:
-- `tool_select` / `args` / `multi_tool`
-- `error_repair` (schema repair after bad args)
-- `no_invented_tools`
-- `long_horizon` multi-step
-- `refuse` safety on tools
-
-Gap vs Chadrock claim surface:
-- No public **HermesAgent-20** or broader Tool-Eval disputed set in our pack
-- No multi-turn tool→observation→tool loop scored end-to-end beyond single-response smoke
-- No published tok/s @ agent-typical ctx (8k–32k) with tools attached
-
-### Layer C — Runtime / Spark
-Documented high-leverage stack:
-- Engine: poolside `llama.cpp` branch `laguna` (+ isfinite patch as measured)
-- Flags: `--jinja`, sensible ctx (8k agent day-to-day; 256k only when needed)
-- Optional: DFlash draft GG UF if VRAM headroom
-- Sampling for tools: temp 0–0.2 agent smoke; poolside long-ctx tip temp 0.7 top_p 0.95 only when exploring 1M rope
-- Client: OpenAI base_url **must include `/v1`**, User-Agent set, strict tool schema, never execute unknown names
-
-### Layer D — Distribution (HF trending mechanics)
-Observed patterns that move Laguna quants:
-1. **Binary people can pull** (Unsloth/official GGUF) → raw DL gravity
-2. **Parent model already trending** → child repos ride search / “Tree” clicks
-3. **Specific hardware story** (Strix Halo, Spark, MLX) with **tables + sha + repro**
-4. **Named agent benches** (Tool-Eval, HermesAgent-20, Aider) with deltas
-5. **Active discussions / collections / Spaces / blog cross-posts** within 24–72h of publish
-6. Tags: `gguf`, `text-generation`, `base_model` relation, `license`, concrete hardware tags
-
-Our S5 constraint: **digests + method, no 90GB re-host.** That lowers DL ceiling. Trending must come from:
-- **Utility density of the card** (copy-paste Spark + Hermes path)
-- **Cross-link from X / Reddit / Discord / HF discussions on parent + Unsloth threads**
-- **Unique measured residue** no one else has (Spark GB10 numbers + Hermes-class smoke JSON)
-- Optional later: Space that hits our public endpoint or notebooks that pull official weight via `huggingface_hub` (no repack)
-
-## Honest competitive strategy
-
-### Win condition A (recommended) — “best Hermes-class Laguna on Spark”
-Measurable:
-- agent_smoke ≥ **39/40**, stretch **40/40**
-- New `hermes_agent_smoke` v2: 20–40 cases shaped like Hermes toolsets (terminal, read_file, web_search stubs, multi-turn repair)
-- Card #1 section: **Hermes-class local wiring** (config keys, tool loop, failure matrix)
-- Scoreboard row only from `results/*` files
-- Secondary row: unlabeled comparison notes vs Unsloth UD-Q4 *if* re-measured on same harness (optional, post-freeze if time)
-
-### Win condition B — “cluster top-10 Laguna quants by attention”
-After launch fungal growth:
-- Be in likes/DL top of *personal* third-party Laguna packs (today Chadrock 18 / CRACK 13 are the bar)
-- Collection: “Laguna local agents”
-- Reply with evidence on poolside + Unsloth Discussion spikes day-of
-
-### Win condition C — global HF trending top-10
-**Reject as primary KPI.** Requires org blast radius or weight host. Use only as aspirational side-effect if Poolside wave + a viral Spark clip align.
-
-## Quant decision tree (locked)
-
-```
-Need new GGUF?
-├─ Official Q4 loads on Spark and smoke ≥38/40 → NO. Ship digests + recipe.
-├─ Official fails load or smoke/tg ≥10% worse vs alternative you can prove → YES DIY,
-│    only with: method + imatrix commit + full sha256 + both smokes + REPRODUCE.md upgrade
-└─ “Looking unique for HF” alone → NO (violates diy_gguf lock + integrity)
-```
-
-If DIY ever unlocks (post-launch research track):
-1. Start from **official F16 or poolside imatrix**, not reverse Unsloth
-2. Protect signal path ≥ Q8_0 (match first-party philosophy)
-3. Build **agentic imatrix corpus**: tool-call transcripts, JSON schemas, multi-turn tool results, code+shell traces — not only wiki text
-4. Candidate mixtypes: Q5_K_M (quality), IQ4_XS only if Spark memory forced; never hero a boutique name without harness delta
-5. Run same agent_smoke + hermes_agent_smoke before any public claim
-
-## Hermes-specific integration map
-
-| Hermes need | Laguna pack lever |
-|-------------|-------------------|
-| OpenAI tools | llama-server `--jinja` + OpenAI `/v1` |
-| Many tools / tool search | smoke with 8–20 schemas; stress no_invented |
-| Terminal + files loop | multi-turn cases with tool role messages |
-| Reliability under repair | close `repair_04`; add 2nd-order repair cases |
-| Config discoverability | document `OPENAI_BASE_URL`, model id echo, timeouts |
-| Skills / long sessions | measure kv/cache behavior @ 8k+; document ctx defaults |
-| Honest branding | “Hermes-class / tool-agent compatible” — **not** Nous endorsement |
-
-Sample client already at `hermes/sample_client.py` + `hermes/README.md`. Expand with:
-- multi-turn repair demo
-- hermes `config.yaml` snippet pointing at Spark
-- failure→fix matrix from MEASURED.md
-
-## Trending launch playbook (Aug 2 freeze → Aug 3 12:00 WIB)
-
-### Pre-freeze (done or do)
-- [x] Public scrub · positive only · personal framing
-- [x] Diagnose + fix `repair_04` / `long_06` (harness; not weights)
-- [x] Re-run smoke → **40/40**; MEASURED + measured.json from live files only
-- [x] Add `research/` brief (this file)
-- [ ] short public “Why official Q4 for agents” section on card
-- [ ] Hermes wiring page on card (copy-paste)
-- [ ] Tags + base_model relation clean
-- [ ] SHA pins + engine sha frozen
-
-### Launch hour
-1. Meaningful card revisiontimestamp (mod time spike)
-2. Short announcement paths founder owns (not Manwë codename): personal X after rename gate lifts; HF Discussion on **parent** Laguna-S-2.1 and Unsloth GGUF pointing to Spark measurements
-3. Cross-link GitHub mirror ↔ HF
-4. One clear hero metric: **40/40 agent_smoke on DGX Spark · official Q4_K_M · Hermes-class client**
-5. Do **not** claim vs Chadrock HermesAgent-20 until we run the same harness
-
-### 48h after
-- Answer every HF comment with evidence
-- Optional: Space “agent smoke viewer” (JSON render) — low cost engagement
-- Watch Unsloth/Poolside threads for template bugs; patch docs same day
-
-## Gap closure backlog (priority)
-
-### P0 — quality residual
-1. ~~`repair_04`~~ — **closed:** case poison `{not-json` in history → server HTTP 500; Hermes client sanitizes prior tool args (keeps `_raw` for judge)
-2. ~~`long_06`~~ — **closed:** judge accepts `type: tool_call` + `any_of_tools` (no KeyError)
-3. ~~Re-smoke~~ — **40/40 · 97.25s · 2026-07-28 19:00 WIB**; official Q4 only
-
-### P1 — Hermes power
-4. `eval/hermes_agent_smoke/` 20 cases (names only inspired by tool categories; no Hermes trademark claim)
-5. Multi-turn tool observation tests
-6. Card section + config snippet for Hermes local provider
-7. Optional: side-measure Unsloth UD-Q4_K_XL on same harness (label third-party)
-
-### P2 — attention
-8. Launch-day parent discussions + collection
-9. Spark chalkboard clip / blog (personal)
-10. Space JSON explorer
-
-### P3 — post-launch research (not launch blocker)
-11. Agentic imatrix corpus design
-12. Ablate Q5_K_M / spine-protect only if P0 saturation and DIY gate opens
-13. DFlash speculative path on Spark if measured tok/s gain on agent loops
-
-## What not to do
-- Re-upload 68–96GB GGUF bare (S5 + lock)
-- Invent smoke scores or competitor beats
-- Claim Nous / Hermes endorsement
-- Chase IQ3 “flex” that tanks tool JSON
-- Spend freeze window on new quant fashion while 2 fails remain
-- Promise global top-10 trending as a committed milestone
-
-## Sources (live pulls)
-- poolside/Laguna-S-2.1 + Laguna-S-2.1-GGUF model cards (official Q4 recipe, imatrix, DFlash, jinja)
-- unsloth/Laguna-S-2.1-GGUF + Unsloth Dynamic 2.0 docs (template fixes, dynamic layer precision)
-- jcbtc Chadrock ROCmFP4 card (HermesAgent-20 / Tool-Eval narrative)
-- HF trending API + models search likes sort (2026-07-28)
-- Local pack: README, SPARK, LAUNCH_AUG3, REPRODUCE, results/MEASURED, agent_smoke.json, launch_lock.json
-- arXiv signals: TinyAgent (2409.00608), agent memory walls (2509.09505), TriAxialKV (2605.17170), PolyKV (2604.24971), UltraQuant (2606.20474)
-
-## Verdict for founder intent
-
-| Intent | Verdict |
-|--------|---------|
-| Most powerful **agentic** quantized Laguna *we* stand behind | **Official Q4_K_M + tighten agent/Hermes harness + Spark recipe** |
-| Most powerful **DIY boutique quant** | Blocked by lock; only if measured ≥10% barrier |
-| Top-10 **Laguna pack cluster** attention | Achievable with evidence card + launch distribution |
-| Top-10 **global HF trending** | Not a primary commit; side-effect only |
-
-Next single move: **P0 close the two smoke fails and wire Hermes multi-turn sample — not a new quant.**

research/device-quant-matrix-aug3.md

@@ -1,118 +0,0 @@
-# Device → quant matrix (Aug 3 accelerated track)
-
-Status: **pre-launch parallel track** (unlocked 2026-07-28).
-Personal pack only. Headline remain **Spark + official Q4_K_M**.
-
-Promo: pack docs OK now · public trending/social still **2026-08-03 12:00 WIB**.
-
-## Bottom line
-
-Full **Laguna-S-2.1** (~118B MoE) does **not** fit founder Macs.
-
-**Founder Mac lock (2026-07-28):** MacBook Pro / Mac Studio in this house are **≤32G** (live Studio probe: M2 Max · **32G**).  
-That means: **no full Laguna weights on Mac**. Mac role = **client to Spark**, not weight host.
-
-We still publish a community ladder (48–64G+ boxes other people own) as *pointers* — never as “our Mac runs it.”
-
-| Device class | RAM (usable) | Prefer quant | Source | Total ~disk | Expect | Pack claim by Aug 3 |
-|--------------|--------------|--------------|--------|-------------|--------|---------------------|
-| DGX Spark GB10 | ~121 Gi | **official Q4_K_M** | poolside | **96.0 GB** | agent bar | **Stand-behind · measured 40/40 · ~21 t/s** |
-| **Founder MacBook / Mac Studio (≤32G) · full S** | **≤32G** | **No full S weights** | — | — | OS + app already fill RAM | **Mac = OpenAI client → Spark :8000** · full-S weight non-fit |
-| **Founder Mac ≤32G · XS parallel** | **32G** (Studio live) | **XS Q4_K_M** (~18.9 Gi) | `poolside/Laguna-XS-2.1-GGUF` | **18.882 GiB** | disk candidate only | **separate model 33B-A3B** · **0 Mac smoke** · not S · see [`Laguna-XS-2.1-Mac-Agentic-fit-2026-07-28.md`](./Laguna-XS-2.1-Mac-Agentic-fit-2026-07-28.md) |
-| Community fat Mac / mini (64–128G) *others* | 64–128G | Unsloth **UD-IQ4_XS** or Bartowski **IQ4_XS** | unsloth / bartowski | **~58 / ~63 GB** | near-Q4 if engine OK | community pointer only · **not founder-measured** |
-| Community Mac/PC 48–64G *others* | ~48–64G | Unsloth **UD-IQ3_S** (first) | unsloth | **~48 GB** | quality dip | Spark same-harness IQ3_S **38/40** pointer · not Mac-local |
-| PC workstation 64–96G CUDA | 64–96G | Bartowski **Q4_K_S** / **IQ4_XS** / Unsloth **UD-Q4_K_S** | bartowski / unsloth | **~69 / ~63 / ~69 GB** | if VRAM+RAM enough | pointer + CUDA note |
-| Laptop 16–32G (any brand) full S | 16–32G | **No full S SKU** | — | — | distill / XS parallel | **full S non-fit** · XS is different product |
-| iPhone / iPad / Android phone·tablet | ~4–12G class | **No full Laguna-S SKU** | — | — | MoE weights alone ~40GB+ | **explicit non-fit** — bonsai/SLM only; see [`phone-bonsai-ios-android-2026-07-28.md`](./phone-bonsai-ios-android-2026-07-28.md) |
-
-Sizes are **sum of GGUF shards** (HF live probe 2026-07-28). Runtime RAM ≫ disk; leave OS + KV headroom.
-
-### Founder “Laguna Mac” definition
-
-| Phrase | Means | Does not mean |
-|--------|-------|---------------|
-| Laguna Mac (default) | Mac as **Hermes / OpenAI client** talking to Spark **S** | Full **S** weights in Mac unified memory |
-| Laguna Spark | Weight host + measured bar for **S** (Q4 40/40) | — |
-| Laguna XS on Mac | Separate **33B-A3B** track · Q4 ~19G candidate | A quant of S · measured ship · “S on Mac” |
-| Community 64G ladder | Docs for strangers with fat boxes running **S** quants | Founder hardware claim |
-
-## Candidate IDs (pull scripts use these)
-
-| sku_id | repo | path pattern | ~GB | target devices |
-|--------|------|--------------|-----|----------------|
-| `official-q4km` | `poolside/Laguna-S-2.1-GGUF` | `laguna-s-2.1-Q4_K_M.gguf` | 96.0 | Spark stand-behind |
-| `unsloth-ud-q4k-xl` | `unsloth/Laguna-S-2.1-GGUF` | `UD-Q4_K_XL/*` (3 shards) | 73.4 | 96–128G workstation / fat Mac |
-| `unsloth-ud-iq4-xs` | `unsloth/Laguna-S-2.1-GGUF` | `UD-IQ4_XS/*` (3 shards) | 57.6 | 64–96G Mac / PC |
-| `unsloth-ud-iq3-s` | `unsloth/Laguna-S-2.1-GGUF` | `Laguna-S-2.1-UD-IQ3_S.gguf` | 48.4 | 64G Mac tight / 48–64G |
-| `unsloth-ud-q2k-xl` | `unsloth/Laguna-S-2.1-GGUF` | `Laguna-S-2.1-UD-Q2_K_XL.gguf` | 39.7 | 48G class experiment |
-| `bartowski-iq4-xs` | `bartowski/Laguna-S-2.1-GGUF` | `Laguna-S-2.1-IQ4_XS/*` | 63.3 | alt IQ4 for Radio / compare |
-
-## Promotion gate (unchanged quality bar)
-
-Promote a smaller SKU onto the **card scoreboard** only if:
-
-1. Same `eval/agent_smoke` 40 cases (or named subset with disclaimer)
-2. Same method (temp, max tokens, tools schema)
-3. Host labeled honestly (never print Mac/PC numbers as "Spark")
-4. sha256 + exact HF path
-5. Delta vs Spark official Q4_K_M published
-6. Fail cases listed
-7. agent_smoke ≥ official − 2 (**≥38/40** if official is 40) **or** research footnote only
-8. Founder go for any public claim change after freeze
-
-Else: keep under `research/` + optional card "community ladder" table with **unmeasured** tag.
-
-## Serve notes by OS
-
-| Host | Engine | Notes |
-|------|--------|-------|
-| Spark | poolside `llama.cpp` `laguna` + CUDA | default pack path · **only weight host we stand behind** |
-| Founder Mac ≤32G | **no local llama-server for full Laguna** | client: `OPENAI_BASE_URL=http://<spark>:8000/v1` · Tailscale OK |
-| Community fat Mac Metal (64G+) | llama.cpp Metal **or** MLX if available | stranger path only; may need laguna fork — test load first |
-| PC CUDA | laguna fork or upstream ≥ Laguna support | ngl defendable on 24GB+ VRAM only for *active* path; MoE still RAM-heavy |
-| PC ROCm / laptop iGPU | experimental | do not claim ship |
-
-## Phone / mobile (locked non-fit)
-
-**Quant success on Mac mini / MacBook / PC does not imply iPhone or Android.**
-
-- Full Laguna-S-2.1 is ~118B MoE. Smallest honest GGUF class we measured still **~45 Gi on disk** (UD-IQ3_S).
-- Mobile NPU/RAM envelopes are roughly **4–12 GB**, with OS + UI overhead.
-- No App Store / Play path claims for full weights.
-- Path if founder wants mobile later: **distill / SLM / on-device small model** — separate product, not this pack's quant ladder.
-- Do not market “quantized to phone” from desktop IQ3/IQ4 wins.
-
-## What we will not say by Aug 3
-
-- "Runs on MacBook Pro / Mac Studio" without saying **≤32G = client only**
-- "Our Mac runs full Laguna" / any founder-Mac weight claim
-- "Runs on any MacBook" / 16–32G laptop full Laguna weights
-- "Runs on iPhone / Android" for full Laguna (even quantized)
-- Replacing official Q4 as default without measure win
-- Bulk multi-quant LFS rehost of Unsloth tree into our HF model
-- Global HF top-10 from device count alone
-- Mac/PC numbers labeled as Spark
-- IQ3_S / IQ4_XS as "Laguna Mac" headline (those are **community ≥48G**, not founder Mac)
-
-## Aug 3 pack shape
-
-1. **Headline**: Spark official Q4_K_M measured row  
-2. **Laguna Mac story**: Hermes/OpenAI **client → Spark** (copy-paste) · weight non-fit ≤32G in one line  
-3. **Community ladder**: 48–64G+ pointers only · no founder-Mac measure row  
-4. **Scripts**: `scripts/pull_sku.sh <sku_id>` (Spark / fat-box strangers)  
-5. **Measured smaller row(s)**: only if same-harness JSON under `results/sku_<id>/` (Spark today)  
-6. **Hermes-class suite**: live **27/27** on Q4 · `results/hermes_agent_smoke.json`  
-
-## Schedule (accelerated)
-
-| When | Work |
-|------|------|
-| Jul 28–29 | Matrix + pull scripts + docs; Hermes v2 live on Q4; IQ3_S pointer done |
-| Jul 30 | Card: Mac ≤32G client path + non-fit; stranger REPRODUCE |
-| Jul 31–Aug 1 | Buffer · no Mac weight chase |
-| Aug 2 18:00 | Freeze claims |
-| Aug 3 12:00 | Promo go-live |
-
-## Authority
-
-Personal founder pack. Third-party quants remain Unsloth/Bartowski IP — we **point + measure**, default **do not rehost**.

research/laguna-xs-2.1-mac-fit-2026-07-28.md

@@ -1,110 +0,0 @@
-# Laguna-XS-2.1 · founder Mac ≤32G fit track
-
-Status: **prepared research** (2026-07-28).  
-Awaiting three-jury on prepared pack, then founder gates.  
-**Not** the Aug 3 S pack headline. **Not** a measured ship claim.
-
-Personal founder track only. Dual roadmap → [`roadmap-dual-s-and-xs-2026-07-28.md`](./roadmap-dual-s-and-xs-2026-07-28.md)
-
-## Bottom line
-
-| Track | Model | Host | Status |
-|-------|--------|------|--------|
-| **A · stand-behind** | Laguna-**S**-2.1 ~118B MoE Q4_K_M | **DGX Spark only** | measured 40/40 + hermes 27/27 |
-| **B · Mac local candidate** | Laguna-**XS**-2.1 33B-A3B Q4_K_M | founder Mac ≤32G | **disk candidate** · **0 local smoke** |
-
-Full **S** stays Spark-only. Mac is **not** “S via XS”.
-
-## Live HF probe (2026-07-28)
-
-### Base
-- `poolside/Laguna-XS-2.1`
-- sha: `205dc65dd4bda946c50da6b7522b215734fa107b`
-- Shape (card): **33B total · 3B active/token · MoE**
-- Poolside claim: “compact enough to run on a Mac with **36 GB** of RAM”
-- Founder Studio coin: **32G** exactly (`hw.memsize` 34359738368) → **tighter than vendor 36G**
-
-### Official GGUF
-- `poolside/Laguna-XS-2.1-GGUF` sha `1a37c0a5fb8c7a18e6106decb6be6327d1b63fa6`
-
-| File | Bytes | GiB | sha256 | Founder Mac ≤32G |
-|------|------:|----:|--------|------------------|
-| `Laguna-XS-2.1-Q4_K_M.gguf` | 20274300032 | **18.882** | `1ac7079101fca5a6df8c5a7523a3c30ea7d1c0e4b1258090e7d6d4039287f6cb` | **disk candidate** (runtime unproven) |
-| `Laguna-XS-2.1-BF16.gguf` | 66930226304 | **62.334** | `d1d99abe30c37749ec1b1ae2681cec74caa550ac6243d2ea70b61b1ff9d187ca` | **non-fit** |
-
-Upstream recommended default: **Q4_K_M**.
-
-Safetensors base ~14 shards ≈ **~67G** class — non-fit on ≤32G.
-
-### Engine dependency
-- llama.cpp Laguna support: PR `ggml-org/llama.cpp#25165`
-- Prefer `poolsideai/llama.cpp` branch `laguna` on Mac Metal — **do not assume stock brew llama runs XS**
-- Ollama (`ollama pull laguna-xs-2.1`) = alt route until smoke
-
-## Founder Mac envelope (Studio live)
-
-| Fact | Value |
-|------|-------|
-| Model | Mac14,13 · M2 Max |
-| RAM | **32G** exactly |
-| System Data free (prep) | ~40 Gi · 91% used — **prefer WD_BLACK for pull** |
-| WD_BLACK free | ~1.7 Ti |
-| Full S Q4 | ~96G · **non-fit** |
-| Full S IQ3_S pointer | ~48G · **non-fit** as local weight |
-| XS Q4_K_M | ~18.9G disk · **candidate** |
-
-Runtime RAM ≫ disk. Leave OS + browser + Hermes + KV headroom.  
-**Disk-fit ≠ load-fit ≠ agent-smoke-fit.**
-
-## Prep lock
-
-`results/xs_mac_track_lock.json` — status `prepared_research`
-
-## Honesty rules
-
-**Say**
-- “XS Q4_K_M is the official ~19G Mac-class GGUF for Laguna-**XS**-2.1”
-- “Founder Mac client still talks to Spark for full **S**”
-- “XS is a **separate** 33B-A3B model, not a quant of S”
-
-**Do not say**
-- “Laguna runs on Mac” without naming **XS**
-- “S on Mac via XS”, “S-lite”, “S-Mac”
-- Any t/s, agent_smoke, or Hermes pass for XS until same-harness JSON
-- Poolside SWE-bench / Terminal-Bench as founder measured bar
-- BF16 / safetensors base as Mac-local
-
-## Measurement gate (before any claim)
-
-1. Working engine on founder Mac · record commit
-2. `sha256` matches lock after download
-3. Load proof: RSS + free RAM after load · ctx used
-4. Same `eval/agent_smoke` 40 (or named subset + disclaimer)
-5. Optional hermes-class 27
-6. Host labeled **Mac Studio 32G** (never as Spark)
-7. Fail list + separate delta table (not diluted into S headline)
-8. Founder go before any public/pack claim move
-
-Results dir when measured: `results/sku_xs-official-q4km/`
-
-## Suggested first pull (NOT run — needs founder go)
-
-```bash
-# ~19G → external SSD preferred
-huggingface-cli download poolside/Laguna-XS-2.1-GGUF \
-  Laguna-XS-2.1-Q4_K_M.gguf \
-  --local-dir "/Volumes/HFR WD_BLACK SN850X/models/laguna-xs-2.1"
-shasum -a 256 "/Volumes/HFR WD_BLACK SN850X/models/laguna-xs-2.1/Laguna-XS-2.1-Q4_K_M.gguf"
-# expect 1ac7079101fca5a6df8c5a7523a3c30ea7d1c0e4b1258090e7d6d4039287f6cb
-```
-
-Do **not** pull BF16 on this Mac.
-
-## Relation to Aug 3 S pack
-
-| Item | Owner |
-|------|--------|
-| `hizrianraz/laguna-s-2.1-spark` freeze/launch | **Roadmap A only** |
-| XS Mac notes | this file + `xs_mac_track_lock.json` |
-| multi_device_track | S community pointers; XS is **extra parallel** |
-| XS may delay S freeze | **no** |

research/phone-bonsai-ios-android-2026-07-28.md

@@ -1,284 +0,0 @@
-# Phone bonsai — fit Laguna agent identity to iOS / Android
-
-Status: research (2026-07-28). Personal. Not a scoreboard claim.
-
-Premise lock: **full Laguna-S-2.1 (even IQ3_S ~ half-size of Q4) does not fit phone RAM**.
-Mac mini / MacBook / PC quant success does **not** carry to iPhone or Android.
-Phone path = **bonsai** (distill / SLM / stack split), not another GGUF of the same MoE.
-
-Stand-behind headline unchanged: Spark official Q4_K_M 40/40.
-
----
-
-## Bottom line
-
-| Claim | Verdict |
-|-------|---------|
-| Quant full Laguna onto iPhone/Android | **No** |
-| Same 40/40 agent_smoke bar on-device | **No** (bar too big; model too big) |
-| Ship a **Laguna-shaped small agent** on phone | **Yes**, separate SKU family |
-| Best shape | **Bonsai student 0.5B–2B** + tool router + optional cloud teacher |
-| Best runtimes | iOS: MLC-LLM / llama.cpp Metal · Android: MLC / llama.cpp + ExecuTorch where PyTorch path |
-| Identity | New card prefix e.g. `laguna-bonsai-*` — never rebrand full Laguna |
-
----
-
-## Why full Laguna cannot phone-fit
-
-### RAM envelope (order of magnitude)
-
-Weights only (no KV), approximate GB:
-
-| params | Q4_K_M | Q3 / IQ3 | BitNet~1.58 |
-|--------|--------|----------|-------------|
-| 0.5B | 0.28 | 0.20–0.23 | ~0.12 |
-| 1B | 0.57 | 0.40–0.45 | ~0.25 |
-| 2B | 1.1 | 0.8–0.9 | ~0.50 |
-| 3B | 1.7 | 1.2–1.4 | ~0.75 |
-| 7–8B dense | 4–5 | 3–3.6 | ~1.8–2.0 |
-| **Laguna-S-2.1 MoE Q4** | **~90–96G** | IQ3 still tens of GB | n/a |
-
-Practical app budget after OS (soft):
-
-| device class | device RAM | usable app ~ | safe model+KV |
-|--------------|------------|--------------|---------------|
-| iPhone 15/16 class | 8G | ~3–3.5G | **≤ ~2.5–2.7G** |
-| iPhone Pro class | 12G | ~5–5.5G | **≤ ~4.5G** |
-| Android 8G | 8G | ~3–3.5G | **≤ ~2.5G** |
-| Android flagship 12G | 12G | ~5–5.5G | **≤ ~4.5G** |
-
-Add KV @ 4k–8k, tokenizer, UI, OS jetsam headroom → **phone hard ceiling is ~1–3B @ 4-bit**, not 100B-class sparse MoE on disk.
-
-Sparse MoE still **stores all experts**. Activation sparsity ≠ download/RAM sparsity unless you **prune experts / nest** (FlexMoE-class) and re-export a tiny subnetwork — still a new model, not “quant the GGUF harder.”
-
-### iOS-specific
-
-- Jetsam kills apps that grow past device class budget; no swap like desktop.
-- Background freeze; long agent loops are foreground or push-driven.
-- Store path: ANE / Metal preferred for App Store polish; C++ via MLC or llama.cpp is proven for indie / TestFlight.
-- Core ML conversion helps ANE for supported ops; exotic MoE kernels often **fail conversion** → keep phone student dense/simple.
-
-### Android-specific
-
-- Huge RAM skew (6G–16G+). Must ship **tiered assets** or download-on-first-run by RAM class.
-- NNAPI / Vulkan / OpenCL backends vary by OEM; MLC + llama.cpp are the portable bets.
-- ExecuTorch (Meta) is the PyTorch-native production path (Meta on-device apps); great if student stays in ET-known archs.
-- Thermal + big.LITTLE: sustained tokens crack after tens of seconds; design short tool turns, not 2k-token monologues.
-
----
-
-## What “bonsai” means here
-
-Bonsai ≠ max Q-drop on the same tree.
-
-Bonsai = **genetic reduction + shape prune + water small pot**:
-
-1. **Smaller genome** — student 0.5B / 1B / 2B dense (or BitNet 2B).
-2. **Shape for the pot** — deep-thin MobileLLM-style, GQA, shared embed, short default context (2k–4k on device).
-3. **Teach from Laguna** — distillation / imitation on **agent traces**, not only chat Wikipedia-MJ.
-4. **Split the organism** — on-device planner / tool-caller; heavy reasoning optional cloud teacher (Laguna Spark) with privacy gate.
-5. **Never claim the tree is the forest** — separate HF / scoreboard row; humble discipline on agent bar.
-
----
-
-## Research stack that matters (2024–2026)
-
-### Architecture for sub-billion / ~1B on-device
-
-| work | id | takeaway for us |
-|------|----|-----------------|
-| MobileLLM | 2402.14905 | Deep-thin + embed share + GQA beats naive wide-small. Architecture > just “more tokens” under 1B. |
-| MobileLLM-Pro ~1B | 2511.06719 | SOTA-class ~1B, 128k via positional distillation, merges, minor 4-bit hit. Template for phone **general** LM. |
-| MobileLLM-R1 | 2509.24945 | Sub-B **reasoning** with open recipes; CoT can emerge well below “big model only.” |
-| MobileLLM-Flash | 2603.15954 | **Latency-guided** NAS / prune from backbone; ExecuTorch-friendly; attention skipping for long ctx. Industry deploy shape. |
-| MiniCPM 1.2B/2.4B | 2404.06395 | SLM can match older 7–13B class on many tasks with good recipe. |
-| MiniCPM-V | 2408.01800 | Multimodal **on phone** exists — but dual-modality budget fights pure agent tools; later track. |
-| TinyLLM | 2412.15304 | 30–120M **task-curated** can beat larger generalists on narrow tasks — backs ultra-narrow phone specialists. |
-
-### Extreme compression
-
-| work | id | takeaway |
-|------|----|----------|
-| BitNet b1.58 2B4T | 2504.12285 | Native ~1.58-bit 2B open weights; real memory/energy wins if runtime is bitnet.cpp-class. |
-| bitnet.cpp | 2410.16144 | Fast lossless ternary on CPU/ARM — relevant Android / iOS CPU fallback. |
-| FBI-LLM | 2407.07093 | Fully binary via autoregressive distillation — research edge, hardware story still maturing. |
-| EdgeRazor | 2605.04062 | Mixed-precision QAD; sub-2-bit MobileLLM/Qwen6e-1 students — if we push below 4-bit quality. |
-
-### MoE → smaller (only if we ever bonsai *from* Laguna MoE body)
-
-| work | id | takeaway |
-|------|----|----------|
-| FlexMoE | 2606.27866 | One train → **nested** deployable subnets across budgets. Closest “prune Laguna into device ladder” research. |
-| MoE++ | 2410.07348 | Zero-compute experts — arch idea, not a free phone port of poolside MoE. |
-
-Do **not** expect expert-pruning alone to bring Laguna-S-2.1 to 2G without catastrophic agent regression. Treat MoE prune as research; default phone path = **dense student**.
-
-### Tools / agents with small models
-
-| work | id | takeaway |
-|------|----|----------|
-| Small LLMs Are Weak Tool Learners | 2401.07324 | Single small LM fails full agent stack; **split** planner / caller / summarizer across small specialists. **Core bonsai pattern.** |
-
-So for phone agent:
-
-- Student-A: intent + plan (short JSON)
-- Student-B: tool call format only
-- Student-C: summarize / speak to user
-- Optional: cloud Laguna for hard multi-step
-
-This is how you keep Hermes-*shaped* behavior without 90G on device.
-
-### Speculative / test-time
-
-- Speculative decoding with **tiny draft on-device** verifying against mid student helps tok/s; draft can be 100–300M.
-- Latent recurrent depth (2502.05171) = spend compute not tokens — interesting if thermal allows loop depth, research-only for Aug track.
-
----
-
-## Runtime map (ship-relevant)
-
-| stack | iOS | Android | notes |
-|-------|-----|---------|-------|
-| **MLC-LLM** | Metal A-series | OpenCL Adreno/Mali | Universal compile; OpenAI-compat engine; indie-friendly. |
-| **llama.cpp** | Metal | Android NDK / Vulkan | GGUF path; matches our desktop harvest methodology. |
-| **ExecuTorch** | yes | yes | PyTorch-native; Meta prod; LLM examples for Llama-class. Prefer if we train/export in PT. |
-| **MediaPipe Tasks** | yes | yes | Strong for vision/audio tasks; LLM story secondary to MLC/ET. |
-| **Core ML** | ANE | n/a | Convert when ops allow; fail-closed for exotic MoE. |
-| **bitnet.cpp** | CPU | CPU/ARM | Only for BitNet students. |
-| **MLX** | — | — | Mac; not phone. |
-
-Recommendation:
-
-- **Prototype path:** Q4/Q5 GGUF of dense student → llama.cpp iOS/Android samples (same smoke harness flavor, reduced bar).
-- **Product path:** MLC or ExecuTorch once architecture freezes.
-- **Do not** start phone path by converting Laguna MoE GGUF.
-
----
-
-## Honest capability ladder (agent)
-
-Desktop Laguna Spark: **40/40** agent_smoke + 27/27 Hermes-class (measured).
-
-Phone expectations if we bonsai correctly:
-
-| SKU class | size | likely role | ~/40-style agent bar |
-|-----------|------|-------------|----------------------|
-| bonsai-nano | 0.5B Q4 | classify, rewrite, single tool | **ship_min phone bar** (new suite ~12–20 cases), not 40/40 |
-| bonsai-mini | 1B Q4 | plan + 1–2 tools | partial multi-step |
-| bonsai-edge | 2B Q4 / BitNet2B | on-device default agent | better multi-step; still << Laguna |
-| hybrid | student + Laguna cloud | full agent when online | cloud holds 40/40; device holds offline floor |
-
-Define a **phone_agent_smoke** (separate harness): shorter tools schema, stricter latency, thermal budget, offline/online modes. Never scoreboard-merge with Spark 40/40.
-
----
-
-## Distillation recipe (actionable)
-
-### Data (the real moat)
-
-Mine from Laguna teacher (Spark):
-
-1. Agent trajectories that **passed** harness (tool JSON, multi-step, error repair).
-2. Hard negatives: invented tools, broken JSON, overlong.
-3. Synthetic repairs: corrupt → Laguna fix pairs.
-4. Privacy scrub before any leave-box export.
-
-Avoid: generic UltraChat-only SFT — tools won’t stick (see 2401.07324).
-
-### Loss stack
-
-1. **SFT** on clean tool traces (strict schema).
-2. **On-policy distill** where student acts, teacher grades / relabels (best agent transfer).
-3. Optional **token KD** (KL to teacher) on non-tool chat for fluency.
-4. Optional preference (DPO/odds-ratio) on “good tool vs bad tool” pairs.
-
-### Train budget (order)
-
-- Student init: Qwen2.5-0.5B/1.5B-Instruct or MobileLLM-Pro / Gemma-3-1B class — pick one ecosystem and stay.
-- LoRA first (days), full FT if LoRA plateaus.
-- Quant-aware or post-train Q4_K_M/IQ4_XS export via llama.cpp / MLC convert.
-- Always restore desktop teacher for eval of **teacher** quality; student has own card.
-
-Gates (align Scientium-style joy when personal track hardens):
-
-- Gate P0: student builds + loads iOS sim **or** Android emulator with < budget RAM.
-- Gate P1: phone_agent_smoke ship_min.
-- Gate P2: human 20-turn soak + thermal (surface temps / throttle).
-- No “Laguna on iPhone” public line. Ever. Name it bonsai.
-
----
-
-## Hybrid privacy pattern (recommended product)
-
-```
-[User on phone]
-    → local bonsai (default)
-         → simple tools (calendar, notes, local files perimeter)
-    → hard / long / coding / multi-tool
-         → optional escalate to Spark Laguna (TLS, founder network only)
-         → return summary only to device
-```
-
-Privacy principle lock: sensitive founder inference stays on-box / sovereign; phone bonsai is personal endpoint class, not fleet A2A dump.
-
----
-
-## Non-goals (phone)
-
-- Scoreboard promote any phone number as Laguna-S-2.1
-- Claim IQ3 Mac success ⇒ iPhone
-- App Store “100B on phone” vibes
-- Full 8k–128k context default on 8G phones
-- Unsloth multi-quant rehost of Laguna for mobile
-
----
-
-## Aug 3 vs later
-
-### By 2026-08-03 12:00 WIB (docs + honesty only)
-
-- Device matrix already: phone = **non-fit full model**
-- This note linked from post-freeze path + matrix
-- Optional: one **bare** public line on HF card: “Mobile: distill/SLM track only; full weights not phone-fit.”
-- No phone weight publish required for Aug 3 clock
-
-### Post-freeze build order (if founder greenlights compute)
-
-1. Freeze phone_agent_smoke v0 (12–20 cases) + ship_min.
-2. Choose student base (default lean: **Qwen2.5-1.5B-Instruct** or **MobileLLM-Pro-class 1B**).
-3. Harvest Laguna tool traces on Spark (offline JSONL).
-4. SFT + schema constrain → GGUF Q4 → MacBook smoke first (faster loop).
-5. MLC or llama.cpp iOS sim + Android emulator RAM probe.
-6. Only then TestFlight / internal APK.
-7. Hybrid escalate toggle last.
-
----
-
-## Decision packet (one decision)
-
-**D1 — Phone track posture after Aug 3**
-
-- A) **Docs-only** through promo clock (recommended default): non-fit declared; bonsai plan parked.
-- B) **Open bonsai spike** post-freeze: trace harvest + 0.5–1.5B SFT week, no public claim.
-- C) **Hybrid product spike**: bonsai offline floor + Laguna cloud escalate design (needs privacy review).
-
-Default if silence: A until founder opens B/C.
-
----
-
-## Sources (anchors)
-
-- MobileLLM 2402.14905 · MobileLLM-Pro 2511.06719 · MobileLLM-R1 2509.24945 · MobileLLM-Flash 2603.15954  
-- MiniCPM 2404.06395 · MiniCPM-V 2408.01800  
-- BitNet b1.58 2B4T 2504.12285 · bitnet.cpp 2410.16144 · FBI-LLM 2407.07093 · EdgeRazor 2605.04062  
-- FlexMoE 2606.27866 · TinyLLM 2412.15304  
-- Small LLMs weak tool learners 2401.07324  
-- ExecuTorch 2605.08195 · MLC-LLM (llm.mlc.ai)  
-- Pack priors: `device-quant-matrix-aug3.md`, `post-freeze-smaller-device-path.md`  
-- Measured IQ3_S agent_smoke 38/40 on Spark does **not** change phone non-fit
-
----
-
-## One-line canon
-
-**Phone gets a bonsai student taught by Laguna — never Laguna shoehorned into a phone.**

research/post-freeze-smaller-device-path.md

@@ -1,74 +0,0 @@
-# Smaller-device path — Mac mini / MacBook / PC / laptop
-
-Status: **UNLOCKED for Aug 3 parallel track** (2026-07-28).
-Still personal surface. Public promo clock unchanged: **2026-08-03 12:00 WIB**.
-
-Canonical matrix → [`device-quant-matrix-aug3.md`](./device-quant-matrix-aug3.md)  
-Pull helper → `scripts/pull_sku.sh`
-
-## Bottom line
-
-**Two roadmaps:** [`roadmap-dual-s-and-xs-2026-07-28.md`](./roadmap-dual-s-and-xs-2026-07-28.md)
-
-**Founder Macs = MacBook Pro / Mac Studio ≤32G** (live Studio: M2 Max 32G).  
-Full **S** weights **do not load** there. Default “Laguna Mac” = **client to Spark S**.  
-**XS** is a **separate** 33B-A3B track for possible local Mac weights (Q4 ~19G) — prep only, unmeasured.
-
-By Aug 3 (Roadmap A): Spark measured S bar + honest Mac client path + community fat-box S pointers.  
-Roadmap B (XS) must **not** dilute or delay that freeze.
-
-Do **not** swap the Spark official S Q4_K_M headline.  
-Do **not** chase 64G Mac hardware we do not own.  
-Do **not** call XS “S on Mac”.
-
-## Order of work (accelerated)
-
-1. Hold Spark official **S** Q4_K_M (40/40 + ~21 tok/s + digests).
-2. Pack stranger reproduce path (Spark-weighted S).
-3. Document **founder Mac ≤32G non-fit for full S weights** + OpenAI client → Spark.
-4. Keep IQ3_S / IQ4 as **community S** pointers (Spark-measured SKU only).
-5. Promote a **S** SKU to the scoreboard only if ≤ −2/40 vs official **and** host labeled.
-6. **XS prep** (parallel): lock + matrix + dual roadmap — pull/smoke only after founder go; prefer WD_BLACK.
-
-## Device targets
-
-| Device class | RAM | Candidate | Engine | Aug 3 intent |
-|--------------|-----|-----------|--------|--------------|
-| DGX Spark | ~121G | Official **S** Q4_K_M (~96G) | poolside laguna CUDA | **stand-behind measured** |
-| **Founder Mac · full S** | **≤32G** | **no full S weights** | Hermes / curl → Spark | **Laguna Mac = client** |
-| **Founder Mac · XS parallel** | **32G** | official **XS** Q4_K_M (~18.9G) | laguna Metal / Ollama? | **research** · 0 smoke · see [`Laguna-XS-2.1-Mac-Agentic-fit-2026-07-28.md`](./Laguna-XS-2.1-Mac-Agentic-fit-2026-07-28.md) |
-| Community workstation 64–128G | 64–128G | **S** UD-Q4_K_XL / IQ4_XS | CUDA / Metal | stranger pointer only |
-| Community Mac/PC 48–64G | 48–64G | **S** UD-IQ3_S (Spark 38/40) | Metal/CUDA | stranger pointer; not founder Mac |
-| Laptop 16–24G full S | 16–24G | no full S | — | distill / bonsai / or XS separate |
-
-## Gate checklist (any scoreboard SKU)
-
-- [ ] Same `eval/agent_smoke` 40 cases
-- [ ] Same method (temp, max tokens, tools schema)
-- [ ] Host labeled honestly (never Mac numbers as "Spark")
-- [ ] sha256 + exact source path
-- [ ] Delta vs Spark official Q4_K_M
-- [ ] Fail cases listed
-- [ ] ≥ official − 2 **or** research footnote only
-- [ ] Founder go for post-freeze claim edits
-
-## Non-goals
-
-- Bulk multi-quant LFS rehost of Unsloth/Bartowski trees
-- "Runs on MacBook / Studio" weight claims for ≤32G founder fleet
-- Waiting on / inventing a 64G Mac row for launch
-- "Runs on any MacBook" claims
-- Nemotron swap as Laguna headline
-- Global HF top-10 from device count alone
-- Public X/trending before Aug 3 12:00 WIB
-
-## Authority
-
-Personal founder pack. Third-party quants stay at source; we point + measure.
-
-## Phone / tablet
-
-Full Laguna (even IQ3) is **non-fit** for iPhone and Android. Mac/PC quant success does not carry to mobile. Distill/SLM only.
-
-Deep research (bonsai path): [`phone-bonsai-ios-android-2026-07-28.md`](./phone-bonsai-ios-android-2026-07-28.md)
-

research/quant-comparison-scoreboard-2026-07-28.md

@@ -1,112 +0,0 @@
-# Quant comparison scoreboard — Laguna-S-2.1 (2026-07-28)
-
-Purpose: same-family quant map + **measured Spark evidence** for our stand-behind choice.
-Not a global HF top-10 claimsheet. Global trending needs downloads/likes/recency — our wedge is **measured agent runtime on Spark + official digest binding**.
-
-## Our stand-behind (measured live)
-
-| Field | Value |
-|-------|-------|
-| Pack | `hizrianraz/Laguna-S-2.1-Spark-Agentic` |
-| Weights | official `poolside/Laguna-S-2.1-GGUF` · `laguna-s-2.1-Q4_K_M.gguf` |
-| sha256 | `a8b55c75714ea73fd90ec85de5defdc0b8d88ca0ad2108343cdd8fc22f7583e4` |
-| Host | DGX Spark GB10 · 121 Gi |
-| Engine | poolside llama.cpp `04b2b72` laguna + isfinite patch |
-| Serve | `:8000` · ctx 8192 · ngl -1 · jinja · fa on · alias `local-laguna` |
-| Gen128 (authoritative) | **~21.47 tok/s** · multi suite · `results/MEASURED.md` · pack tip `bf82eab` |
-| Gen128 (earlier same-day stamp) | 21.016 tok/s · 128 tok · 6.091s — band sample only, **not** a second claim |
-| Prefill marks (measured multi) | 2k ~1.6s · 8k ~4.8s (server_bench_live) |
-| agent_smoke | **40/40 (100%)** · 84.86 s · temp 0.0 |
-| hermes_agent_smoke v2 | **27/27 (100%)** · 100.1 s · `results/hermes_agent_smoke.json` |
-
-## Same-family GGUF landscape (HF metadata)
-
-| Source | Likes | Downloads | Role vs us |
-|--------|------:|----------:|------------|
-| `poolside/Laguna-S-2.1-GGUF` | 157 | 90106 | official |
-| `unsloth/Laguna-S-2.1-GGUF` | 222 | 129601 | unsloth |
-| `bartowski/Laguna-S-2.1-GGUF` | 7 | 5235 | bartowski |
-| `lmstudio-community/Laguna-S-2.1-GGUF` | 1 | 2991 | lmstudio |
-
-### Size-class picks (single file or last shard size shown for multi-part — **sum shards for full disk**)
-
-| Class | Official Poolside | Unsloth (example) | Bartowski (example) | LM Studio | Fit note |
-|-------|-------------------|-------------------|---------------------|-----------|----------|
-| Agent default Q4 | **`Q4_K_M` 96.03 GB** (our measure) | `Laguna-S-2.1-UD-Q4_K_XL-00003-of-00003.gguf` ~23.42 GB shard/file UD-Q4_K_XL | `Laguna-S-2.1-Q4_K_M-00002-of-00002.gguf` ~31.94 GB shard/file | `Laguna-S-2.1-Q4_K_M-00002-of-00002.gguf` ~31.22 GB shard/file | Spark 128G OK for official Q4_K_M |
-| Dense Q8 | `Q8_0` 128.75 GB | multi-part Q8_0 | multi-part Q8_0 | multi-part Q8_0 | Heavy on 128G; optional |
-| Smaller third-party | — | `Laguna-S-2.1-UD-IQ4_XS-00003-of-00003.gguf` ~7.74 GB shard/file / `Laguna-S-2.1-UD-Q3_K_M-00003-of-00003.gguf` ~4.04 GB shard/file | IQ/Q3 family present | — | Laptop/Mac candidates **after** freeze |
-
-## Fair comparison method (locked)
-
-1. **Same harness**: `eval/agent_smoke` (40 cases) + `scripts/bench_server.py` + gen128 chat.
-2. **Same host**: DGX Spark only for headline numbers.
-3. **Same engine**: poolside `laguna` fork at pinned SHA when model needs Laguna ops; stock llama.cpp only if it loads cleanly — note fork delta.
-4. **Digest-bind**: publish sha256; no silent re-quant.
-5. **No server swap during launch lock** without explicit founder go: live Laguna Q4 stays up.
-
-## What we will / will not claim
-
-| Claim | Status |
-|-------|--------|
-| First/only Laguna quant | **No** — Poolside/Unsloth/Bartowski exist |
-| Best universal tok/s | **No** without same-harness wins |
-| Spark-measured agent smoke 40/40 on official Q4_K_M | **Yes** (evidence in `results/`) |
-| Spark-measured Unsloth UD-IQ3_S agent smoke ≥38/40 (not headline) | **Yes** · 38/40 · `results/sku_unsloth-ud-iq3-s/` |
-| Full Laguna on iPhone / Android | **No** — explicit non-fit |
-| Global HF top-10 overall | **Not a commit** — optimize pack cluster + agent evidence |
-| Top tier *among Laguna Spark packs / measured agent runtimes* | **Target** |
-
-## Cross-family (Nemotron) note
-
-- On disk Spark: `Nemotron-3-Super-120B-Q4_K.gguf` ~66G.
-- **Not swapped in** this window (would drop live Laguna serve).
-- Post-freeze same-harness pulse is the fair path if founder wants scientific cross-base compare.
-
-## Laptop / Mac mini / PC path (Aug 3 accelerated — parallel track)
-
-**Unlocked 2026-07-28** for pack docs + pointer wiring before promo clock.
-Does **not** move public promo. Live Q4 restored on `:8000` after one brief IQ3 window.
-
-### IQ3_S same-harness (done 2026-07-28)
-
-| Field | Value |
-|-------|-------|
-| SKU | Unsloth `Laguna-S-2.1-UD-IQ3_S.gguf` |
-| sha256 | `8a9ab3f8b3ff1723441cd251e873b295a7ef086d78dbae7515e5e27c8382b002` |
-| Host | DGX Spark (not Mac) |
-| agent_smoke | **38/40 · 95% · 71s** |
-| ship gate (≥38) | **met** |
-| Headline? | **No** — row is pointer + delta |
-| Evidence | `results/sku_unsloth-ud-iq3-s/` |
-| Caveat | Spark runner lacked sanitize / any_of_tools; fails = closed harness IDs on Q4 |
-
-### Remaining order
-
-1. Hold Aug 3 lock: official Q4_K_M + 40/40 + hermes 27/27 + pack docs (**still lead**).
-2. Optional: re-smoke IQ3 with **fixed** harness when GPU free (expect possible 40/40; do not block claims).
-3. Pull `UD-IQ4_XS` if disk allows; smoke only when free.
-4. Never relabel Mac/PC numbers as Spark; never claim iPhone/Android for full Laguna.
-5. DIY imatrix only if it **beats** official/Unsloth on agent_smoke.
-
-## DFlash optional row (measured 2026-07-29) — **DO_NOT_PROMOTE**
-
-| Field | Value |
-|-------|-------|
-| Mode | Q4_K_M + `laguna-s-2.1-DFlash-BF16.gguf` · `--spec-type draft-dflash` |
-| Host | DGX Spark GB10 · engine `04b2b72` |
-| Gen128 | **15.286 tok/s** · 128 tok · 8.374s |
-| Prefill | 2k **2.253s** · 8k **5.556s** |
-| Metrics gen gauge | ~17.3 tok/s predicted |
-| Spec path | slots `speculative=true` · types `none,draft-dflash` |
-| Decision | **DO_NOT_PROMOTE** — slower than plain Q4 baseline (~21.5 t/s gen128) |
-| Evidence | `results/dflash_2026-07-29/` |
-
-Do **not** claim DFlash speedup on this pack. Headline remains plain Q4 serve.
-
-## Live bench stamp
-
-- **Authoritative multi suite (quote this):** gen128 **21.47** tok/s · smoke 40/40 · hermes 27/27 · pack tip `bf82eab` · `results/MEASURED.md` / `results/measured.json`
-- Earlier same-day band sample (not a second claim): gen_bench gen128 **21.016** · prefill 2k 1.022s / 8k 1.521s · unix 1785240774.865332
-- dflash_gen128_tok_s: 15.286 · decision DO_NOT_PROMOTE · 2026-07-29
-- Serve captain: `-c 8192 -ngl -1 --jinja -fa on --alias local-laguna` (SPARK serve blocks aligned 2026-07-29 post-jury fix)
-

research/roadmap-dual-s-and-xs-2026-07-28.md

@@ -1,106 +0,0 @@
-# Dual roadmap — Laguna S (Spark) + Laguna XS (Mac)
-
-Prepared 2026-07-28. Personal founder tracks only.  
-Two roadmaps. One is never a rename of the other.
-
-## North star split
-
-| | **Roadmap A — S / Spark** | **Roadmap B — XS / Mac** |
-|--|---------------------------|-------------------------|
-| Model | Laguna-**S**-2.1 · 118B-A8B | Laguna-**XS**-2.1 · 33B-A3B |
-| Host | **DGX Spark only** (weights) | Founder Mac ≤32G (weights candidate) |
-| Default SKU | official S Q4_K_M ~96G | official XS Q4_K_M ~18.9G |
-| Measured bar today | 40/40 + hermes 27/27 · ~21 t/s | **none** (disk evidence only) |
-| Pack / freeze | `hizrianraz/Laguna-S-2.1-Spark-Agentic` · freeze Aug 2 18:00 · launch Aug 3 12:00 WIB | research parallel; optional later `laguna-xs-*` surface |
-| Founder Mac role for S | **client → Spark :8000** | n/a |
-| Coupling | — | **must not** delay or dilute A |
-
-## Roadmap A — S / Spark (stand-behind)
-
-### Done
-- Official Q4 bind + sha256
-- agent_smoke 40/40 · hermes v2 27/27 on Spark
-- weight_host lock: `dgx-spark-only`
-- founder_mac lock: `client-only-no-local-weights` for full S
-- IQ3_S pointer 38/40 (not headline)
-- Launch lock + personal HF surface
-
-### To freeze (2026-08-02 18:00)
-1. Clean tree + pushed tip for any final STAMP
-2. README / MEASURED / matrix honest; no Mac-S weight claim
-3. Client path smoke from Mac → Spark documented
-4. No DIY GGUF unless measured win
-5. No public promo before Aug 3 12:00
-
-### Non-goals A
-- Local full-S weights on ≤32G Mac
-- Rebrand S as XS
-- Community quants as founder headline
-
-### Kill / hold A
-- Dirty WT for “final”
-- Conflicting jury on host honesty
-- Loss of Q4 restore after alt smoke
-
-## Roadmap B — XS / Mac (parallel research)
-
-### Prepared (this session)
-- Live HF probe: base sha `205dc65…` · GGUF sha `1a37c0…`
-- Q4_K_M sha256 `1ac7079101fca5a6df8c5a7523a3c30ea7d1c0e4b1258090e7d6d4039287f6cb` · 18.882 GiB
-- BF16 non-fit (62.3 GiB) pinned
-- Host: Studio 32G live vs vendor “36G” gap noted
-- Engine: poolside laguna fork / PR #25165 risk
-- Locks: `results/xs_mac_track_lock.json`
-- Notes: `research/Laguna-XS-2.1-Mac-Agentic-fit-2026-07-28.md`
-- Matrix row + definition table updated
-
-### Ordered next (founder-gated each step)
-1. **Jury GO** on prepared XS path + dual roadmap (this consult)
-2. **Founder go** to pull Q4 only → prefer WD_BLACK  
-   ` /Volumes/HFR WD_BLACK SN850X/models/laguna-xs-2.1 `  
-   (System Data ~40 Gi free — soft; external has room)
-3. Build/verify Metal llama on laguna fork (or prove Ollama path)
-4. Load proof: RSS + free RAM + ctx
-5. Same-harness `eval/agent_smoke` → `results/sku_xs-official-q4km/`
-6. Optional hermes-class 27
-7. Fail list + delta vs **not** S headline (separate table)
-8. Founder go before any public/HF claim sentence
-
-### Non-goals B
-- Pull BF16 on Mac
-- DIY re-quant before official fails a real gate
-- Phone/tablet from XS
-- Any claim that XS results are S results
-- Couping B dates into A freeze
-
-### Kill criteria B
-- Cannot load under 32G with usable ctx after OS+Hermes
-- Engine cannot run XS without endless fork chase past freeze bandwidth
-- Smoke << research bar and no learnable delta
-- Disk pressure threatens S freeze work → **park B**
-
-## Shared honesty rules
-
-1. Name the model family every time: **S** or **XS**
-2. Disk size ≠ runtime fit ≠ agent smoke
-3. Poolside public benches ≠ founder hermes harness
-4. Always restore S Q4 on Spark after any alt experiment
-5. Analysis / jury ≠ ship clearance
-
-## Artifact map
-
-| Artifact | Roadmap |
-|----------|---------|
-| `results/launch_lock.json` | A |
-| `results/MEASURED.md` · smoke JSON | A |
-| `results/xs_mac_track_lock.json` | B |
-| `research/Laguna-XS-2.1-Mac-Agentic-fit-2026-07-28.md` | B |
-| `research/device-quant-matrix-aug3.md` | A+B (labeled rows) |
-| `research/post-freeze-smaller-device-path.md` | A + community; B pointer |
-| `research/roadmap-dual-s-and-xs-2026-07-28.md` | A+B (this file) |
-| `prompts/*-dual-roadmap-*.md` | external analysis of both |
-
-## One-line each
-
-- **A:** Ship honest Spark S pack Aug 3; Mac is client.
-- **B:** Prove whether official XS Q4 actually works on founder 32G Mac — separate surface, after pull+smoke, no story theft from A.

results/LAST_GREEN_PIN.md

@@ -12,9 +12,9 @@
 | Gen128 | **~21.47 tok/s** |
 | agent_smoke | **40/40** · 84.86s · temp 0.0 |
 | hermes v2 | **27/27** · 100.1s · temp 0.0 |
-| Pack tip (measured) | `bf82eab` |
+| Measure tip (provenance) | `bf82eab` — not necessarily current docs tip |
 | Serve | `-c 8192 -ngl -1 --parallel 1 --alias local-laguna --jinja -fa on` |
-| Evidence | `results/MEASURED.md`, `results/launch_lock.json`, `results/multi_throughput.json` |
+| Evidence | `results/MEASURED.md`, `results/measured.json` (`throughput` + `quote_gen_tok_s`), `results/launch_lock.json`, `results/server_bench.json` |
 
 ## Explicit rejects
 

results/MEASURED.md

@@ -1,10 +1,10 @@
 # Laguna-S-2.1 on DGX Spark — measured 2026-07-29
 
-personal Spark measurement note — accel multi-bench
+Independent Spark measurement note — multi-bench
 
 ## Stamp
 - When: **2026-07-29 13:22:05 UTC+07:00**
-- Host: `Hizrian-Razs-DGX-Spark`
+- Host: NVIDIA DGX Spark (GB10)
 - Pack git: `bf82eab` (bf82eab5fd6c1fb04e863f0c4b05b5658dec4aee)
 - Headline quant: **Q4_K_M** (held)
 
@@ -40,7 +40,7 @@ Prefill-oriented server_bench (OK replies):
 ## Locks held
 - diy_gguf: false
 - weight_host: Spark-only
-- founder Mac: client-only
+- Mac full-S: client-only
 - public_promo_before_launch: false
 - XS not in S freeze
 

results/agent_smoke.json

@@ -700,7 +700,7 @@
             "type": "function",
             "function": {
               "name": "write_file",
-              "arguments": "{\"path\":\"sandbox/brand.txt\",\"content\":\"personal spark pack\"}"
+              "arguments": "{\"path\":\"sandbox/brand.txt\",\"content\":\"independent spark pack\"}"
             },
             "id": "i4JTNEZ9UPL8cgzzZaeiGzCYwdtU3XJQ"
           }
@@ -734,7 +734,7 @@
     "suite_version": 1,
     "generated_at_utc": "2026-07-29T06:06:55.380678+00:00",
     "host": {
-      "hostname": "Hizrian-Razs-DGX-Spark",
+      "hostname": "dgx-spark",
       "platform": "Linux-6.17.0-1026-nvidia-aarch64-with-glibc2.39",
       "python": "3.12.3"
     },

results/freeze_gate_model_card_lock_set_2026-07-29.json

@@ -7,7 +7,7 @@
   "repo_id": "hizrianraz/Laguna-S-2.1-Spark-Agentic",
   "freeze_at_wib": "2026-08-02T18:00:00+07:00",
   "launch_at_wib": "2026-08-03T12:00:00+07:00",
-  "decision": "gate_filled_continue_prep_not_freeze",
+  "decision": "gate_filled_docs_rewritten_continue_prep_not_freeze",
   "measure_tip": {
     "pack_git_head": "bf82eab5fd6c1fb04e863f0c4b05b5658dec4aee",
     "agent_smoke": "40/40",
@@ -22,51 +22,49 @@
   },
   "model_card": {
     "path": "README.md",
-    "sha256": "b1ef36a5edcc75fc8cbadd9aa0701f1cb3624a90114e9a72d73162f7e40a36d8",
-    "bytes": 10275,
+    "sha256": "9f2603efe7b292c7b7d01386a5d91779aeff089a1d9d6c8bd0bfbc30ecc2748e",
+    "bytes": 9691,
     "fields_filled": [
       "HF YAML frontmatter",
+      "professional table lead",
       "base identity pinned SHAs",
       "official GGUF sha256 table",
       "quant guide + Mac client-only",
       "stranger path",
       "scoreboard live tip 40/40 \u00b7 27/27 \u00b7 21.47 t/s",
       "measured detail + locks",
+      "scope boundaries",
       "disclaimer + attribution"
     ]
   },
   "lock_set": {
     "LAUNCH_AUG3.md": {
-      "sha256": "175c38ccea42405ee3480d299fc3f4c0ef2c3129c3258910a6618034f9dc78a5",
-      "bytes": 5393
+      "sha256": "4904d7cd71ce5a2a72bb063f5ba98f8f60fea3b1f6594213f6e914e57777b079",
+      "bytes": 5676
     },
     "results/launch_lock.json": {
-      "sha256": "330d839e77b00e445d4eba86e7525624e374affa15952973f322412d5d1e4f69",
-      "bytes": 2985
+      "sha256": "b27870a88f396e1dd8aea98618980965e1053ea93e4df3b4c5174bbcf2b7dc0f",
+      "bytes": 3174
     },
     "results/MEASURED.md": {
-      "sha256": "480057b2999e498a88d29ece101dd363724523d10df59b09e979dfc3eb641062",
-      "bytes": 1668
+      "sha256": "18f67a17e00e5aa9c05d8f3d4ab17b77f41d9d7c2e1e53d2d64d79b428242a80",
+      "bytes": 1870
     },
     "results/measured.json": {
       "sha256": "878baffe0f8da0fdc756060d6c65a79177bdd61fb2476270907a0bc5aeb96380",
       "bytes": 3008
     },
     "SHA256SUMS": {
-      "sha256": "ffa7d1c0e882b4c5deac5c748aa0654881351e735102d71a1471749ad6210b1c",
-      "bytes": 1253
-    },
-    "results/xs_mac_track_lock.json": {
-      "sha256": "dc2953263c7a3c7c7f732cbec50a648dd37a76a06a4e9d383c7a7d19743f2f9a",
-      "bytes": 3144
+      "sha256": "90e86226dde3a830f91aaaa53437c631348d297be5e3a89fdfa87dd838f278b0",
+      "bytes": 355
     },
     "results/freeze_readiness_2026-07-29_accel.json": {
       "sha256": "a1011e2f002e03d96c283d3521a2ed19581d744723809cbd9788d1ee808ab962",
       "bytes": 1504
     },
     "results/freeze_notes_2026-07-29.md": {
-      "sha256": "3b8494601350512707d3b6c932e7260643ac218ee6e5d7e0c0205e400e2257ff",
-      "bytes": 1182
+      "sha256": "94071b175f983992a0de7d275deb25d1b2f8df2df959a9f04f8eab3771cf6e27",
+      "bytes": 2288
     },
     "results/agent_smoke.json": {
       "sha256": "587f51f3f7966d9c416e9fbab2a9fc70034106a875e4a18114dd31f2570dd67a",
@@ -80,7 +78,7 @@
   "locks_held": {
     "diy_gguf": false,
     "weight_host": "dgx-spark-only",
-    "founder_mac": "client-only-no-local-weights",
+    "mac_le_32g": "client-only-no-local-weights",
     "public_promo_before_launch": false,
     "xs_not_in_s_freeze": true,
     "branding": "personal-only"
@@ -97,5 +95,11 @@
     "stranger-path dry",
     "push filled card+locks to GH/HF when network ready",
     "hold tip through Aug 2 18:00 freeze"
-  ]
+  ],
+  "xs_in_s_freeze": false,
+  "note": "S freeze lock_set is S-only. XS track is parallel research, not hashed into S freeze. Hashes are interim after professional README rewrite; final freeze rehash still required at freeze gate on clean pushed HEAD. rehash_at_freeze remains true.",
+  "rehash_at_freeze": true,
+  "rehashed_at_wib_doc": "2026-07-29T18:21:21+07:00",
+  "rehashed_at_wib": "2026-07-29T18:21:21+07:00",
+  "rehash_reason": "professional README rewrite + interim lock_set rehash; final freeze rehash still required at Aug 2 18:00 local on clean pushed HEAD"
 }

results/freeze_notes_2026-07-29.md

@@ -5,8 +5,8 @@
 - Freeze gate: **model_card + lock files**
 - Status: **FILLED** at `2026-07-29T14:43:26+07:00`
 - Receipt: `results/freeze_gate_model_card_lock_set_2026-07-29.json`
-- Freeze clock: still **2026-08-02 18:00 WIB** (do not freeze early)
-- HF go-live: **2026-08-03 12:00 WIB**
+- Freeze clock: still **2026-08-02 18:00 local** (do not freeze early)
+- HF go-live: **2026-08-03 12:00 local**
 
 ## Live tip bound into card + locks
 
@@ -23,7 +23,7 @@
 
 - diy_gguf: false
 - weight_host: Spark-only
-- founder Mac: client-only (no local S)
+- Mac ≤32 GB: client-only (no local S)
 - public_promo_before_launch: false
 - XS: parallel, not in S freeze
 - branding: personal-only
@@ -33,15 +33,23 @@
 
 1. ~~Optional DFlash status row final~~ → **done 2026-07-29 · DO_NOT_PROMOTE** (`results/dflash_2026-07-29/`)
 2. ~~Stranger-path dry re-run~~ → **DRY PASS** (`results/stranger_path_dry_2026-07-29.md`) · serve flag aligned to `-ngl -1`
-3. ~~Baseline serve re-health after DFlash teardown~~ → **HEALTH_OK + chat OK** 2026-07-29 15:50 WIB (`results/baseline_rehealth_2026-07-29.md`)
+3. ~~Baseline serve re-health after DFlash teardown~~ → **HEALTH_OK + chat OK** 2026-07-29 15:50 local (`results/baseline_rehealth_2026-07-29.md`)
 4. HF tip push of DFlash/last-green/stranger/jury receipts (docs-only)
-5. Clean pushed HEAD on freeze day (**2026-08-02 18:00 WIB**)
-6. Public go-live **2026-08-03 12:00 WIB** (no early promo)
+5. Clean pushed HEAD on freeze day (**2026-08-02 18:00 local**)
+6. Public go-live **2026-08-03 12:00 local** (no early promo)
 
 ## Non-actions
 
 - No accidental freeze today
-- No public promo before Aug 3 12:00 WIB
+- No public promo before Aug 3 12:00 local
 - No XS inside S freeze artifact
 - No re-bench required unless weights/runners change
 - No DFlash speedup claim
+
+## 2026-07-29 jury-fix rev (post dual seat)
+
+- **S freeze lock_set is S-only.** Dropped `results/xs_mac_track_lock.json` from S `lock_set`.
+- XS remains parallel `prepared_research` under its own lock; **not** an S freeze hash input.
+- **Rehash at freeze:** on clean pushed HEAD (not dirty WT), recompute sha256 for every S lock_set member and stamp freeze gate. Docs tip may move; measure tip stays `bf82eab` until smoke/weights change.
+- Gen headline sole: **~21.47 tok/s** from multi suite in `results/measured.json`. No 22.50 on launch surfaces.
+- SHA256SUMS is GNU **2-column** for `sha256sum -c`.

results/freeze_readiness_2026-07-29_accel.json

@@ -22,7 +22,7 @@
   "locks_held": {
     "diy_gguf": false,
     "weight_host": "dgx-spark-only",
-    "founder_mac": "client-only-no-local-weights",
+    "mac_le_32g": "client-only-no-local-weights",
     "public_promo_before_launch": false,
     "xs_not_in_s_freeze": true,
     "branding": "personal-only"
@@ -34,7 +34,7 @@
     "final clean HEAD push on freeze day"
   ],
   "blockers_to_freeze_now": [
-    "not yet freeze clock (Aug 2 18:00 WIB)"
+    "not yet freeze clock (Aug 2 18:00 local)"
   ],
   "receipt": "results/freeze_gate_model_card_lock_set_2026-07-29.json"
 }

results/hermes_agent_smoke.json

@@ -577,7 +577,7 @@
     "suite_version": 2,
     "generated_at_utc": "2026-07-29T06:08:57.571350+00:00",
     "host": {
-      "hostname": "Hizrian-Razs-DGX-Spark",
+      "hostname": "dgx-spark",
       "platform": "Linux-6.17.0-1026-nvidia-aarch64-with-glibc2.39",
       "python": "3.12.3"
     },

results/hf_publish.json

@@ -1,9 +1,9 @@
 {
   "repo_id": "hizrianraz/Laguna-S-2.1-Spark-Agentic",
   "url": "https://huggingface.co/hizrianraz/Laguna-S-2.1-Spark-Agentic",
-  "commit": "4296543781f6318edd356e83754479cd65f326a1",
-  "latest_commit": "4296543781f6318edd356e83754479cd65f326a1",
-  "local_git_head": "921ff6fb573036bdb2f2f81debe4c1cc15ae54ab",
+  "commit": "57aa78584ad62ed5c75c0cf8b158aebb70b26da8",
+  "latest_commit": "57aa78584ad62ed5c75c0cf8b158aebb70b26da8",
+  "local_git_head": "63488eccb16efbaf670aabee31b22faa662aab2c",
   "private": false,
   "siblings": [
     ".gitattributes",
@@ -28,7 +28,9 @@
     "hermes/README.md",
     "hermes/sample_client.py",
     "prompts/_mirror/README-SPARK-HIZRIANRAZ.md",
+    "prompts/chatgpt-analyze-laguna-s-2026-07-29.md",
     "prompts/chatgptwork-laguna-s-plus-xs-2026-07-28.md",
+    "prompts/claude-analyze-laguna-s-2026-07-29.md",
     "prompts/claudecowork-laguna-s-plus-xs-2026-07-28.md",
     "research/Laguna-XS-2.1-Mac-Agentic-fit-2026-07-28.md",
     "research/agentic-quant-hermes-hf-trending-2026-07-28.md",
@@ -39,9 +41,18 @@
     "research/quant-comparison-scoreboard-2026-07-28.md",
     "research/roadmap-dual-s-and-xs-2026-07-28.md",
     "results/.gitkeep",
+    "results/LAST_GREEN_PIN.md",
     "results/MEASURED.md",
+    "results/accel_receipts_2026-07-29.md",
     "results/agent_smoke.json",
     "results/agent_smoke.log",
+    "results/baseline_rehealth_2026-07-29.md",
+    "results/dflash_2026-07-29/STATUS.md",
+    "results/dflash_2026-07-29/measured.json",
+    "results/dflash_2026-07-29/meta.txt",
+    "results/dflash_2026-07-29/models.json",
+    "results/dflash_2026-07-29/server_bench.json",
+    "results/dflash_2026-07-29/throughput.json",
     "results/dual_audit_consolidation_2026-07-28.md",
     "results/engine_sha.txt",
     "results/freeze_gate_model_card_lock_set_2026-07-29.json",
@@ -57,6 +68,8 @@
     "results/sku_unsloth-ud-iq3-s/agent_smoke.json",
     "results/sku_unsloth-ud-iq3-s/agent_smoke.log",
     "results/sku_unsloth-ud-iq3-s/sku_meta.json",
+    "results/stranger_path_dry_2026-07-29.md",
+    "results/three_jury_post_dflash_2026-07-29.md",
     "results/xs_mac_track_lock.json",
     "scripts/_stamp_accel_receipts.py",
     "scripts/bench_server.py",
@@ -64,35 +77,39 @@
     "scripts/pull_sku.sh",
     "scripts/serve_spark.sh"
   ],
-  "n_files": 57,
+  "n_files": 70,
   "launch_locked": true,
   "launch_at_wib": "2026-08-03T12:00:00+07:00",
   "freeze_at_wib": "2026-08-02T18:00:00+07:00",
   "freeze_gate_status": "filled",
   "freeze_gate_receipt": "results/freeze_gate_model_card_lock_set_2026-07-29.json",
   "measure_tip_pack_git": "bf82eab5fd6c1fb04e863f0c4b05b5658dec4aee",
+  "docs_pack_git_head_at_packet": "63488eccb16efbaf670aabee31b22faa662aab2c",
   "measured_json_on_hf": true,
   "agent_smoke_on_hf": true,
   "hermes_agent_smoke_on_hf": true,
+  "dflash_status_on_hf": true,
+  "last_green_pin_on_hf": true,
+  "analyze_prompts_on_hf": true,
   "freeze_gate_on_hf": true,
-  "freeze_readiness_on_hf": true,
   "launch_lock_on_hf": true,
   "readme_on_hf": true,
-  "updated_at_wib": "2026-07-29T14:52:10+07:00",
-  "note": "docs/receipts only; weights untouched; freeze gate filled; launch lock still Aug 3 12:00 until founder re-lock",
+  "jury_md_on_hf": true,
+  "jury_json_on_hf": false,
+  "jury_json_policy": "local_and_github_only_by_pack_gitignore_design",
+  "spark_ngl_fix_on_hf": true,
+  "scoreboard_dualgen_fix_on_hf": true,
+  "updated_at_wib": "2026-07-29T16:20:55+07:00",
+  "note": "docs tip after jury fix rev2; SPARK -ngl -1 + dual-gen close on HF; three_jury JSON intentionally not public (gitignore); MD jury GO is public; weights untouched; DFlash DO_NOT_PROMOTE; launch Aug 3 12:00 local",
   "uploaded_paths": [
-    "README.md",
-    "LAUNCH_AUG3.md",
-    "SHA256SUMS",
-    ".gitignore",
-    "results/launch_lock.json",
-    "results/measured.json",
+    "SPARK.md",
+    "research/quant-comparison-scoreboard-2026-07-28.md",
+    "results/stranger_path_dry_2026-07-29.md",
+    "results/three_jury_post_dflash_2026-07-29.md",
+    "results/LAST_GREEN_PIN.md",
     "results/MEASURED.md",
-    "results/agent_smoke.json",
-    "results/hermes_agent_smoke.json",
-    "results/freeze_notes_2026-07-29.md",
-    "results/freeze_readiness_2026-07-29_accel.json",
-    "results/freeze_gate_model_card_lock_set_2026-07-29.json",
-    "results/xs_mac_track_lock.json"
-  ]
+    "LAUNCH_AUG3.md"
+  ],
+  "upload_return": "https://huggingface.co/hizrianraz/Laguna-S-2.1-Spark-Agentic/commit/57aa78584ad62ed5c75c0cf8b158aebb70b26da8",
+  "jury_overall": "GO_PASS_CONTINUE_PREP"
 }

results/launch_lock.json

@@ -26,7 +26,7 @@
   "hermes_agent_smoke_artifact": "results/hermes_agent_smoke.json",
   "hermes_agent_smoke_runner_sha256": "20c1e52a8a22306b81309c08984d6875cfff46e7382a46b674c1a6345030293b",
   "quote_gen_tok_s": 21.47,
-  "quote_gen_label": "gen128 @ 128 completion \u00b7 multi_throughput",
+  "quote_gen_label": "gen128 @ 128 completion \u00b7 measured.json multi suite",
   "measured_json": "results/measured.json",
   "measured_md": "results/MEASURED.md",
   "multi_device_track": "unlocked_pointer_docs",
@@ -34,8 +34,8 @@
   "iq3_not_headline": true,
   "weight_host": "dgx-spark-only",
   "weight_host_lock": "2026-07-28-spark-only",
-  "founder_mac": "client-only-no-local-weights",
-  "founder_mac_ram_class_gb": 32,
+  "mac_le_32g": "client-only-no-local-weights",
+  "mac_le_32g_ram_class_gb": 32,
   "diy_gguf": false,
   "branding": "personal-only",
   "public_promo_before_launch": false,
@@ -54,12 +54,12 @@
     "results/MEASURED.md",
     "results/measured.json",
     "SHA256SUMS",
-    "results/xs_mac_track_lock.json",
     "results/freeze_readiness_2026-07-29_accel.json",
     "results/freeze_gate_model_card_lock_set_2026-07-29.json"
   ],
   "measure_pack_git_head": "bf82eab5fd6c1fb04e863f0c4b05b5658dec4aee",
   "measure_pack_git_short": "bf82eab",
   "docs_pack_git_head_at_fill": "8ccc4536f4628791bd5fcb20e2e69e0e0bd187f4",
-  "docs_note": "measure tip sticks at bf82eab (smoke/weights unchanged); docs tip moves with builds"
+  "docs_note": "measure tip sticks at bf82eab (smoke/weights unchanged); docs tip moves with builds. S freeze lock_set is S-only; XS parallel track is separate prepared_research pointer, not a freeze hash input. Rehash lock_set artifacts at freeze on clean pushed HEAD.",
+  "lock_set_policy": "s_only_rehash_at_freeze"
 }

results/measured.json

@@ -1,6 +1,6 @@
 {
   "ts": 1785306125.3880131,
-  "host": "Hizrian-Razs-DGX-Spark",
+  "host": "dgx-spark",
   "engine": "poolsideai/llama.cpp 04b2b72",
   "model": "laguna-s-2.1-Q4_K_M.gguf",
   "sha256": "a8b55c75714ea73fd90ec85de5defdc0b8d88ca0ad2108343cdd8fc22f7583e4",
@@ -54,7 +54,7 @@
     "temperature": 0.0,
     "model": "local-laguna",
     "base_url": "http://127.0.0.1:8000/v1",
-    "host": "Hizrian-Razs-DGX-Spark",
+    "host": "dgx-spark",
     "out": "results/agent_smoke.json",
     "runner_sha256": "3bb81080879ddf78d3d6295e333c78b1ecc1d84ba5a9e9dc19a0e677f54da48f",
     "run_manifest_schema": "laguna.run_manifest/v1",
@@ -66,7 +66,7 @@
     "temperature": 0.0,
     "model": "local-laguna",
     "base_url": "http://127.0.0.1:8000/v1",
-    "host": "Hizrian-Razs-DGX-Spark",
+    "host": "dgx-spark",
     "out": "results/hermes_agent_smoke.json",
     "runner_sha256": "20c1e52a8a22306b81309c08984d6875cfff46e7382a46b674c1a6345030293b",
     "run_manifest_schema": "laguna.run_manifest/v1",

results/sku_unsloth-ud-iq3-s/README.md

@@ -14,7 +14,7 @@ Do not brand “same-harness vs Q4 40/40” until IQ3 re-runs on the fixed runne
 | bytes | 48428911520 (~45.1 Gi) |
 | Engine | poolside llama.cpp laguna `04b2b72` |
 | Serve | `:8000` briefly (Q4 stopped) · ctx 8192 · alias `local-laguna-iq3s` |
-| agent_smoke | **38/40 · 95% · 71s** (2026-07-28 20:10 WIB) |
+| agent_smoke | **38/40 · 95% · 71s** (2026-07-28 20:10 local) |
 | fails | `repair_04`, `long_06` — same IDs as prior Q4 harness bugs; older runner |
 | ship gate | ≥ official−2 → **met** on runner that ran |
 | Q4 restored | yes, alias `local-laguna` on `:8000` after measure |

results/xs_mac_track_lock.json

@@ -26,12 +26,12 @@
     "bytes": 66930226304,
     "gib": 62.334,
     "sha256": "d1d99abe30c37749ec1b1ae2681cec74caa550ac6243d2ea70b61b1ff9d187ca",
-    "role": "non_fit_founder_mac"
+    "role": "non_fit_mac_le_32g"
   },
   "target_host": {
-    "class": "founder_mac_le_32g",
+    "class": "mac_le_32g_le_32g",
     "live_studio": {
-      "model": "Mac14,13",
+      "model": "Apple Silicon Mac",
       "ram_bytes": 34359738368,
       "ram_gb": 32
     },
@@ -48,17 +48,17 @@
     "agent_smoke_local": null,
     "hermes_smoke_local": null,
     "tps_local": null,
-    "gate_before_any_claim": "same-harness agent_smoke on founder Mac + sha256 + host label"
+    "gate_before_any_claim": "same-harness agent_smoke on Mac \u226432 GB + sha256 + host label"
   },
   "storage_policy": {
-    "prefer_pull_volume": "/Volumes/HFR WD_BLACK SN850X/models/laguna-xs-2.1",
+    "prefer_pull_volume": "/path/to/external/models/laguna-xs-2.1",
     "avoid_system_data_when_full": true,
     "system_data_free_gi_at_prep": 40,
-    "do_not_pull_without_founder_go": true,
+    "do_not_pull_without_operator_go": true,
     "do_not_pull_bf16_on_mac": true
   },
   "namespace": {
-    "notes_in_s_pack": "research/ only — not S freeze artifact",
+    "notes_in_s_pack": "research/ only \u2014 not S freeze artifact",
     "optional_future_surface": "hizrianraz/Laguna-XS-2.1-Mac-Agentic (not created)",
     "forbidden_aliases": [
       "S on Mac",
@@ -70,7 +70,7 @@
   "relation_to_s_pack": {
     "s_repo": "hizrianraz/Laguna-S-2.1-Spark-Agentic",
     "s_weight_host": "dgx-spark-only",
-    "s_founder_mac": "client-only-no-local-weights",
+    "s_mac_le_32g": "client-only-no-local-weights",
     "s_freeze_wib": "2026-08-02T18:00:00+07:00",
     "s_launch_wib": "2026-08-03T12:00:00+07:00",
     "xs_launch_coupling": false,
@@ -79,14 +79,14 @@
   "honest_language": {
     "allowed": [
       "XS Q4_K_M is official ~19G GGUF for Laguna-XS-2.1",
-      "founder Mac client still talks to Spark for full S",
+      "Mac \u226432 GB client still talks to Spark for full S",
       "XS is a separate 33B-A3B model"
     ],
     "forbidden": [
       "Laguna runs on Mac (unqualified)",
       "S on Mac via XS",
       "any unmeasured t/s or agent_smoke for XS",
-      "Poolside public benches as founder measured bar",
+      "Poolside public benches as independently measured bar",
       "BF16 / safetensors as Mac-local"
     ]
   },

scripts/_stamp_accel_receipts.py

@@ -10,7 +10,7 @@ import time
 from datetime import datetime, timedelta, timezone
 from pathlib import Path
 
-WIB = timezone(timedelta(hours=7))
+local = timezone(timedelta(hours=7))
 
 
 def sha_file(p: Path) -> str:
@@ -45,7 +45,7 @@ def summary_from(obj: dict, fallback_pass: str, fallback_el: float):
 
 def main() -> None:
     root = Path(__file__).resolve().parents[1]
-    now = datetime.now(WIB)
+    now = datetime.now(local)
     agent = json.loads((root / "results/agent_smoke.json").read_text())
     hermes = json.loads((root / "results/hermes_agent_smoke.json").read_text())
     server = json.loads((root / "results/server_bench.json").read_text())
@@ -130,7 +130,7 @@ def main() -> None:
     lines = [
         f"# Laguna-S-2.1 on DGX Spark — measured {day}",
         "",
-        "personal Spark measurement note — accel multi-bench",
+        "Spark measurement note — accel multi-bench",
         "",
         "## Stamp",
         f"- When: **{stamp}**",
@@ -177,7 +177,7 @@ def main() -> None:
         "## Locks held",
         "- diy_gguf: false",
         "- weight_host: Spark-only",
-        "- founder Mac: client-only",
+        "- Mac ≤32 GB: client-only",
         "- public_promo_before_launch: false",
         "- XS not in S freeze",
         "",

scripts/measure_layer_b_v3.sh

@@ -0,0 +1,68 @@
+#!/usr/bin/env bash
+# Run Layer B hermes v3 suite against a live captain serve.
+# Does NOT start GPU/weights. Does NOT rewrite launch_lock.json.
+# Fail-closed if /v1/models is down.
+set -euo pipefail
+
+ROOT="$(cd "$(dirname "$0")/.." && pwd)"
+BASE_URL="${OPENAI_BASE_URL:-http://127.0.0.1:8000/v1}"
+BASE_URL="${BASE_URL%/}"
+MODEL="${OPENAI_MODEL:-local-laguna}"
+OUT="${LAGUNA_LAYER_B_OUT:-$ROOT/results/hermes_agent_smoke_layer_b_v3.json}"
+CASES="$ROOT/eval/hermes_agent_smoke/cases_layer_b_v3.json"
+UA="${LAGUNA_UA:-laguna-layer-b-measure/0.1}"
+
+echo "pack=$ROOT"
+echo "base_url=$BASE_URL model=$MODEL"
+echo "cases=$CASES"
+echo "out=$OUT"
+
+if [[ ! -f "$CASES" ]]; then
+  echo "missing cases: $CASES" >&2
+  exit 2
+fi
+
+echo "probe /models …"
+code="$(curl -sS -m 8 -o /tmp/laguna_models_probe.json -w '%{http_code}' \
+  -H "User-Agent: ${UA}" \
+  -H "Authorization: Bearer ${OPENAI_API_KEY:-sk-local}" \
+  "${BASE_URL}/models" || true)"
+if [[ "$code" != "200" ]]; then
+  echo "serve down (HTTP ${code:-curl-fail}) at ${BASE_URL}/models" >&2
+  echo "restore on Spark first:" >&2
+  echo "  ./scripts/serve_spark.sh" >&2
+  echo "  # or LAST_GREEN_PIN restore block" >&2
+  exit 3
+fi
+echo "models ok: $(head -c 180 /tmp/laguna_models_probe.json)"
+echo
+
+python3 "$ROOT/eval/hermes_agent_smoke/run_hermes_smoke.py" \
+  --base-url "$BASE_URL" \
+  --model "$MODEL" \
+  --cases "$CASES" \
+  --temperature 0 \
+  --out "$OUT"
+
+echo
+echo "wrote $OUT"
+if command -v python3 >/dev/null; then
+  OUT="$OUT" python3 - <<'PY'
+import json, os
+from pathlib import Path
+p = Path(os.environ["OUT"])
+d = json.loads(p.read_text())
+n = d.get("passed", d.get("n_pass", d.get("n_ok", "?")))
+t = d.get("n", d.get("n_total", d.get("total", "?")))
+if (n == "?" or t == "?") and isinstance(d.get("results"), list):
+    cases = d["results"]
+    t = len(cases)
+    n = sum(1 for c in cases if c.get("pass") or c.get("ok") or c.get("passed"))
+elif (n == "?" or t == "?") and isinstance(d.get("cases"), list):
+    cases = d["cases"]
+    t = len(cases)
+    n = sum(1 for c in cases if c.get("pass") or c.get("ok") or c.get("passed"))
+print(f"summary: {n}/{t}  (freeze bar unchanged: hermes v2 27/27 + agent_smoke 40/40)")
+print("do not rewrite launch_lock.json from this artifact")
+PY
+fi

scripts/pull_official_gguf.sh

@@ -15,7 +15,7 @@ if [[ ! -f "${SUMS}" ]]; then
   exit 2
 fi
 
-# Pack format: sha256  byte_size  filename  (3-column). Accept 2-col sha256sum too.
+# Pack format: GNU sha256sum 2-column (sha256  filename). Comments OK. awk accepts legacy 3-col too.
 expected="$(
   awk -v f="${FILE}" '
     $0 ~ /^[[:space:]]*#/ { next }

scripts/serve_spark.sh

@@ -1,6 +1,6 @@
 #!/usr/bin/env bash
 # Serve official Laguna-S-2.1 GGUF on DGX Spark via poolside llama.cpp (laguna).
-# personal DGX Spark serve helper
+# DGX Spark serve helper
 set -euo pipefail
 
 ROOT="${LAGUNA_ENGINE:-$HOME/src/llama.cpp-laguna}"