# External seat — ChatGPT — 2026-07-29 evening WIB

**Seat:** chatgpt (external adversarial, dual pack)
**Evidence claimed:** local + GH + HF; not live Spark; not raw multi_throughput file
**Ingested:** external review paste 2026-07-29 ~17:23 WIB
**Manwë live re-probe after paste:** XS HF bleed already cleared (see dual synth)

## S verdict (seat)

**SHIP_WITH_NITS** — on schedule, not freeze-ready as-is.

**One-line story (seat):** Personal, unaffiliated DGX Spark runtime pack for Poolside’s official Q4_K_M GGUF: pinned llama.cpp serve, 40/40 agent_smoke, 27/27 Hermes-class smoke, and 21.47 tok/s gen128.

## XS verdict (seat)

**HOLD_PREP** — local/GH mostly honest; live HF contaminated at seat time (Manwë re-probe later: fixed).

## BANNED (union-ready)

- First/only Laguna quant
- Fastest/best Laguna GGUF
- DFlash speedup
- Full Laguna-S on Mac/mobile
- XS is S-lite / S-Mac / part of S freeze
- Laguna runs on Mac (without naming XS)
- XS stand-behind / 32 GB fit
- Any XS smoke or tok/s before raw evidence
- BF16/safetensors fits locally

## S contradictions named

1. SPARK.md older smoke durations vs MEASURED 84.86s / 100.1s
2. SPARK ~21 quote vs LAUNCH_AUG3 22.50 parenthetical — sole headline 21.47
3. DFlash used -ngl 99 vs baseline -ngl -1 — reject still fair, not perfect A/B
4. engine_sha local_patch wording drift vs SPARK.md
5. bf82eab labeled pack/live tip — should be measurement provenance tip

## Dual gen

**Narrow yes** if 21.016 stays band-only and 22.50 + older ~21.1 leave headline surface.

## Stranger

**FAIL overall (seat):** no checkout of engine 04b2b72; invalid three-column `sha256sum -c` flow; smoke uses `--model laguna-q4` after `--alias local-laguna`; dry receipt skipped clean rebuild.

## S RISKS_TOP5 (seat)

1. Sibling XS HF serving S card+lock (seat-time; re-probe clear)
2. Advertised stranger path can fail before headline
3. Freeze manifest stale post-rev2 + includes XS lock inside S lock set
4. LAST_GREEN_PIN cites missing multi_throughput.json; baseline health MD-only
5. Competitive wedge narrow — Spark ops evidence, not new quant / beat Poolside-Unsloth-Bartowski; IQ3 below fold

## Freeze checklist (seat) — shortened keepers

- No freeze before 2026-08-02 18:00 · no promote before 2026-08-03 12:00 WIB
- Remove S material from public XS (done at re-probe)
- Clean stranger path: engine pin + checksum flow + alias match
- Raw source for 21.47 or drop dead pointer
- Bound claims only: 21.47 · 40/40 @ 84.86s · 27/27 @ 100.1s
- Final S-only hash-valid manifest after doc edits
- Clean git · pushed HEAD · GH/HF parity · measure tip ≠ docs tip labels
- Jury MD public · jury JSON intentionally absent

## XS overclaims / bleed (seat)

- Live HF README+lock were S (re-probe: honest XS)
- local lock still has `stand_behind_sku` while disclaiming ship
- “No weight pull” stale vs Synology placement receipt (storage ≠ Mac runtime)
- Hermes fixture injects ~21 tok/s host assertion
- Agent cases “independent spark pack” + SPARK.md refs

## Confidence (seat)

- S: med (no live Spark / no raw throughput file)
- XS: high on bleed at seat time

## Disconfirming (seat)

Aug 3 story embarrassed if XS HF still presents S while stranger PASS was never end-to-end.