# Dual external seats — ChatGPT + Claude — synth 2026-07-29 17:25 WIB

**Inputs**
- Claude cowork dual: `external_seat_claude_cowork_2026-07-29.md` (filed earlier)
- ChatGPT dual: external paste → `external_seat_chatgpt_2026-07-29.md`
- Manwë live HF re-probe after ChatGPT paste: XS README + lock **honest XS** (bleed closed)

## Joint verdict

| Pack | ChatGPT | Claude | Manwë joint |
|---|---|---|---|
| S Spark | SHIP_WITH_NITS | PASS_WITH_NITS | **SHIP_WITH_NITS** — schedule OK · **not freeze-ready** |
| XS Mac | HOLD_PREP | HOLD_PREP→PREP_NITS after HF swap | **PREP_NITS** on public surface · still **no measure/pull authority** |

Clocks unchanged: freeze **2026-08-02 18:00 WIB** · HF public **2026-08-03 12:00 WIB**.

## Agreement (both seats)

1. Headline tuple holds: **40/40 · 27/27 · 21.47** on official Q4 · independent Spark pack.
2. DFlash stays **DO_NOT_PROMOTE**.
3. Stranger path is **not** checkout-clean (checksum form · alias/model mismatch · dry skipped rebuild).
4. Freeze gate receipt stale + **XS lock file inside S lock_set** must go before freeze.
5. LAST_GREEN_PIN dead `multi_throughput.json` pointer.
6. XS must never read as S-lite / Development of S freeze.
7. IQ3 / competitive claims stay below fold.

## ChatGPT-only (Claude lighter or silent)

1. SPARK.md **old smoke durations** vs MEASURED **84.86 / 100.1**.
2. LAUNCH **22.50** parenthetical still on headline surface.
3. DFlash **-ngl 99 vs -1** — reject fair, not perfect causal A/B; disclose.
4. engine_sha **patch prose drift** (`math.h`+`::isfinite` vs `cmath`+`std::isfinite`).
5. Label **bf82eab = measurement provenance tip**, not moving docs HEAD.
6. Baseline re-health is **MD attestation only**.
7. XS fixtures: Hermes **~21 tok/s** inject · agent **spark pack** wording · `stand_behind_sku` key · stale no-pull vs placement receipt.
8. Confidence split: S med / XS high.

## Claude-only (still open after HF fix)

1. S `hf_publish.json` stale (if still true on tip).
2. Ranked nits list oriented to post-bleed cleanup order.

## Live state delta since Claude seat

| Item | Claude time | Now |
|---|---|---|
| XS HF README/lock S-bleed | CONFIRMED S | **CLEARED** — XS unmeasured card + prep lock |
| S stranger docs | open | still open |
| Freeze lock_set has XS file | open | still open |
| multi_throughput.json | missing | still missing |

## Ranked fix plan (docs-only unless noted)

### P0 — before any freeze talk

1. **Stranger path executable**
 - Fix SHA256SUMS consumer: either 2-col file for `sha256sum -c` **or** document `awk '{print $1, $3}'` / dedicated check script (3-col is intentional LFS+bytes format — **don't lie with bare -c**).
 - README smoke: `--model local-laguna` to match serve alias (or document both).
 - REPRODUCE/BUILD: pin checkout of engine `04b2b72` + named host patch.
 - Re-label stranger receipt: **docs-dry**, not full stranger PASS, until clean clone path runs on Spark (GPU run = operator/Spark time, not required tonight if docs honest).

2. **Freeze manifest S-only + rehash after edits**
 - Drop `results/xs_mac_track_lock.json` from S `lock_set`.
 - Regenerate hashes on final doc set only at freeze window (not now as false “filled forever”).

3. **Dead evidence pointers**
 - LAST_GREEN_PIN: remove `multi_throughput.json` **or** restore raw file from authoritative suite.
 - Prefer point at existing measured/multi suite paths that actually exist.

### P1 — claim surface (same commit OK)

4. LAUNCH_AUG3: drop **22.50** parenthetical from gen quote line.
5. SPARK.md table: sole public gen line **~21.47**; move ~21.1 to historical/band; smoke rows → **84.86s / 100.1s** bound pair.
6. DFlash STATUS/measured: disclose flag mismatch; no absent server-log cites; DO_NOT_PROMOTE unchanged.
7. engine_sha ↔ SPARK patch one-liner align.
8. Everywhere: `bf82eab` = **measurement provenance tip**.

### P2 — XS (no Mac measure until after S go-live)

9. HF already clean — issue fresh `hf_publish.json` receipt when next XS docs tip lands.
10. Rename/remove `stand_behind_sku`; fix stale no-pull prose vs placement receipt (storage only).
11. Neutralize S/Spark facts in XS `eval/**/cases.json` before any future run.
12. Alias standard already `local-laguna-xs` in serve_mac.sh — herd runners/sample_client to same.

### P3 — leave alone

- Weight bytes / engine binary / captain flags / smoke JSON bodies
- Jury JSON public
- Freeze/launch clocks
- Company vault / Neptune gates

## Bound public claims (locked)

- Spark agent_smoke **40/40** · **84.86s**
- hermes **27/27** · **100.1s**
- gen128 **~21.47** tok/s (authoritative multi suite)
- official Q4_K_M digest-bound
- personal unaffiliated pack

## Forbidden (union)

All phrases from both seats’ banned lists (see ChatGPT seat file).

## Disconfirming joint view

The launch embarrasses if (a) stranger instructions still cannot reproduce, or (b) any public surface re-bleeds S into XS, or (c) freeze manifest still hashes sibling XS locks into S.

## Recommended next (one decision)

**Docs fix-pass now (P0+P1 on S tree)** → commit/push GH → HF docs-only S tip → optional stranger GPU only if you want a true PASS receipt before Aug 2.

XS: no measure, no pull action — cleanup queue only after S tip stable.