results/three_jury_post_dflash_2026-07-29.md

Three-jury consult — post-DFlash (2026-07-29) · rev2 fix-pass

Packet: three_jury_post_dflash_2026-07-29
Mode: Manwë three-seat appellate · fix-if-not-go
Measure tip: bf82eab · Engine: 04b2b72 · Weight: official Q4_K_M
Live Spark: HEALTH_OK · flags match last-green (-ngl -1 · jinja · fa on · alias local-laguna)

Bottom line

GO · PASS_CONTINUE_PREP

Named jury gaps from rev1 are fixed in-tree this tick:

1. SPARK default serve / DFlash / llama-bench: -ngl 99 → -ngl -1
2. Scoreboard [dual gen] closed: ~21.47 authoritative · 21.016 band-only
3. Stranger dry re-verified PASS
4. DFlash remains DO_NOT_PROMOTE
5. Ship clock still 2026-08-03 12:00 WIB · freeze 2026-08-02 18:00 WIB

Not freeze day. No claim expansion.

Seats

Fable5 — measure

• PASS
• 15.286 < ~21.5 forces DO_NOT_PROMOTE
• Baseline re-health HEALTH_OK after DFlash teardown
• Dual gen claim closed in scoreboard
• Caveat kept: no full 40/40 re-smoke on post-rename tip this window (measure remains tip-bound bf82eab)

Sol5.6 — serve

• PASS
• Live captain = last-green flags (probed)
• SPARK stranger serve blocks now match captain
• DFlash off default path
• XS HF pending_open is OK (S can ship alone)

Composer2.5 — pack/claims

• PASS
• Forbidden claims listed
• Stranger dry PASS after SPARK fix
• HF must re-tip after this commit (docs-only) — in next unblocked

Lórien chair

Not re-queued this tick (prior timeout ≠ fail). Appellate stands.

External

• ChatGPT: prompts/chatgpt-analyze-laguna-s-2026-07-29.md
• Claude: prompts/claude-analyze-laguna-s-2026-07-29.md
  Founder paste → return findings become extra seats.

Allowed claims

• Spark agent_smoke 40/40 official Q4_K_M
• hermes 27/27
• ~21.5 t/s gen128 plain Q4 Spark (authoritative multi suite 21.47)
• official digest-bound Q4 (no bare re-upload)

Forbidden

• DFlash speedup
• first/only Laguna quant
• best universal tok/s
• full Laguna phone
• IQ3 headline
• global HF top-10

Next

1. Commit + push GH tip (this fix)
2. HF docs-only tip (SPARK + scoreboard + jury + stranger)
3. Run external prompts when ready
4. Optional T-3 polish only if external opens a real gap