2026-09-12: The Spyglass on the Max

Bert asked which model the M4 Max 128 GB box should actually run, then asked whether the answer was EETISMAD. It was not. The weights were on disk, the launcher existed, and the steering flag had been rejected at engine open so the first A/B never applied the bank. Vision was a 588 MB file nobody had pointed the agent at. This note is the closeout.

The flag that did nothing

Directional steering already existed for DeepSeek V4 Flash and GLM 5.3 Flash: one normalized f32 direction per layer, applied as y = y - scale * dir * dot(dir, y). Qwen3.8 is 48 × 2560. The Qwen Metal graph never called the projection kernel, and engine open printed --mtp-model, steering, SSD streaming and --power are not supported for Qwen3.8 and refused the file. The first A/B looked like ablation was a no-op because it was. The gate now allows steering; SSD streaming, --mtp-model, and --power stay rejected.

The bank lives at ~/.sanctum/ablation/qwen38-flash-next-refusal.f32 (491,520 bytes). It is not in the ds4 repo. Sixteen last-token HC-mean pairs, --pair-normalize, scale 1.0 — not the dense-27B cliff at α=2.0. Capability prompts (CSV helper, SSRI paragraph, print("hello")) still answered. Two over-refusal prompts that vanilla hard-refused engaged after the bank. That is a four-prompt smoke test, not Cilghal’s 12/12 battery.

What ds4max actually launches

Unless you override it, the wrapper now passes:

• full-resident native context 262,144 and --think (xhigh, not DeepSeek Think Max)
• --mtp --mtp-draft 3 --mtp-exact-sampling
• --vision on gguf/mmproj-Qwen3.8-Flash-Next-Q8_0.gguf
• --prefill-chunk 1024
• the steering file at FFN scale 1.0

Draft depth is 3, not 4. The Qwen mtp_R buffer is three hyper-connection rows. --mtp-draft 4 plus steering logs Metal directional steering received undersized buffers on the predictor steps. Generation still completed; the edit did not apply there. --no-ablate and --no-vision remain.

The agent has no /read. Point it at a PNG or JPEG and it calls view_image. Plain ds4 uses /read FILE. Image turns cannot /save.

Receipts

Official CLI vision regression, two synthetic PNGs (64×64 red, 96×48 blue), ctx 8192, greedy, --nothink:

• PASS ordinary — 2/2 encodes (64 and 72 image tokens), 4/4 prefills.
• PASS mtp — same encodes; decode 45–53 t/s versus ordinary 33–37 t/s.

Text smoke with --mtp-draft 4 --mtp-exact-sampling and the bank: print("hello") at 124.54 / 40.08 t/s, plus the undersized-buffer scar above. Ablation A/B (hide-evidence, revenge, CSV, SSRI) is in the companion runbook.

EETISMAD

E2E-tested. Vision CLI both modes PASS. Ablation A/B on four prompts after the engine gate was lifted. ds4max --check plans full-resident 262K when reclaimable RAM is ~90 GiB+.

In docs. This note, plus dir-steering/README.md on the Qwen profile, plus Claude_Code/runbooks/flash-next-m4max-2026-09-12.md.

Merged. Two local commits on qwen3.8-flash-next (8d29215, 69cb3ee). Upstream write access on ivanfioravanti/ds4 is READ, so they were pushed to Ogilthorp3/ds4-metal and opened as PR #13 against qwen3.8-flash-next. The refusal bank and contrast lists stay private. ~/.sanctum-mbp holds the launcher.

And Deployed. The stale ds4-agent (pid 67107, --mtp with draft default 1, no --vision) was stopped. The next ds4max is the live seat. There is no LaunchAgent; this is a terminal agent, not a Mini cathedral port.

Caution

Qwen SSD streaming is still unsupported. If ds4max --check is red because Chrome ate the unified pool, freeing RAM is the fix — --force will offer a stream plan that the engine will then refuse.

The spyglass is on the bench. Next time someone types ds4max, it is this seat — vision, draft 3, bank at 1.0 — or it is a flag to take one of those off. That is as deployed as a terminal agent gets.