Study · not public
1 / 4
I · Position

Owned metal.
Metered judgment.
Power for the rest.

A private inference bench for principals who already think in unit cost. Volume stays in the room. Outside intelligence is reserved for decisions that earn it.

~14 TBNVMe across the workhorses
64 GBsystem memory on the bench
Largein-house models · heavy review
i7 · CPU PCIe · 16 GT/s ×4
Samsung 990 PRO Heatsink 2 TB
Root complex/ + weights
Operating system and model weights on one lane
i7 · archive · 16 GT/s ×4
Samsung 990 PRO 4 TB
PCH ×4paradise
Internal archive
Vault · ORA · 16 GT/s ×4
Corsair MP600 CORE XT 4 TB
Gen4 ×4library
Cold store · media spine
Vault · LABORA · 16 GT/s ×4
Predator GM7 2 TB
Gen4 ×4active work
Projects · model park

Assayed July 2026. Capacity is indicative and set per engagement. This bench takes a small number of houses, not a queue.

The Code Jeweller · Marbella
II · Path

The board is the product.

MEG Z690 UNIFY. A heatsink 990 PRO on the CPU’s own PCIe root — full 16 GT/s ×4 — and that drive is also the system volume. Not a hub. Not a laptop SSD asked to play server.

i7 topology · simplified
i7-13700 CPU root 990 PRO HS 2TB / + WEIGHTS · ×4 990 PRO 4TB archive RX 6700 XT 12GB Vault · Corsair 4TB · Predator 2TB · ×4 each

Large in-house models run here, including a heavyweight reviewer on i7. How they are packed is craft. That they run on owned metal is the point.

Z690MEG Unify · CPU M.2
2× 990PRO · 2 TB + 4 TB
×4every workhorse NVMe · 16 GT/s
27B+class · heavy review on i7

Link state measured 30 July 2026. Board MS-7D28.

II of IV
III · Proof

A third of a million
evaluations.
One owned pass.

The commercial question is not peak theatre. It is how long real bulk takes when the path is local and the bill is electricity.

326,531items classified
~22 hplanned
~10.6 hactual · 100%
$0vendor tokens
A metered frontier path prices this run in four figures and puts you behind someone else's queue. Run here: planned twenty-two hours, delivered ten and a half — capacity you already own, cost already sunk.
~1.0–1.6Btokens/month · sustained ceiling
~0.4Bbusiness-hours tier
~0.6Bovernight batch tier

Reproducible on request — dates, not adjectives.

III of IV · every number wears a date
IV · The chain

One queue.
Three weights.
Judgment last.

Bulk drafts fan across the mesh. A heavyweight reviewer sits on i7. The daily seat keeps a strong midweight for orchestration. Outside models only where the work earns them.

1 · Draft swarm
~8B class · vault (+ seat when invited)
short bulk~20–40 tok/s / node
Classify · extract · draft-ugly · triage
2 · Heavy review · i7
~27B class · compact weights · 12 GB GPU
Vulkan measured~28 tok/s
Verify · adjudicate · long context
3 · Code lane · i7
~16B class specialist
fits 12 GBmechanical code
Scaffolds · refactors · configs
4 · Seat · M1
~9B class midweight · 16 GB unified
orchestrate~20–40 tok/s class
Daily driver · joins swarm only when invited
~28tok/s · heavy ~27B on i7
~4.1item/s · 3-node bulk classify
~80tok/s · mesh gen sum (order)
~6concurrent draft streams · knee
Approximate envelope from dated in-house measurements (July 2026), not a cloud SLA. Heavier 70B-class only via multi-node split over copper — capacity win, not a speed win (~3–6 tok/s class). Frontier stays the signature on judgment.

Rates: heavy reviewer ~28 tok/s on i7 (Vulkan, dated 17 Jul). Draft tier ~21–40 tok/s per node. Real-prompt bulk classify ~4.09 items/s mesh-wide. Swarm concurrency knee ~6. Model names and packing method on enquiry.

IV of IV · Ora et Labora
Swipe · arrows · 1–4