fold in external review (N-5): decode-rate corrected to unknown/plan-at-5-tok/s, cache semantics fixed (hits not re-prefill), KV headroom, feral-cats 7-slot gotcha + engineering debt, five-strategy table, benchmark gate replaces the N-4 lean, TCO

Review source: GPT-5.6 pass shared by jhestyr 2026-07-17. Accept/pushback record in NOTES N-5.

[ezra]

BUILDS.md

@@ -1,8 +1,12 @@
 # local-horse — the two priced builds (as of 2026-07-17)
 
 ⚠️ **Prices dated 2026-07-17, ±15% on used parts and worse on RAM/GPU (see NOTES N-3). Re-price before buying.**
+⚠️ **Reviewed 2026-07-17 (external pass, NOTES N-5): decode-rate and fit numbers corrected below; any
+purchase is now gated on the benchmark plan at the bottom of this file.**
 Budget target: ~$20k. Constraint: no Mac. Both shapes serve an OpenAI-compatible endpoint on the tailnet
-(llama.cpp / vLLM, headless Debian, Dockge like everything else).
+(headless Debian, Dockge like everything else). The inference stack is **part of the spec, not a detail**
+(N-5 #4): a reproducible build names *model file + quant + engine + version + expert placement + context
+length + prefix-cache config* — "llama.cpp / vLLM" hand-waving is not a design.
 
 ---
 
@@ -24,12 +28,17 @@ What it runs (MoE trick: hot path/attention in VRAM, cold experts in system RAM)
 
 | Model | Fit | Speed (est.) | Capability, plainly |
 |---|---|---|---|
-| GPT-OSS-120B (5B active) | Entirely in VRAM @ 4-bit | 100+ tok/s | Reliable daily hand — summaries, scripts, small fixes. ~Sonnet-3.5 tier. |
-| **Qwen3-Coder 480B-A35B** @ 4-bit (~270GB) | Hybrid VRAM+RAM | ~15–25 tok/s | **The prize.** Real agentic coding on a known repo. Sonnet-4-era. |
-| GLM 5.2 quantized | Hybrid (footprint unverified) | similar | Current open SWE-Bench champ; same tier or a notch up. |
-| Kimi K2.6 (1T) @ Q3 | Barely (608GB total, tight) | single digits | Party trick, not a daily driver. |
+| GPT-OSS-120B (5B active) | Entirely in VRAM @ 4-bit | 100+ tok/s | Reliable daily hand — summaries, scripts, small fixes. Tier claims need re-testing on the exact quant (N-5 #5). |
+| **Qwen3-Coder 480B-A35B** @ 4-bit (~270GB) | Hybrid VRAM+RAM | **UNKNOWN — plan at ~5 tok/s.** Naive PCIe offload ≈2–3; CPU-executed experts could hit low teens but is unproven on Milan/AVX2 (N-5 #1) | The prize *if* the hybrid path works. Benchmark before believing. |
+| GLM 5.2 quantized | Hybrid (footprint unverified) | same unknown as above | Current open SWE-Bench champ (full precision — quant untested). |
+| Kimi K2.6 (1T) @ Q3 | Barely (608GB total, tight) | low single digits | Party trick, not a daily driver. |
 | K3 (2.8T) | No | — | Cloud only. Its **distills** should land in this envelope. |
 
+**KV-cache headroom (N-5 #3):** the 96GB card is NOT all weight capacity. Qwen3-480B (62 layers,
+8 KV heads, 128-dim) needs ≈24GiB of KV at 100K context in BF16 (≈12 FP8), plus CUDA workspaces and
+buffers — realistic weight budget on-card is **~60–75GB**, which pushes *more* experts to RAM and
+worsens the offload math above. Size every fit calculation at your target context, not at zero.
+
 ## Build 2 — The Feral Cats (8× used RTX 3090, 192GB VRAM)
 
 Everything model-resident → prefill screams. Also screams literally; needs a **240V circuit** (~3kW load),
@@ -39,13 +48,20 @@ Everything model-resident → prefill screams. Also screams literally; needs a *
 |---|---|---|
 | GPUs | 8× used RTX 3090 24GB (~$850 ea; GDDR6X dodged the DRAM spike) | ~$6,800 |
 | CPU | EPYC 7402 used (Rome — just needs the PCIe lanes) | ~$300 |
-| Board | ASRock Rack ROMED8-2T (7× PCIe x16) | ~$700 |
+| Board | ASRock Rack ROMED8-2T — **7× PCIe x16: the 8th GPU needs a bifurcated slot** (N-5 #4) | ~$700 |
 | RAM | 256GB DDR4 refurb | ~$1,000 |
 | Risers/frame | Bifurcation risers + open frame or 4U conversion | ~$500 |
 | Power | 2× 1600W PSU + sync board, 240V | ~$650 |
 | Storage/misc | NVMe, fans, zip ties, regret | ~$550 |
 | **Total** | | **~$10.5k–12k** |
 
+**Engineering debt this table hides (N-5 #6) — required before this is a plan, not a sketch:** lane
+map (which slots run x16 vs bifurcated), Above-4G decoding / MMIO validation that 8 large-BAR devices
+actually boot, riser signal integrity at PCIe 4.0, 16–24 PCIe power connectors with transient headroom,
+an airflow design, a power-cap strategy, and a spare-GPU replacement procedure. Also a correction: 3090s
+DO support **pairwise** NVLink (no 8-way fabric) — four bridged pairs may help some topologies if the
+framework understands the nonuniform layout.
+
 | Model | Fit | Speed (est.) | Capability, plainly |
 |---|---|---|---|
 | **Qwen3-235B-A22B** @ 4-bit (~130GB) | Fully in VRAM — the rig's sweet spot | fast decode AND prefill | Strong all-rounder one tier below Coder-480B. Best agentic feel per dollar here. |
@@ -56,18 +72,52 @@ Everything model-resident → prefill screams. Also screams literally; needs a *
 
 ## The trade in one sentence
 
-Workhorse reaches the **best** open coding models slowly-but-surely and behaves like an appliance; the
-cats run one tier down blisteringly fast for half the money, paid for in a 240V run, fan roar, and jank.
-For agentic loops (99% context re-reads — NOTES N-1) all-VRAM prefill genuinely matters, but model
-ceiling matters more: a smarter slow model wastes fewer turns than a fast one going in circles.
+Workhorse *might* reach the best open coding models (unproven — N-5 #1) and behaves like an appliance;
+the cats run one tier down fast for half the money, paid for in a 240V run, fan roar, jank, and the
+engineering debt above. For agentic loops (cache-dominated — NOTES N-1) what matters most is **keeping
+the KV cache resident and un-evicted across tool calls**; all-VRAM helps there, but model ceiling still
+matters: a smarter slow model wastes fewer turns than a fast one going in circles.
 
 ## Why one big card and not 8 (the physics, kept for re-derivation)
 
 1. **Power/noise:** 8×350W = 2.8kW load — over a 15A/120V circuit's ceiling; plus ~150W idle, forever.
-2. **Interconnect tax:** split models cross PCIe (~32GB/s) every token vs 1.8TB/s on-card; consumer cards
-   have no P2P/NVLink (disabled on 4090+). Multi-GPU wins at batch *throughput* (many users), not
-   single-user latency — and the counter has one user.
-3. **MoE changed the buy:** ~35B active params need one fast GPU + cheap capacity, not 8 GPUs of compute.
+2. **Interconnect tax:** split models cross PCIe (~32GB/s) every token vs 1.8TB/s on-card; 3090s have
+   only pairwise NVLink (no 8-way fabric; P2P disabled on 4090+). Multi-GPU wins at batch *throughput*
+   (many users), not single-user latency — and the counter has one user.
+3. **MoE changed the buy:** ~35B active params need one fast GPU + cheap capacity, not 8 GPUs of compute —
+   *if* the expert-offload path performs (N-5 #1: currently unproven on this CPU generation).
+
+## Five strategies, not two (N-5 #8)
+
+| Strategy | Role | Est. cost |
+|---|---|---|
+| One DGX Spark (128GB coherent, 273GB/s, ARM64) | Low-ops control case; slow but simple | ~$4k |
+| Two Sparks (256GB aggregate) | Distributed-inference experiment | ~$8k |
+| **One RTX Pro 6000, strong 80–120B model** | Fast, simple, proven parts — **the defensible baseline** | ~$11–15k |
+| Workhorse hybrid (this doc's Build 1) | Experimental 480B CPU/GPU execution — benchmark-gated | ~$15–19k |
+| **Local 80–120B + cloud escalation for hard turns** | Likely economic optimum; pairs with the baseline row | baseline + subscription |
+
+The workhorse is honestly a *poor man's DGX Station* (748GB coherent memory, outside budget) without
+the coherent interconnect — whether the workaround performs is the entire open question.
+
+## Benchmark gate — do this BEFORE buying anything (N-5, replaces the N-4 lean)
+
+1. Rent or borrow equivalent hardware (cloud RTX Pro 6000 instances exist; a Spark is ~$4k as a control).
+2. Run the **exact quantized artifacts** — Qwen3-Coder 80B-A3B first (≈96% of 480B quality; if it holds
+   on our repos, one GPU suffices and the hybrid question is moot), then 120B, then 480B hybrid.
+3. Measure: uncached prefill, cached TTFT, decode rate, **cache-hit rate across real tool-call loops**,
+   max stable context, wall power, and **successful agent turns per hour** on our actual repos — not
+   SWE-bench, not tok/s alone. Include: tool-call validity, patch correctness, instruction retention at
+   long context, recovery after failed commands, hallucinated-completion rate.
+4. Treat **5 tok/s as the conservative workhorse case** until a receipt exists.
+5. Only then pick a row from the strategy table.
+
+## TCO beyond the parts bill (N-5 #9)
+
+Subscription makes current marginal Claude cost **$0** — this box is bought for privacy, control, and
+unlimited background tokens, not savings. At ~$0.15/kWh: 700W continuous ≈ $920/yr; 3kW continuous ≈
+$3,940/yr; realistic duty cycle (idle 100–150W, bursty load) lands a few hundred $/yr for the workhorse.
+Add: cooling, a 240V circuit if cats, dead used silicon, and admin time.
 
 ## Sources (as read 2026-07-17)
 
@@ -76,3 +126,6 @@ ceiling matters more: a smarter slow model wastes fewer turns than a fast one go
 - Hardware/pricing: videocardz.com (RTX Pro 6000 list $13,250) · thundercompute.com (build pricing) ·
   corewavelabs.com + pcserverandparts.com (DRAM crisis) · tomshardware.com RAM price index · memory.net
 - K3: venturebeat.com · cnbc.com · marktechpost.com · simonwillison.net · platform.kimi.ai quickstart
+- External review (2026-07-17): GPT-5.6-thinking pass via jhestyr's friend — full accept/pushback record
+  in NOTES N-5; its cites included vLLM offload docs, KTransformers AVX2 issues, Qwen3-Coder config,
+  NVIDIA DGX Spark/Station specs, and mykolaaleksandrov.dev on Claude Code breaking llama.cpp prefix cache

NOTES.md

@@ -16,8 +16,11 @@ cached-in · $3/M in · $15/M out). Repricing the adventure game's real build to
 
 - **1,147M cache-read** + 18M uncached-in + 3.5M out → **≈ $451 at K3 rates** (±30%, session-level
   granularity) vs ≈ $1,307 at Claude list prices. Actual marginal cost: $0 (subscription).
-- The load-bearing fact: agentic coding is **~99% context re-reads, ~1% output**. Cached-input price
-  (cloud) / prefill speed + KV-cache persistence (local) dominate everything.
+- The load-bearing fact: agentic coding is **~99% cache reads, ~1% output** — cache-*dominated*.
+  (Corrected per N-5: cache reads are **hits**, not reprocessing. Local first-order requirements =
+  KV-cache persistence + hit rate + capacity; raw prefill bandwidth matters on the misses only.
+  Operational gotcha: harnesses that mutate the prompt *prefix* silently break llama.cpp/vLLM prefix
+  caching — a slower box that keeps a 100K-token cache warm beats a faster box that keeps evicting it.)
 - K3 itself is never local: ~1.4TB weights at 4-bit. "Local like that" = best open model that fits a box.
 
 ## N-2 (2026-07-17) — The open-model landscape, mid-2026
@@ -46,7 +49,7 @@ mid-2026; SK Hynix sold out through 2026, fabs pivoted to HBM). Consequences:
 - RTX Pro 6000 Blackwell 96GB: Nvidia list jumped 55% in 16 months to $13,250; street/Max-Q hunts lower.
 - Relief not expected before **2027**. Waiting is a legitimate strategy; re-price at buy time.
 
-## N-4 (2026-07-17) — Decision lean (not a decision)
+## N-4 (2026-07-17) — Decision lean (not a decision) — ⚠️ SUPERSEDED BY N-5
 
 Ezra's recommendation if/when the itch turns real: **the workhorse** (Milan + 512GB DDR4 refurb + one
 RTX Pro 6000 Max-Q, ~$15–19k) over the 8×3090 rig — a smarter model that's slower wastes fewer turns
@@ -55,3 +58,38 @@ second card slot) without replacing bones. Integration shape: headless Debian, l
 OpenAI-compatible endpoint on the tailnet; per [[mill-is-home-for-workloads]] it's a Mill annex, not a
 new species. Sensible trigger to revisit: **after 2026-07-27** (K3 weights + first distills) or when
 DDR5 unclenches.
+
+**Superseded 2026-07-17 (same day) by N-5:** the workhorse's decode-rate assumption didn't survive
+external review. Lean is now benchmark-gated; see N-5.
+
+## N-5 (2026-07-17) — External review (GPT-5.6 via jhestyr's friend) — what survived, what didn't
+
+jhestyr shared BUILDS.md out (repo made public for it) and brought back a GPT-5.6-thinking review.
+Graded on merits; most of it holds. **Accepted corrections** (BUILDS.md updated to match):
+
+1. **Workhorse decode rate was wrong.** My "15–25 tok/s" for Qwen3-Coder-480B silently assumed
+   CPU-executed experts. Naive vLLM/llama.cpp offload ships ~11GB of active expert weights over PCIe
+   (~25GB/s) *per token* → **2–3 tok/s ceiling**. KTransformers-style CPU-expert execution could reach
+   low teens against 205GB/s DDR4 — but its fast paths want AVX-512/AMX, and **Milan is AVX2** (native
+   AVX2 support still maturing as of Feb 2026). Honest number: **unknown; plan at ~5 tok/s** until
+   benchmarked on this exact combo. The workhorse is a hypothesis, not a validated build.
+2. **Cache semantics were backward** (fixed in N-1): cache reads = hits, not re-prefill. Persistence/
+   hit-rate/capacity are first-order; prefill bandwidth is the miss path. Corollary gotcha: Claude
+   Code-style harnesses can mutate the prompt prefix and silently kill prefix-cache reuse.
+3. **KV-cache headroom ignored:** Qwen3-480B at 100K context ≈ 24GiB BF16 (12 FP8) of KV — the 96GB
+   card really offers ~60–75GB for weights, pushing MORE experts to RAM. Compounds #1.
+4. **Feral cats under-engineered:** ROMED8-2T has **7** x16 slots — my table put 8 GPUs on it with no
+   bifurcation plan, lane map, Above-4G/MMIO validation, or power engineering. Also: 3090s DO have
+   pairwise NVLink (no 8-way fabric) — my "no NVLink" line was too flat.
+5. **Capability tiers not decision-grade:** "Sonnet-4-era" conflates full-precision benchmarks with an
+   aggressive quant on an exotic execution path. The metric that matters: **successful agent turns per
+   hour** on OUR repos. And test the **80B-A3B first** (≈96% quality) — if it holds, one GPU suffices.
+6. **Economics:** subscription makes marginal Claude cost $0 — the box is justified by privacy/control/
+   unlimited background tokens, not savings. Add TCO: ~$0.15/kWh → 700W continuous ≈ $920/yr.
+   (My pushback: our duty cycle isn't continuous; idle ~100–150W. Real number is a few hundred $/yr.)
+
+**Pushback recorded (minor):** the 2–3 tok/s "upper bound" assumes uniformly distributed expert
+selection with zero locality; hot-expert pinning skews that upward in practice — but direction stands
+and the burden of proof is ours. **New lean (replaces N-4):** *benchmark before buying* — the friend's
+strongest frame: **one RTX Pro 6000 running a strong 80–120B local model, with cloud escalation for
+hard turns, is more defensible than either full build.** Benchmark gate lives in BUILDS.md.

README.md

@@ -15,9 +15,15 @@ re-derive it: two build shapes, the models each runs, honest capability tiers, a
 that will invalidate the numbers over time.
 
 The trigger was real data: repricing the lassiter-creek-adventure-game's build (10 sessions, Jul 5–17,
-from local transcripts) showed **1.15B cache-read tokens vs 3.5M output** — agentic coding is ~99%
-re-reading context. That shape drives every hardware conclusion here (prefill/memory-bandwidth is the
-bottleneck, not generation speed).
+from local transcripts) showed **1.15B cache-read tokens vs 3.5M output** — agentic coding is
+cache-dominated. The hardware conclusion that follows (corrected by external review, NOTES N-5): the
+first-order requirements are **KV-cache persistence, hit rate, and capacity** — cache reads are *hits*,
+not reprocessing; raw prefill bandwidth matters on the misses, not on the 99%.
+
+**Status 2026-07-17 (late):** externally reviewed (jhestyr's friend via GPT-5.6 — NOTES N-5). Two
+original numbers did not survive: the workhorse decode-rate estimate and the feral-cats slot count.
+**Any purchase is now gated on the benchmark plan in BUILDS.md** — the workhorse is a hypothesis, not
+a validated build.
 
 Standing constraints from jhestyr: **no Mac** ("not interested in the mac approach"). Not that fancy a
 man. Any coffee is coffee as long as it's not burnt.