{
"schema_version": "article-claims.v1",
"article_id": "what-consumer-hardware-actually-does",
"article_path": "site/articles/what-consumer-hardware-actually-does/index.html",
"verification_date": "2026-08-28",
"article_date_marker": "Updated 28 August 2026.",
"canonical_summary": "Twenty-nine indexed local-model labs plus the August rebuild and Qwen 3.8 campaigns",
"summary_paths": [
"site/articles/what-consumer-hardware-actually-does/index.html",
"site/articles/index.html",
"site/index.html",
"README.md"
],
"forbidden_open_gap_phrases": [
"Retracted, not resolved",
"Whether the two cards are actually equal",
"Step 5 is the fix, and it hasn't happened yet"
],
"claims": [
{
"id": "era3-rebuild",
"status": "confirmed",
"verified": "2026-08-28",
"source": {
"label": "Rebuilt-OMEN campaign fingerprint and retained burn-in record"
},
"assertions": [
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": "Completed 20 August:"
}
]
},
{
"id": "phase-specific-constraints",
"status": "corrected",
"verified": "2026-08-28",
"source": {
"label": "Rebuilt-OMEN load, concurrency and depth campaigns"
},
"assertions": [
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": "The binding constraint changes by phase"
}
]
},
{
"id": "concurrency-knee-eight",
"status": "confirmed",
"verified": "2026-08-28",
"source": {
"label": "Rebuilt-OMEN concurrency ladder and dense-model discriminator"
},
"assertions": [
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": "78 GB of commit headroom"
}
]
},
{
"id": "vulkan-mmv-dispatch",
"status": "confirmed",
"verified": "2026-08-28",
"source": {
"label": "vulkancliff public evidence package",
"url": "https://github.com/djcdevelopment/vulkancliff"
},
"assertions": [
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": "19.7 to 131.5 tok/s"
}
]
},
{
"id": "coopmat-restored",
"status": "resolved",
"verified": "2026-08-28",
"source": {
"label": "Rebuilt-OMEN single-card A/B and 6.5-hour soak"
},
"assertions": [
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": "to 1451/1474 tok/s"
}
]
},
{
"id": "flash-attention-crossover",
"status": "corrected",
"verified": "2026-08-28",
"source": {
"label": "Rebuilt-OMEN flash-attention context ladder"
},
"assertions": [
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": "16k–24k crossover"
}
]
},
{
"id": "q8-kv-new-stack",
"status": "superseded",
"verified": "2026-08-28",
"source": {
"label": "Rebuilt-OMEN fp16/q8 KV context ladder"
},
"assertions": [
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": "5–6% tax at 32k and 64k"
}
]
},
{
"id": "headless-symmetric-scaling",
"status": "resolved",
"verified": "2026-08-28",
"source": {
"label": "Rebuilt-OMEN headless two-server scaling campaign"
},
"assertions": [
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": "1.85× scaling"
}
]
},
{
"id": "model-rung-bakeoff",
"status": "current",
"verified": "2026-08-28",
"source": {
"label": "Rebuilt-OMEN three-model server bake-off"
},
"assertions": [
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": "95.2 tok/s"
}
]
},
{
"id": "combined-load-stability",
"status": "confirmed",
"verified": "2026-08-28",
"source": {
"label": "Rebuilt-OMEN combined-load soak and finale"
},
"assertions": [
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": "5.1% coefficient of variation"
}
]
},
{
"id": "operating-point-inversion",
"status": "current",
"verified": "2026-08-28",
"source": {
"label": "Qwen 3.8 promotion campaign: gated jobs-per-hour matrix at sixteen saturated clients"
},
"assertions": [
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": "5.49×"
}
]
},
{
"id": "per-card-thermal-ceiling",
"status": "confirmed",
"verified": "2026-08-28",
"source": {
"label": "Qwen 3.8 campaign watchdog abort records and per-adapter thermal telemetry"
},
"assertions": [
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": "96 °C"
}
]
},
{
"id": "flash-placement-ladder",
"status": "current",
"verified": "2026-08-28",
"source": {
"label": "Flash-Next placement study: eight-rung CPU/GPU layer ladder with per-rung validity"
},
"assertions": [
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": "27.70 tok/s"
}
]
},
{
"id": "mmap-host-placement",
"status": "confirmed",
"verified": "2026-08-28",
"source": {
"label": "Flash-Next load attempts with and without memory mapping and direct I/O"
},
"assertions": [
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": "0.76 GB of headroom left"
}
]
},
{
"id": "mtp-output-divergence",
"status": "current",
"verified": "2026-08-28",
"source": {
"label": "Speculative-decoding output comparison at temperature zero across matched cells"
},
"assertions": [
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": "zero identical responses"
}
]
},
{
"id": "single-stream-crossover",
"status": "current",
"verified": "2026-08-28",
"source": {
"label": "llama-bench single-stream sweep, four arms including a two-binary control"
},
"assertions": [
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": "181.6"
},
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": "0.40%"
}
]
},
{
"id": "placement-and-harness-cost",
"status": "current",
"verified": "2026-08-28",
"source": {
"label": "Paired single-card and layer-split arms, and bench-versus-server at matched placement"
},
"assertions": [
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": "about 7% of generation speed"
},
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": "19% on the fast model"
}
]
},
{
"id": "kv-park-restore-cost",
"status": "current",
"verified": "2026-08-28",
"source": {
"label": "Rotation-physics probes: KV slot save and restore across a full server restart, with byte-identical greedy output"
},
"assertions": [
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": "2.68 GB of KV state"
},
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": ">102.8 s"
}
]
},
{
"id": "direct-io-cold-load",
"status": "current",
"verified": "2026-08-28",
"source": {
"label": "Rotation-physics load ladder, seven trials per model, cold read-into-memory versus warm memory-mapped reload"
},
"assertions": [
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": "inert at the Windows file layer"
},
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": "8.2–8.3 s"
}
]
},
{
"id": "co-residency-quad",
"status": "current",
"verified": "2026-08-28",
"source": {
"label": "Four heterogeneous models resident two per card, solo and four-way concurrent canaries"
},
"assertions": [
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": "128 tok/s aggregate"
},
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": ">52.9"
}
]
},
{
"id": "hetero-pair-holds-solo",
"status": "current",
"verified": "2026-08-28",
"source": {
"label": "Heterogeneous one-model-per-card pair under concurrent fire, held against each model's solo rate"
},
"assertions": [
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": "99.2 and"
}
]
},
{
"id": "experts-on-host-point",
"status": "current",
"verified": "2026-08-28",
"source": {
"label": "Expert-weights-on-host placement of the 125B, solo and beside the live production model; deep-prompt prefill in this placement remains unmeasured"
},
"assertions": [
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": "10.6 tok/s at +6.3 GB"
},
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": "Shallow-prompt decode is what was measured"
}
]
},
{
"id": "sm-tensor-negative",
"status": "current",
"verified": "2026-08-28",
"source": {
"label": "Tensor-split canaries against layer-split baselines, plus peer-access probes reporting no direct path between the two cards"
},
"assertions": [
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": ">−72%"
},
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": "cannot reach each other"
}
]
},
{
"id": "depth-recall-fidelity",
"status": "current",
"verified": "2026-08-28",
"source": {
"label": "Verbatim-line recall probes on organic documents at two depths, both models at temperature zero in production shapes; one task family, eight probes per cell"
},
"assertions": [
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": "5 of 8 at 32K"
}
]
},
{
"id": "prefix-miss-306x",
"status": "current",
"verified": "2026-08-28",
"source": {
"label": "The campaign's accidental controlled experiment: shared-prefix versus unique-prompt time to first token at 26K depth"
},
"assertions": [
{
"path": "site/articles/what-consumer-hardware-actually-does/index.html",
"contains": "306×"
}
]
}
]
}