Skip to the content.

Benchmark sheets — directory index

Navigation only, not authority. Numbers are only authoritative in ../../BENCHMARK-AUTHORITY.md (the governed authority ledger), and every claim carries a tag in ../../CLAIMS.md. Charts live in ../../BENCHMARK-GALLERY.md. Each sheet below keeps its own provenance labels (MEASURED / MODELED / WITNESSED / PENDING / GATED); this page just tells you where to look.

Fleet, cache & session value

File Kind One-liner Headline
ABLATE-RESULTS.md results The deterministic, $0, no-model half of the self-ablation benchmark harness (epic #607). vDSO fast path serves 7 of 12 calls from cache, cutting 520 tokens (937 → 417) on the frozen tau2-airline-smoke trace
SERVED-INLINE-DEDUP-RESULTS.md results WITNESSED replay measuring how often the vDSO served-inline fast path (--vdso-proxy-fill) actually fires on a named agent read pattern, decomposed by win class (#1350). WITNESSED: 5/14 (35.7%) read-only proposals served inline on read-only-shaped names (all cross-turn re-reads); 0/14 on Claude-native Read/Grep names — the read-only NAME gate does not recognize them, so keep --vdso-proxy-fill opt-in
CDB-RESULTS.md results How the fak context debugger attaches to a finished session as a core image and faults in only the working set. A real ~350k-token-class session decomposes into an 18 KB page table over a 1.2 MB swap device; a follow-up demand-pages 1.8–6.2% of resident bytes (measured, witnessed by go tests + committed cdb-report.json)
FANOUT-BENCH-RESULTS.md results Prices one master goal decomposed into N sub-agents swept from 1 to 1024, measuring cross-agent dedup and shared-prefix KV reuse. MEASURED: 1,024 real agents complete one goal end-to-end in 364 ms on a laptop with no GPU (headline 72.8× parallel speedup is a MODELED projection)
FLEET-5X200-7B-10MIN-RESULTS.md results Measured on an Apple M3 Pro (Qwen2.5-7B Q8, llama.cpp Metal): the 5-agent × 200-turn fleet lands at ~8.2 min. MEASURED: batched + shared prefix fleet = ~8.2 min on M3 Pro — under the 10-minute bar (naive re-prefill ≥ 4.0 hours); host Metal forward, not the pure fak kernel
FLEET-SWEEP-RESULTS.md results Measures the turns-by-agents turn-tax surface across a 1-50 grid, quantifying the cross-agent cache uplift. At the T=50×A=50 corner the read-only fleet deletes 2344 of 2500 calls (94%), cross-agent uplift +370 — measured tier-2 kernel events, not modeled
FULLSPAN-TAILLOAD-RESULTS.md results One process, all four clock bands in a single causal trace with deny-consequence classes, plus the µs decide fold’s tail quiet vs under same-process load. OBSERVED (research-grade): same-process load leaves decide p50 nearly flat but degrades p99 ×2.1–2.5 (13.5–14 µs → 30–34 µs) and triples the >100 µs share; trace walks B6→B4→B2→B0 with every DENY carrying retry_turn/forked_outcome/clean_stop
GLM52-FAK-KERNEL-CACHE-VALUE-RESULTS.md pending PENDING — Results not yet collected; describes the result packet shape for the live GLM-5.2 cache-value run. Cache value (reused tokens): PENDING; offline WITNESSED-derived prefill-elimination floor A/C 17.9× → 23.4× (workers 1→16)
GLM52-FAK-KERNEL-CACHE-VALUE-RUNBOOK.md runbook End-to-end path to observe fak’s OWN in-kernel KV-prefix cache value on a real solved ticket via a GLM-5.2 gateway Status: the observation seam is SHIPPED and tested; the live GLM-5.2 number is the box residual — nothing here invents a tok/s or a reuse figure
GUARD-HOP-OVERHEAD-PENDING.md pending Harness + row structure landed; the live MEASURED wall-clock is PENDING (hardware-gated). PROJECTED per-turn 2.90–19.42 µs (8 calls/turn); per-session 0.14–0.97 ms (50 turns)
MEMORY-DREAM-CLEANUP-RESULTS.md results An offline pass that re-screens, repairs, and prunes a finished session’s core image against today’s gate.  
SELF-TAX-TREND.md trend Living trend companion to the self-tax row in BENCHMARK-AUTHORITY.md; a dated, net-true-labeled series. ~0.55 ns/op · 0 allocs · FLAT N=1→1000 drivers (WITNESSED, adjudication read-path floor)
SESSION-VALUE-STACK-RESULTS.md results Measures the fused agent kernel’s net work-reuse value-add (60.3x vs naive, 4.1x vs tuned) on a 50-turn session on M3 Pro. 60.3× net value-add vs naive (A/C), 4.1× vs tuned single-tenant (B/C) — MEASURED for the live arms; arm A computed-from-measurement (validated live)
ULTRA-LONG-CONTEXT-RESULTS.md results Reread-elimination win at the >100k-token regime, proven as an exact contention-free work floor ~10x vs naive on a single >100k session, ~40x+ on a 5-agent fleet each >100k (exact contention-free work floor, not a wall-clock)

Context safety & quarantine

File Kind One-liner Headline
KV-QUARANTINE-BRIDGE-RESULTS.md results Context-MMU quarantine verdict on a tool result’s bytes also evicts its K/V span from the attention cache. max|Δ| evict-vs-never = 0.000e+00 — the evicted cache is bit-identical to one that never saw the poison (go test witnessed; wiring witness uses a synthetic model)
LIVE-RESULTS.md results A live multi-turn A/B where a real model drives the same airline task twice, counting turns, tokens, and quarantines. The injection reached the baseline’s context in 100% of runs (5/5) and the kernel quarantined it in 100% of runs (5/5).
RECALL-RESULTS.md results Proves a tool result quarantined by the context-MMU cannot be paged into a new context across the process boundary. Core image: 4 pages (2 benign, 2 sealed), 442 bytes CAS — the quarantine survived the persist→reload boundary; 5/5 skeptic claims CONFIRMED
TOOL-RESULT-TREE-KV-RESULTS.md results Pins the precise claim that fak alone removes a tool result from a sequence middle bit-identically to never-saw. fak’s middle-span removal is bit-identical to never-saw: TestKVQuarantineEqualsNeverSaw matches HF token-for-token, max|Δ|=0 (sibling-isolation 0.000e+00; fixture-gated HF rung; not yet wired into the live fak agent loop)
TURN-TAX-RESULTS.md results Measures fak’s deterministic safety floor on a live kernel trace plus the extra error-recovery turns its one-shot syscall path eliminates when self-hosted. Safety floor (the moat): 1 injection quarantined, 1 destructive op denied (baseline 1/1 vs fak 0/0) on the 14-call turntax-airline slice — real kernel counters, engine-agnostic; efficiency upside −9 turns is a cache-favorable self-host-only slice (real-world ~0.7% rate → 0.33 turns/session)

In-kernel model correctness & parity

File Kind One-liner Headline
FAK-NATIVE-QWEN35-RESULTS.md results fak’s own in-kernel forward pass runs the Qwen3.5/3.6 hybrid Gated-DeltaNet end-to-end, 0.8B to 27B GGUF on M3. Clean headline: decode 1.2 tok/s (64 tok / 54.12 s), prefill 0.6 tok/s (29 tok / 48.27 s) vs the llama.cpp-Metal bar 7.29 / 51.55 and the 3× goal 2.7 (MEASURED on Apple M3 Pro)
IN-KERNEL-MODEL-RESULTS.md results A kernel-owned in-process CPU forward pass for SmolLM2-135M, proven rung-by-rung bit-for-bit against HuggingFace. Every rung proven against HuggingFace transformers: argmax exact 3/3 prompts, greedy generation token-for-token identical, KV quarantine and prefix reuse max|Δ| = 0.000e+00
M3-LLAMACPP-RESULTS.md results Adds an arm64 NEON Q8 kernel that flips int8 to 1.9x faster than f32, runs Qwen2.5-1.5B, and measures the gap to llama.cpp fak now lands in the same order of magnitude as llama.cpp’s CPU path (decode ~2.2x behind, prefill ~6.5x behind) — not at parity
MODEL-BASELINE-RESULTS.md results Measures the in-kernel forward pass against HF and llama.cpp on CPU, then closes it to decode parity. Prefill per-token slope (raw compute): llama.cpp Q8_0 0.337 ms/tok vs fak Q8 0.346 ms/tok = 1.03x — parity (measured 2026-06-17 under live 6-session fleet load)
MODEL-BATCHING-RESULTS.md results Measures in-kernel multi-user batched decode over per-user KV caches, scaling throughput while bit-identical to serial Step. Aggregate decode throughput: 2916 tok/s at B=960 = 151.9× the unbatched f32-serial baseline and 41.3× the real single-stream Q8 decode (measured, native, under live fleet load)
QWEN25-7B-RESULTS.md results fak’s first 7B dense run on M3 Pro Q8, reporting the honest throughput gap vs llama.cpp Metal plus full greedy parity. 8.7 tok/s decode (115.2 ms/tok) — fak is 0.083× prefill / 0.50× decode vs llama.cpp Metal
QWEN35-0.8B-RESULTS.md results End-to-end in-chat f32 hybrid-GDN path for Qwen3.5-0.8B, proving the architecture works on a tiny model before scaling up. One Qwen3.5-0.8B f32 modelbench number, and it is in the Authority (#113): pinned cold load time 1203ms (M3 Pro)
QWEN36-LOAD-PROFILE-440.md results Phase-attributed load profile of the Qwen3.6-27B GGUF->Q8 path, with a same-host arena-reuse before/after. Witnesses the #440 page-churn fix at -4.0% peak RSS / -4.0% page faults (measured 2026-06-26)
QWEN36-PARITY-RESULTS.md results The witnessed llama.cpp Metal reference for Qwen3.6-27B on M3 Pro, the speed bar fak’s own engine targets. Metal, full offload: prefill 51.55 tok/s, decode 7.29 tok/s (witnessed; fak decode progression 0.1 → 0.9 → 1.2 tok/s, not yet speed-parity)
QWEN36-PARITY-ROLLUP-2026-06-28.md rollup One authoritative reconciliation of the Qwen3.6-27B-vs-llama.cpp parity status: PROVEN vs not yet, with gated repros Correctness parity PROVEN at architecture level but REFUTED at 27B scale (token-3 argmax flip); speed parity not yet — 1.2 tok/s vs the 7.29 tok/s llama.cpp-Metal bar (recorded prior Mac witnesses, none re-measured here)
ROPE-SCALING-RESULTS.md results NTK-by-parts RoPE rescale letting an 8K-trained model attend at 128K, HF-faithful and byte-identical when unset. llama3 inv_freq rescale — 8K→128K, proven against HF’s own reference; device lanes NOT yet wired (known gap)
SLIDING-WINDOW-RESULTS.md results Reports fak’s per-layer sliding-window attention bounding cost from O(N) to O(W), shipped as a read-time mask. Measured: 1,000,000 tokens decoded with the KV cache never exceeding 256 positions (window 128), vs the ~67 GB a full 1M-token cache would need for SmolLM2-135M

GPU & hardware platform results

File Kind One-liner Headline
GCP-H100-RESULTS.md results First end-to-end run of fak’s own CUDA engine on a live GCP H100, head-to-head vs llama.cpp CUDA, with an honest gap verdict MEASURED: fak-cuda decodes Qwen2.5-3B at 96.3 tok/s (f32) vs llama.cpp’s 361.6 tok/s (Q8_0) on an H100 — ~3.75x behind on decode, verified on-device (-require-non-reference)
GCP-L4-RESULTS.md results The completed GCP L4 head-to-head for Qwen2.5-3B Q8_0: llama.cpp CUDA, fak-cpu, and fak-cuda on one NVIDIA L4. fak-cuda measured on a real L4: backend selected “cuda”, tier sm_89, 16.6 prefill / 18.7 decode tok/s (f32) vs llama.cpp CUDA Q8_0 at 6,638.3 / 70.8 — the gap is not a win claim
GPU-QWEN-RESULTS.md results fak-CUDA decode parity with llama.cpp f16 on Qwen2.5-1.5B and a Q8_0 path reaching Qwen2.5-3B on an 8 GB RTX 4070, argmax-exact. Equal-precision decode parity, Qwen2.5-1.5B-Instruct: fak-CUDA f16 ~36.6 tok/s vs llama.cpp F16 34.3 (fak ~1.07x), witnessed greedy argmax-exact vs the f32 reference
H100-KERNEL-5X-ROADMAP.md plan Evidence-backed decomposition of the measured fak-on-H100 gap into ranked, code-anchored levers with expected multipliers. Every speedup number below is a projection gated on a measured Hopper run — decode 3.75× behind, prefill ~380× behind (MEASURED baseline; levers GPU-gated “not yet”)
QWEN36-AMD-VULKAN-RESULTS.md results Witnessed run of Qwen3.6-27B Q4_K_M on an AMD/Vulkan Windows desktop, proving the model serves all three fak surfaces. Result: 3/3 surfaces passed (witnessed 2026-06-19 on the AMD/Vulkan desktop node)
VULKAN-AMD-RESULTS.md results Witnesses fak’s Vulkan compute backend reaching numerical parity on a real AMD RX 7600 while still ~60x slower than llama.cpp. Numerical parity on real AMD silicon (argmax-exact greedy decode, prefill-logit cosine = 1.0); throughput NOT at parity — best 2.5 tok/s, ~58-60x slower than llama.cpp CPU (all MEASURED, nothing SIMULATED)

Engine & serving head-to-heads

File Kind One-liner Headline
GPU-SERVER-GLM52-VLLM-AGENTIC-BENCHMARKS.md pending A pending-measurement plan for GLM-5.2 on vLLM through fak’s gateway: agentic benchmarks + 20-task SWE-bench slice. Status: pending measurement. This document carries commands and gates, not results.
VLLM-EP-EPLB-MOE-BASELINE-RUNBOOK.md runbook The reproducible vLLM Expert-Parallel + EPLB MoE serve command and artifact schema — the SOTA serving floor native fak EP-sharded GGUF work must cite (#1733). Status: pending measurement. Carries the EP/EPLB serve command, TTFT/TPOT/ITL + before/after expert-balance capture, and the fak.vllm-ep-eplb-moe-baseline.v1 schema — no number until a real EP node runs it.
LLAMACPP-HEADTOHEAD-RESULTS.md results Measures fak against llama.cpp on CPU across decode, prefill, batched throughput, and shared-prefix, with honest verdicts. single-stream decode is parity (1.12×), batched is parity/slight fak lead (2916 vs ~2816 tok/s); cross-agent shared-prefix is OPEN (preliminary, settings-dependent)
QWEN36-27B-GPU-SERVER-RESULTS.md results fak serving and coding-agent surfaces on Qwen3.6-27B with an 8-GPU SGLang backend, reporting gateway tax across a 1-to-128 concurrency sweep. All three fak surfaces PASS on the 27B; peak 64-concurrency 1085.6 tok/s fak-gateway vs 1451.6 raw SGLang (0.75×), converging to ~3% tax at conc 128 (0.97×)
RADIXATTENTION-RESULTS.md results Head-to-head of fak’s KV-cache reuse against SGLang’s RadixAttention on cache hit rate, with a 4.58x live speedup. 4.58× live wall-clock speedup on SmolLM2-135M Q8 (agents workload); WITNESSED kernel-owned 77–88% cache hit rate inside SGLang’s verified 50–99% band (MEASURED)
VLLM-HEADTOHEAD-RESULTS.md pending A gateway adjudication-tax witness and vLLM engine bench; pending-measurement — every vLLM cell is a placeholder. Status: pending-measurement (GATED scaffold). No vLLM GPU run has landed yet — every numeric cell in the vLLM tables is a placeholder (TBD); the only real vLLM-comparison numbers are the measured SGLang sibling’s (0.75× at peak, ~3% tax at saturation).

External benchmark suites & contracts

SWE-bench, LiveCodeBench, Terminal-Bench, FrontierSWE, and the local coding witness — plus the contract map that governs which of these may ever carry a public result claim.

File Kind One-liner Headline
BENCHMARK-CONTRACT-MAP.md contract The single map binding each mediated fak eval to its official public benchmark, oracle, artifacts, and caveats. Every row with result_claim_allowed=false is a contract or local fixture, not a public-leaderboard result; local smokes quotable only as [SIMULATED]
EVE-EVAL-PARITY-RUNBOOK.md runbook The command path and evidence contract for running a Vercel Eve eval suite raw and fak-routed, before any parity claim (#2605). Status: contract written, no fixture harness yet — no eve parity claim is made
FRONTIERSWE-ENV-ADAPTER.md adapter How fak stands up a co-resident fak serve gateway inside a FrontierSWE task sandbox without claiming a benchmark result.  
FRONTIERSWE-RESULTS.md pending The authority page for fak’s FrontierSWE time-to-solution (TTS) claim. Created empty and gated: no number yet. GATED, no number yet — no wall-clock or turn-count TTS number is recorded until the official grader has produced both arms’ reward.json and score-parity holds
FRONTIERSWE-SCORING-PARITY.md contract How internal/frontierswe’s Go scorer maps field-for-field onto the published leaderboard oracle score_from_reward.py  
FRONTIERSWE-TTS-RUNBOOK.md runbook The exact end-to-end recipe for a raw-vs-fak FrontierSWE time-to-solution (TTS) comparison. No TTS number is claimed until the official grader has run and score-parity holds (GATED)
LIVECODEBENCH-EPIC.md plan Repo-side index mirroring the GitHub epic anchor for first-class LiveCodeBench support (#2085). this index does not carry a score; LiveCodeBench results stay pending run
LIVECODEBENCH-RESULTS.md pending Status: pending run. Authority placeholder recording the fields a real LiveCodeBench result must carry; no pass rate. pass@1: pending run
LIVECODEBENCH-RUNBOOK.md runbook The command path and evidence contract for running LiveCodeBench through fak without turning a dry run into a score claim. Status: runbook assembled; native fak adapter pending child issues — all score cells remain pending run
LIVECODEBENCH-SUBMISSION-PACKET.md packet Assembly index for a future LiveCodeBench leaderboard submission; makes no LiveCodeBench result claim. Status: BLOCKED_PRECREDENTIAL - no result claim, no authority row yet.
LOCAL-MODEL-CODING-WITNESS-2026-06-27.md results Measured results of a small local model behind fak guard on a minimal coding task, with an honest local-vs-frontier A/B. MEASURED — the capability axis was run live: both 3B and 7B Qwen2.5-Coder rungs produced the correct fix (1-fail → 2-pass); governance witnessed via replay-trace (8 adjudicated verdicts, 3 dangerous calls denied)
LOCAL-MODEL-CODING-WITNESS-RUNBOOK.md runbook End-to-end command sequence to measure a small local model behind fak guard on a minimal coding task, with A/B. Status: the path is ASSEMBLED; the measured numbers are PENDING a real run.
SWEBENCH-PURE-KERNEL-RUNBOOK.md runbook Exact end-to-end command sequence to resolve SWE-bench Verified with fak’s native CUDA engine (no SGLang in the path) Resolve-rate on the pure kernel: pending GPU run — the path is ASSEMBLED in code and the harness is witnessed
SWEBENCH-RESULTS.md results A fak-native SWE-bench Verified benchmark directly comparable to the bench tool on cost, cache-reuse, turns, and adjudication THEORETICAL (MODELED) value-stack floor A/C 17.9×→23.4× (workers 1→16); MEASURED in-process adjudication p50 ~2.4 µs vs ~5.8 ms spawn (~2 400×); resolve-rate not MEASURED/VERIFIED (GPU server-gated)
SWEBENCH-VERIFIED-GPU-SERVER-RESOLVE-COMPARE.md results A real coding-agent SWE-bench Verified run driving Qwen3.6-27B through the fak gateway and raw SGLang, official harness. The same model resolves the instance through raw SGLang (1/1) but is blocked by fak’s capability/trust floor (0/1) — overall completion is decided by the floor, not the model.
TERMINAL-BENCH-2.1-FAILURE-TAXONOMY.md taxonomy The general agent behavior that classifies a failed Terminal-Bench task and decides a legal recovery. Status: engine shipped, pending live-run wiring — no Terminal-Bench number claim; result_claim_allowed stays false
TERMINAL-BENCH-2.1-SUBMISSION-PACKET.md packet Assembly index tying every checked-in campaign artifact to the promotion gate; makes no Terminal-Bench result claim. Status: BLOCKED_PRECREDENTIAL — no result claim, no authority row yet; result_claim_allowed=false across every artifact