nickh007/kv-reuse-econ-traces

KV Reuse Econ Traces — a headline and the closed form that predicts it, side by side

65 per-workload first-touch prefill accounting rows: 36 from a synthetic size ramp, 29 from the real Mooncake FAST'25 trace.

Why this dataset exists

We published a 90.0% mean first-touch prefill cut. It is an exact arithmetic identity, not a measured efficiency — and rather than say so in a footnote, this dataset ships the closed form as a column beside the measurement:

first_touch_cut == closed_form_cut       # true on 36 of 36 synthetic rows
closed_form_cut  = 256 * (n // 256) / n  # chunk size × prompt length. That is all.

The value is fixed by two numbers we chose — the LMCache chunk size and the prompt-length ramp. On the eight largest blocks it becomes 0.947 on identical hardware. The pre-registered 0.30 floor was arithmetically unfalsifiable: any prompt ≥ 512 tokens clears it by construction.

The real finding in that run is zero_tenant_leak, true on every row. That is a correctness result. The 90% is accounting.

We ship the refutation as a column because a caveat in prose gets dropped when the number is quoted.

Schema

column	type	meaning
workload_id	string	synthetic-NNN or mooncake-NNN
source	string	synthetic_ramp · mooncake_fast25_trace
prompt_tokens	int/null	prompt length (synthetic rows)
apc_prefill_computed	int/null	tokens prefilled, no-pooling arm
pooled_prefill_computed	int/null	tokens prefilled, pooled arm
cached_tokens	int/null	tokens served from the pre-warmed store
first_touch_cut	float	measured cut, 1 − pooled/apc
closed_form_cut	float/null	256·⌊n/256⌋/n — equal to the measured cut on every synthetic row
zero_tenant_leak	bool	no cross-tenant reuse observed
certificate	string	the committed cert this row came from

The Mooncake rows carry closed_form_cut = null: real prompts have no chosen ramp, which is why their mean cut (0.668) is lower and 6 of 29 fall below the 0.30 floor. Those six are in the data.

Usage

from loader import load

rows = load("kv-reuse-econ-traces", where={"source": "synthetic_ramp"})
assert all(r["first_touch_cut"] == r["closed_form_cut"] for r in rows)   # 36/36

real = load("kv-reuse-econ-traces", where={"source": "mooncake_fast25_trace"})
below = [r for r in real if r["first_touch_cut"] < 0.30]                 # 6 of 29

Provenance

certificate	rows
gpu/fleet_econ_blocks.json	36 synthetic-ramp workloads (A100 + Qwen2.5-7B)
fleet_econ_realtrace_cert.json	29 distinct Mooncake FAST'25 conversation workloads

Third-party data: the 29 mooncake_fast25_trace rows are derived from the Mooncake trace — Apache-2.0, © kvcache-ai, https://github.com/kvcache-ai/Mooncake (USENIX FAST '25). They are per-workload accounting aggregates and carry no prompt or response text. See NOTICE. Attribution travels with the rows if you redistribute them.

Regenerate: python oss/datasets/export_datasets.py; verify with --check.

Honest limits

• No latency, no FLOPs, no throughput was measured. These are token-accounting quantities.
• The store-fetch cost is not netted out. Fetching KV from a remote store is not free and does not appear here.
• The synthetic headline is ramp-dependent, and the dataset lets you demonstrate that in one line.
• zero_tenant_leak is an observation over these workloads, not a proof of isolation.

The commercial edition

The pooling fabric, the cross-instance reuse path, and the fleet controller that these traces characterise are covered by filed claims and licensed separately. The traces, the benchmark, and the refutation of our own headline are free.

Licence

Apache-2.0. Mooncake-derived rows: Apache-2.0, © kvcache-ai — see NOTICE.

---

The rest of the portfolio

25 artifacts, one idea: a measurement you cannot check is a press release. Every tool here reports; none of them gates.

Tools

abstain-bench	how often does a verifier pass input it could not check?
evidence	run the whole portfolio over your repo — the weakest leg, never the mean
floorgen	what must your system remember? an exact lower bound
formal-proof-mcp	a proof kernel for your coding agent
gatecount	exactly how many states does removing this check admit?
gridlock	certify a wait-for relation cannot wedge
honestbench	measure your CI's escape rate
kvleak	cross-tenant leak scanner
kvprobe	model-substitution detector with a measured FPR
preregister	refuses to seal a plan whose conclusion is already fixed
proof-carrying-ci	the whole portfolio as one CI check, with SARIF
proof-to-code-drift	fail the build when the proof stops matching
sf-verify	re-derive admission decisions offline
signoff-cert	certificates that carry their own false-pass bound
tokencount	a token count both parties can recompute

Benchmarks — each recomputes one of our own published numbers from its certificate

illusion-bench	how many broken kernels does your oracle admit?
kv-reuse-econ-bench	recompute our economics headline
llm-tenant-isolation-bench	recompute our isolation figures

Datasets

abstain-corpus	32 inputs a verifier must NOT pass
kv-reuse-econ-traces	per-workload reuse accounting + the closed form ← you are here
kv-tenant-isolation-bench	isolation observations, uninterpretable rows included
llm-precision-fingerprints	precision-labelled logprobs with a negative control

Try it in a browser — no install, no GPU

negative-results-atlas	ten claims we took back
tenant-leak-demo	the residency calculator
wait-for-visualiser	paste a wait-for graph, see the cycle

Documentation

Everything above, explained in one place: https://nickharris808.github.io/evidence-docs/ — the tutorial, what this proves and what it does not, and a CLI reference generated by running --help on every published command.

The commercial edition

Everything above is measure-only and Apache-2.0: it tells you what is true and never acts on it. The enforcement side — binding a partition key at the admission decision, the compiled gate corpus, and the certificate-issuing faucet — is covered by filed patents and licensed separately.

Reading is free. Enforcing is licensed.