nickh007/kv-tenant-isolation-bench

KV Tenant Isolation Bench — every cross-tenant observation, including the ones that refute us

13 measurements of cross-tenant KV-cache reuse on real serving stacks: the data-path leak, its closure, the timing oracle, and a four-run replication — four rows of which are marked uninterpretable because their positive control failed.

Why this dataset exists

Isolation results are published as headlines: "cross-tenant reuse: 208 → 0." What you cannot do with a headline is check whether the measurement could have detected a leak at all. This dataset ships the observations with the field that decides it.

interpretable is the column that matters. Four rows are false. On Phi-3-mini-4k the timing oracle did not reproduce — and not because the model is safe. It has no grouped-query attention, so it stores ~13.7× more KV per token; the victim's primed prefixes needed ~19.8 GiB against ~5.6 GiB available and were evicted before the attacker probed. Those rows measure two cache misses being compared to each other.

A dataset that quietly dropped them would make our result look cleaner than it is. They are here.

Schema

column	type	meaning
observation_id	string	unique id
measurement	string	data_path · timing_oracle · timing_oracle_replication
model	string	model under test
arm	string	cross_tenant · same_tenant · salted · unsalted
cached_tokens	int/null	tokens served from cache (data-path rows)
auc	float/null	attacker distinguisher advantage; 0.5 is chance
auc_ci_low, auc_ci_high	float/null	95% interval (Hanley–McNeil or bootstrap)
n_per_class	int	samples per class
positive_control_ok	bool/null	did the control establish the channel is detectable?
interpretable	bool	false ⇒ this row supports no conclusion in either direction
certificate	string	the committed cert this row was exported from

Usage

from loader import load

# The load-bearing filter. Never aggregate over uninterpretable rows.
rows = load("kv-tenant-isolation-bench", where={"interpretable": True})

leak = [r for r in rows if r["measurement"] == "data_path"]
# unsalted cross-tenant: 208 cached tokens → salted: 0 → same-tenant: 208 (preserved)

Stdlib only. pandas and datasets are optional conveniences, not requirements.

Provenance

Every row carries its source certificate:

certificate	what it measured
gpu/inengine_weave.json	the data path — 208 → 0 cross-tenant, 208 same-tenant preserved (n=1, opt-125m)
gpu/timing_oracle_apc.json	the single-run timing oracle, AUC 1.00 → 0.486
gpu/timing_oracle_replication.json	4 runs, 2 model families × 2 seeds, arm order randomised

Regenerate: python oss/datasets/export_datasets.py. Verify it still matches the certificates: --check.

Honest limits

• The data-path rows are n = 1 per condition, on opt-125m, a 125M toy model.
• The timing result is model-conditional. It reproduced on Qwen2.5-1.5B in 2/2 seeds and in both arm orders; it did not reproduce on Phi-3-mini-4k, where the positive control failed.
• "Zero cache cost" is retracted. Same-tenant reuse is preserved, but the cost of tenant-salting is the cross-tenant sharing it eliminates — and the measured workload contains none of that.
• Cross-tenant reuse of 0 is not a proof of isolation. It is the absence of one leak under one probe set.
• These are our own measurements. They support no claim about any commercial provider.

The commercial edition

This dataset is the diagnosis. Binding a partition key to an authenticated principal at the admission decision — the fix — is covered by filed claims and is licensed separately. See kvleak for the scanner, which is free and always will be.

Licence

Apache-2.0.

---

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
kv-tenant-isolation-bench	isolation observations, uninterpretable rows included ← you are here
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.