nickh007/abstain-corpus

Abstain Corpus — inputs a verifier must not pass

32 inputs for testing whether a verifier claims success on something it could not possibly have checked. 24 are scored; 8 are controls that make the score mean anything.

Every verifier — a linter, an SBOM scanner, an attestation checker, a policy engine, a deadlock detector — has the same tempting shape: return success unless you find a problem. Hand one an empty file, a truncated log, or an artifact with no trust anchor and it finds no problem, so it reports success. It has verified nothing, and said so in green.

This corpus is the input side of that test. The harness that scores against it is abstain-bench; the data is published separately so it is reusable by anyone building a verifier, in any language.

Load it

from datasets import load_dataset

ds = load_dataset("nickh007/abstain-corpus", split="train")
print(len(ds))                                        # 32
scored = ds.filter(lambda r: r["scored"])             # 24
print(scored[0]["name"], scored[0]["why_a_pass_is_unearned"])

Or regenerate it from source — the corpus is constructed, not stored, so every case is reproducible:

pip install abstain-bench
abstain-bench export ./cases

Fields

field	meaning
name	stable identifier
category	which way this input has nothing to check (below)
family	chain, graph, certificate, or any
scored	whether passing it counts as an unearned pass
filename	the name to write it as (.json / .jsonl matters to some parsers)
content	the artifact itself, verbatim
why_a_pass_is_unearned	the reason, in a sentence

Categories

category	n	why a pass is unearned
empty	6	zero records, nodes or bindings. all([]) is True
malformed	5	not the thing it claims to be; the right answer is a diagnosis, not a verdict
out_of_distribution	4	a schema, version or shape the verifier does not know
truncated	3	a hash chain's prefix verifies perfectly; only an anchor reveals the missing tail
no_anchor	3	internally consistent, tied to nothing. Self-consistency is not authenticity
degenerate	3	structurally valid, semantically vacuous: a bound of 1.0, a sample of size zero

The 8 unscored rows, and why they exist

positive_control (3) — valid artifacts a working verifier should accept, one per family. negative_control (2) — clearly broken artifacts it should reject.

These are not decoration. A verifier that exits non-zero on absolutely everything scores a perfect zero unearned passes and is completely useless. Any harness using this corpus must refuse to report a score until the subject accepts a positive control and rejects a negative one — otherwise the headline number is an artefact of the harness rather than a property of the tool.

sound_pass (3) — cases a verifier may legitimately pass, scored in neither direction.

These exist because the first live run of the associated benchmark accused a correct verifier of two unearned passes. {"a": []} — one node waiting on nothing — genuinely cannot deadlock. An acyclic graph whose identifiers contain a right-to-left override is still acyclic; the deception is in how a terminal renders the name, not in the verdict. Both were moved out of the numerator. A corpus that manufactures findings is exactly as untrustworthy as the unearned passes it exists to measure. They are kept and still run, because over-refusal is a real defect too — it is just not this one.

Family scoping matters

A graph checker handed a truncated hash chain will refuse it — because it is not a graph. That is a free correct answer for a reason unrelated to truncation, and counting it inflates the score. Score a subject only on the families whose positive control it accepted.

Honest scope

A 0% unearned-pass rate against this corpus means: on these inputs, in the families the subject demonstrated it handles, it never claimed success on something it could not check. It does not mean the verifier is correct — this covers one failure mode, not soundness — and the corpus is not exhaustive. It covers six ways of having nothing to check; there are certainly others. With 24 scored cases, report an exact binomial interval rather than a bare proportion: 0 of 24 is not "0% ± 0%".

Licence

Apache-2.0. Generated by abstain-bench (abstain_bench.corpus.build_corpus).

---

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 ← you are here
kv-reuse-econ-traces	per-workload reuse accounting + the closed form
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.