Skip to content

A measurement you cannot check is a press release

Twenty-five open-source artifacts. One idea.

Every verifier 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 it has said so in green.

Every tool here is built to refuse that. When a tool cannot check something, it says so, in a way you cannot mistake for a pass.


The rule, in three lines

exit 0    checked, and the property holds
exit 1    checked, and the property does NOT hold
exit 2    NOT CHECKED — and this is never a pass

That third value is the whole portfolio. It is why gridlock refuses to call an empty graph deadlock-free, why kvleak refuses to report a null it cannot interpret, why preregister refuses to seal an experiment whose conclusion is already fixed, and why evidence aggregates to the weakest leg, never the mean.


Why these are one body of work and not twenty-five unrelated repos

They came out of one project — a proof-carrying inference stack — and each one exists because that project got something wrong first. The tools are the scar tissue, extracted and made general.

the tool the mistake that produced it
preregister we sealed a pre-registration whose positive branch was structurally unreachable. prompt_logprobs=0 pins argmax_flips to zero; the sealed rule was argmax_flips > 0. Public, honest, and guaranteed to report the null before a GPU started
gridlock our own checker certified {} — the empty graph — as SAFE, and printed "rank strictly decreases on every edge: yes". all([]) is True
abstain-bench when we first pointed it at a real tool, it accused a correct verifier. Two of its findings were wrong
kvleak a clean cross-tenant result whose positive control had failed — the probe could not have seen a leak if one existed
gatecount "we fuzzed it and found zero escapes" is a statement about a search, not about the system
evidence four checks passing and one that could not run is not "80% verified"
signoff-cert a certificate that does not carry its own false-pass bound is a sticker
honestbench our CI was green. We deleted the evidence it checked. It was still green

That last one is the shortest description of the whole portfolio. If deleting your evidence does not turn your build red, your build was never checking your evidence.


Where to start

  • :material-play: I want to try it


    One command over your own repository, in about thirty seconds.

    → Install · → The tutorial

  • :material-lightbulb: I want to understand the idea


    Why "not checked" is a third value, and what that buys you.

    → The abstention discipline · → Verdicts

  • :material-scale-balance: I want to know what it actually proves


    The honest scope, stated as matched pairs: what a pass means, and what it does not.

    → What this proves

  • :material-book-open: I want the reference


    Every tool, every command, every API.

    → Tools · → CLI · → API


Sixty seconds

pip install "evidence[all]"
evidence audit .

Against a repository with a genuine lock-order inversion:

  gridlock       WEDGES: conn_lock -> stats_lock -> conn_lock   [FAIL]
  honestbench    no evidence manifest found (honestbench n...   [n/a]
  proof-drift    no Lean sources, so there is no proof to ...   [n/a]
  sf-verify      no hash-chained decision log (.jsonl with...   [n/a]
  signoff-cert   no signoff-cert/v1 certificates found          [n/a]
  --------------------------------------------------------------------------
  AGGREGATE: FAILED — 1 fail, 4 n/a
  gridlock is FAILED, and the aggregate is the weakest leg — 0 other constituent(s) passing does not lift it

Note what it did not say. It did not say "1 of 5 checks failed, 80% healthy". Four tools found nothing to look at; that is not four passes.


The map

Run everythingevidence · proof-carrying-ci (the GitHub Action) · formal-proof-mcp (for coding agents)

Check a propertygridlock (deadlock) · floorgen (state floors) · gatecount (over-acceptance) · proof-to-code-drift · tokencount

Check the evidence itselfsignoff-cert · honestbench · sf-verify · preregister

Check a live systemkvleak · kvprobe

Check the checkersabstain-bench · illusion-bench

Reproduce our own numberskv-reuse-econ-bench · llm-tenant-isolation-bench

Full table with one-liners and honest-scope notes: Tools reference.


What none of this does

Nothing here gates anything. Every tool computes a verdict and prints it; the exit code is advisory, and no package ships a mechanism that would withhold, block, or refuse an operation on the strength of its own result. That line is deliberate and it is documented per-package in each repository's CLAIMS-MAP.md. See Measure vs gate.

Nothing here claims a speedup. There is no performance claim anywhere in this portfolio. Where we measured one and it did not survive, we said so.


Apache-2.0. Built by @nickharris808.