The Independent Reliability Registry for Agent Tools
Payment rails prove money moved. Nobulex proves what happened on the other side.
What this register is for
Agent tools are checked for two things today. Is this the package it claims to be, and does it contain a known vulnerability. Both are questions about the code. Neither is a question about the answer it returns.
When the underlying data is missing, stale, malformed, or contradictory, does the tool stop, or does it return something plausible and wrong?
The failure has a name: silent semantic corruption. A response that is well formed and plausible but materially wrong, stale, mis-scoped, truncated, or misattributed. It passes every schema check and every scanner, because nothing about it is malformed. It is simply not true.
This matters because an agent cannot distinguish wrong data from no data. It acts on both. A human reading a stock price that is four days stale may notice. An agent will trade on it.
One test governs this register: does it fail loud, or does it lie quiet?
The mark
An assay office does not certify a silversmith. It tests one object and strikes a set of punches into that object recording what was tested, what standard was met, who tested it, and when. The mark travels with the object, not the maker.
A Nobulex record works the same way. Four punches, and the mark is never about a project name.
The analogy breaks in one place, deliberately. A struck hallmark is permanent, and permanence is the wrong property here. A silver standard does not change after assay. A running server does. So the Nobulex mark is not struck into anything: it is resolved against this register, live, every time it is claimed.
Subject
The exact package, version, commit, configuration, and upstream source that was tested.
Verdict
One of five states. Never a score, never a percentage, never a grade out of ten.
Office
The issuing key. Who ran the suite, under which methodology version.
Date letter
Observation time and validity window. Outside it, the record reads EXPIRED regardless of verdict.
There is no badge image. A graphic sitting on someone else's site is a claim about the past that keeps asserting itself in the present. Status is queried live against this register or it is not a status.
The five verdicts
PASS
Correct within a tolerance that was pinned before execution, against a named ground truth source.
FAIL_SAFE
Could not answer, and said so. The tool refused, errored, or returned an explicit null. Recoverable, and still a failure.
FAIL_UNSAFE
Returned something plausible and materially wrong, with no error signal. Disqualifying.
INDETERMINATE
No ground truth was available to decide the question. Not a pass.
OUT_OF_SCOPE
Not covered by the conditions pinned for this run. Not a pass.
FAIL_SAFE beats FAIL_UNSAFE and is not equivalent to PASS. It is still a failure. A tool that stops when it cannot be sure has left you something to route around. A tool that guesses confidently has not.
The register
Held, not published
3 records issued and withheld.
Gate: right of reply, subject under embargo. A record whose finding is adverse to a named maintainer is not published on the day it is issued. That maintainer first receives the full run artifact and has seven days to answer, and the answer publishes beside the record, unedited. The verdict is fixed before the window opens and the window cannot change it; only a new run under newly pinned conditions can.
One of these is this registry's own withdrawn record. It is adverse to nobody but us, and it would have published the day it was written, except that its subject is the same package as the records now under reply and is the only subject this page would name at all. Publishing it says that Nobulex is withholding adverse findings about one identified project while showing none of the evidence for them. That is the accusation the window exists to prevent, minus everything the maintainer would need to answer it. It publishes when they do. Nothing was ever served to a reader who would now be owed the correction, so holding it costs that reader nothing.
What a held record found is not stated here, not its verdict and not who it is about, because a finding a reader can infer but nobody can check is worse than one published in full. The count is compiled from the records, so a held record can be forgotten about but cannot be hidden.
Each held record is committed to by sha256 in records/held.manifest.json, which is in version control while the records themselves are not. When one publishes, hash it. Matching the value committed on the day it was issued is what makes “the verdict is fixed before the window opens” something a stranger can check rather than something we assert.
0 records published, 3 issued and held, and 0 withdrawn and published, compiled from the records on every build. Every record this registry has issued is currently held, so this page carries its rules and its count and no subject at all. That is the accurate state of it, not an empty template. A held record is in force: it is issued, its verdict is fixed, and it is waiting on its subject rather than on us. Nothing on this page is typed by hand, because the one time it was, it published a subject that could not be resolved.
What a verdict warrants
Stated here, in advance, so that it is not defined for us later by someone disappointed.
A record does say
This exact build, in this configuration, was given these inputs at this time.
It returned these outputs.
Under the pinned conditions, that behavior was correct, safely refused, or materially wrong.
The evidence is reproducible by anyone running the same public suite.
A record does not say
That the tool is secure, or free of vulnerabilities.
That a different version, config, or upstream will behave the same way.
That the vendor is trustworthy. We test objects, not companies.
That the result still holds today. Check the validity window.
That any loss is indemnified. Nobulex is the evidence provider, never the custodian.
Nobulex · buyer funded, never vendor funded
the method is published · the suite will be public · the mark is queried, never carried
disputes and corrections: nobulex.dev@gmail.com