Skip to content
IsoKron

Ethics

Editorial independence and research ethics

How IsoKron keeps its audits, research results and ratings independent of who pays, how its methods change, and how anyone can check and correct what it publishes.

In short

  • Payment never changes a finding, a severity, a rating or a decision to publish.
  • Every audit and rating published under this policy states who paid for the work.
  • IsoKron keeps final editorial control. Vendors and clients correct facts; they do not change findings.
  • The MWS standard and the IsoKron rating are separate. A rating is IsoKron’s statement, not an MWS result.
  • “Not tested” is a state of its own, never counted as meeting a requirement or as a violation.
  • Every result published under this policy carries the command, the raw records and the tester, specification and method versions.
  • Method changes are announced 30 days before they apply.
  • Corrections are dated and public. Nothing is changed silently.
  • A named vendor’s statement is shown beside its result once received, and anyone can dispute a published result.

These rules apply to results, audits and ratings that IsoKron publishes under this policy from its effective date, 18 November 2026. Pages published earlier keep their own funding lines and changelogs.

Version 1.0, 18 November 2026.

Funding and conflicts

  • How IsoKron earns money. IsoKron sells MCP server audits, builds and advisory, and maintains the open-source tools used in its testing. It also tests and publishes on servers whose vendors could become clients. This is a conflict of interest, and the rules below exist to manage it.
  • Payment never changes a result. The same method, the same rules and the same rating method apply whether a vendor has paid, has never paid or has declined to buy. No payment changes a finding, its severity, a rating or a decision to publish.
  • Every result names its funding. Every published audit opens with a funding line. Public research reads as our published audits already do: “Self-funded research by IsoKron … Not commissioned or endorsed by [vendor]”. A client report published at the client’s request reads “Commissioned and paid for by [client], published at [client]’s request”. Every rating states whether the vendor has paid IsoKron for any service.
  • Servers we sent fixes to. When IsoKron rates a server it has sent fix pull requests to, the page says so and links each pull request, and the same method applies. Results on a build that carries IsoKron’s own patches test IsoKron-authored code: those “after” results are labeled so, and are not an independent result.
  • Research is not for sale. No payment starts, stops, delays or alters public research. Public research cannot be requested or bought.
  • Public stays public. If a vendor whose server we have published on later becomes a client, the published results stay as they are.
  • No selling on open findings. We make no sales contact on an upstream thread where we have open findings.
  • One price for the audit itself. The audit has one price, listed on the pricing page, in every mode. The design-partner discount is for a marketing case study and never changes a result. Publishing a report earns no discount, and no price depends on a result.
  • Design-partner case studies. The case study is a marketing piece, separate from the test results. You may approve its text, never the findings, verdicts or ratings: those follow IsoKron’s ethics policy unchanged and are published the same as for anyone. The discount buys the case study, not a result or a publishing decision.

Three audit modes

The method, the evidence and the rules on findings are the same in every mode. What differs is who chooses the server, who pays and who sees the report.

The three audit modes compared: Private audit, Private, then published, Public research
ComparedPrivate auditPrivate, then publishedPublic research
Who chooses the serverThe clientThe clientIsoKron
Who paysThe client, at the one audit priceThe client, at the same priceIsoKron. It cannot be requested or bought
Who sees the reportThe clientThe client, then everyoneEveryone
When it is publishedOnly if the client later asksAt the client’s request, once fixes are re-tested or the agreed window endsWhen IsoKron decides. Security-class findings follow coordinated disclosure
Backing outNothing is publishedThe client may withdraw the request at any time before publication. Once public, it stays publicNot applicable
FingerprintKept by IsoKron; we confirm on request whether a copy is authenticKept by IsoKron, so the published report can be checked against itNot needed: the report is public
Facts and findingsFacts can be corrected; findings cannot be changedThe sameThe same, for the vendor

Editorial control

  • IsoKron keeps final editorial control over every report, result and rating it publishes.
  • A client or vendor may show us an error of fact, with evidence, and we correct it. Nobody can change a finding, its severity, a rating or our conclusions on request.
  • A vendor or maintainer may add a signed, dated statement of up to 300 words. We publish it beside the result, unedited.
  • Failed and void test runs stay in the signed record. A later run never hides an earlier one.
  • No one outside IsoKron reviews or approves a page before it is published. Vendors reply after publication, and their statement is shown on the page.

Public research

  • IsoKron chooses the open-source MCP servers it researches and pays for the work. Results on these servers are published.
  • The exception is security-class findings. They are kept out of the published results and reported privately to the maintainers through the project’s own channel, under our disclosure policy and its 90-day default embargo.
  • Correctness findings go to public issues on the project’s own tracker, filed before the results are published. Maintainers owe IsoKron nothing: not a reply, a fix or a review.
  • We test on throwaway organizations, local instances and test accounts, never on anyone’s production data.

The standard and the rating

The MCP Write-Safety Requirements (MWS) are an open specification with an open-source tester. An MWS result is a set of test records, one per requirement and tool: what was called, what came back and what was stored.

The IsoKron rating is a letter that IsoKron computes from MWS test records by a versioned method, published before the first rating appears. It is IsoKron’s statement, derived from results. It is not an MWS result, it is never shown with the MWS mark, and the tester never prints it.

IsoKron rates one version of one server, not a company or a person.

IsoKron certifies no server and gives no “certified”, “approved” or “passed” label. A rating covers write mechanics on the tools and settings tested; it says nothing about authorization, prompt injection, tools that are destructive by design, production behavior or anything not tested.

A letter is never shown without its coverage beside it: how many requirements were tested of those that apply.

Popularity, stars, downloads, documentation, metadata and payment never enter a rating.

What “not tested” means

MWS gives every test record one of these verdicts, with the meanings the specification defines:

violated
the forbidden behavior was reproduced.
platform
the platform’s own API does it too; the server passes it on.
held
forced, and handled correctly every time.
not observed
attempted, not forced, not seen; not evidence of meeting it.
not tested
not run, with the reason.
not applicable
the condition is false here.
self-attested
a person’s answer to a review question.
  • Each requirement on each write path gets one MWS verdict. A verdict is a test record, not a judgment of the server.
  • “Not tested” and “not observed” are never counted as meeting a requirement. “Not tested” always carries its reason (for example: needs credentials, no write tools, timed out).
  • “Not applicable” is used only under a written applicability rule, whose id is printed with it. In doubt, a requirement applies.
  • A “not observed” search result always carries the scope of the search, and is never described as clean, safe or verified.

These states belong to the IsoKron rating, not to MWS:

Seen, not counted
A violation was recorded, but its evidence failed a validity rule (for example, the run did not complete). It is shown with the rule it failed and does not affect the rating.
Not rated yet
Too little was tested for a letter. This is a state, not a low letter. The page still prints everything found and which coverage conditions were not met.
Flagged
The server behaved differently under the tester than in control runs of the same operations. The page shows the observed difference; it states no motive. A flagged result carries no letter until the difference is explained.

Reproduction

Every published result carries:

  • the exact command that produced it, to run yourself
  • the raw JSON test records
  • the tester and its version, the MWS version and the rating-method version
  • the server version or commit, and the backend it ran against
  • the run date
  • the source: run by IsoKron, or self-run by the maintainer
  • whether the vendor has paid IsoKron for any service

Audit evidence is delivered as an Evidence Pack: signed, independently timestamped, and checkable offline on your own machine with the open-source kronicle-lite tool (how to verify a pack).

We keep the fingerprint of every report we deliver and confirm on request whether a copy is authentic. Nothing about a private engagement is published without the client’s written consent.

A fingerprint shows that a document existed in exactly that form on that date. It is not a certification of the findings or of the server.

Method versions and change notice

  • The rating method has a version number, and every rating names the version it was computed under.
  • A material change to the method is announced at least 30 days before it applies, whichever way it moves letters. The notice lists what changes and, for each published server, the letter it would change to.
  • Every change is listed in the method’s changelog with its date.
  • A published result keeps the versions it was produced under. Results under different method or specification versions are never compared without their version labels.
  • Changes to the MWS specification follow its own public comment process, on the specification page.

Gaming the tester

  • Behaving differently under test is a finding. Where a server behaves differently under the tester than in control runs of the same operations, the result is marked “Flagged” with the observed difference, and no motive is stated. IsoKron decides whether to flag, after inviting the maintainer’s statement, and logs the decision on the page. A flagged result carries no letter until the difference is explained.
  • Private test cases stay in private audits. A private audit may use test cases that are not in the public tester. Any published result that rests on such a test case carries its full definition. Public fixtures are added or rotated with minor versions of the specification.
  • Every run stays in the history. All runs of a version stay in its history, failed and void runs included. Re-running a version until a better letter appears changes nothing: a counted violation in any valid run counts.
  • Self-run results are labeled. A result a maintainer produced with the tester is labeled self-run. Only runs made or reproduced by IsoKron are labeled as IsoKron runs.
  • Objections. Anyone who believes a result was gamed can say so through the dispute process. An allegation is published only once evidence substantiates it.

Citing a result

These rules apply to vendors, clients and anyone else who quotes an IsoKron result.

  • Give the result’s date and the server version it covers, and link to the dated result page, not to a general page.
  • Quote a letter only with its coverage line and its method version, and call it the IsoKron rating.
  • Do not use the MWS mark on or beside a result (mark policy). To say a server was tested against MWS, use the sentence in section 15.5 of the specification.
  • Do not describe a server as certified, approved, verified or passed by IsoKron or by MWS.
  • When a newer result exists, an older one may still be cited with its date.
  • Any image that shows an IsoKron result links to the dated result page.
  • A client may say it worked with IsoKron and may give a testimonial. A client may not state or imply what an audit found, or that it passed, unless the full report is published.
  • Where a result is misquoted in public, we may publish a correction note on the result page.

Right of reply and disputes

A named vendor’s statement is shown beside its result once received, and anyone can dispute a published fact, finding or rating, with no customer relationship needed. The process, the evidence that helps, the timeline and the public log are on the disputes page.

Corrections log

We correct errors of fact, finding or attribution in public. Each published page lists its changes, dated, in its own changelog.

A change is listed here when it altered something a published page had already stated: a fact, a count, a figure, an attribution or how a finding was described, at any time after first publication. Additions, new ratings and status updates are not corrections and stay in each page’s own changelog.

Corrections to published research, newest first
DatePageCorrection
Payload MCP server auditThe sentence that no finding is rated S1 now says it refers to the original scale; on the rubric PAY-03 is Critical.
Grafana MCP server auditWith GF-03 and GF-21 rated High and GF-01 Medium, the introduction calls GF-01 the headline finding rather than the most serious.
GitHub MCP server auditWith GH-14 also rated High, the introduction calls GH-01 the headline finding rather than the most serious.
Lost updates across four MCP serversThe resources_scale sentence now says its check covers one call, not an agent’s older read.
Kubernetes MCP server auditThe resources_scale line now says its version check covers races inside one call, not an agent’s older read.
Lost updates across four MCP serversThe figure marks pairs in which one call was rejected with a conflict, instead of counting them as both changes kept.
Grafana MCP server auditThe summary no longer presents the skipped version check as the cause of the 30 of 30 lost updates, since Grafana’s own save API lost simultaneous saves too. The figure marks pairs in which one call was rejected with a conflict.
GitHub MCP server auditThe retest step in the checking figure now covers only the eight issues it was run on, as the method list says.
GitHub MCP server auditLater on the day of publication, the 26% figure was attributed to the two batch-read tools, and the harness timing and limits were stated.

Changes to this policy

  • 18 November 2026: Version 1.0 published.