When a failing check is the finding
Three proof runs failed in one morning. Each time the easy story was that the test was wrong. Once it was, once it was the product, and once the product was right and the world had changed.

Reason
AgentsQuoxMindAgentic TeamsMirror & QuoxHatQuoxLensRemember
QuoxMemoryQortexCompoundingQuoxPlanCodebase MirrorAct
QuoxFlowQuoxEngineQuoxAgentQuoxChatAutonomyRun
EnterpriseOrganisationsPowers & ToolsmithQlusterQuoxBastionInterfaces
QuoxMCPQuoxCLIQuox ConsoleQuoxBoxQuoxVoiceGovern
HITL ApprovalsQuox SecurityAgent HonestyQuoxVaultAI GovernanceAgentic AIProve
For AuditorsVerifiable AI OpsLoggingQuoxSEOEU AI ActCompliance SuiteChannels
Matrix RoomsDiscord ProTelegram ProQuoxSignalCoreCommsAll products A-ZBuild
QuoxProofDeveloper KitPlugin SDKBring your tool to QuoxQuoxpertQuoxSkillsShip and sell
Browse MarketplaceDownloadsProtocolsDev Suite
Dev ServersQuoxBuildQuoxSlotsShared Skills + RulesQlarityDev workflow
RepoBrainGripeTriageFixLoopProofLoopDoneEngineAll products A-ZThree proof runs failed in one morning. Each time the easy story was that the test was wrong. Once it was, once it was the product, and once the product was right and the world had changed.

Quox proves its own guides by running them. A proof engine asks Quox a question the way a person would, then checks the answer against what the command line says is true. When the check is red, the guide does not go out.
One morning this week, three checks came back red. In each case the tempting explanation was the same: the test must be wrong. That was true once. Once it was the product. Once the product was right and the world under the test had moved. Telling those apart is most of the job, so here is how each one went.
Asked how many commitments had slipped this week, Quox looked and said 0. That was correct: nothing had slipped. The check failed it anyway. The engine treated a matched value of 0 the way it treats a missing value, as no match at all, so a true answer of zero could never pass.
That is a fault in the check, and we fixed the engine. But a fix that makes checks pass more easily is exactly the kind of change that can hide a loosening, so it had to prove it was not one.
Review found an edge before it landed: with the new logic, an empty expected value could have matched any answer at all. A second change closed that, so an empty expectation now passes nothing.
With both in place, the weekly review journey passed 12 of 12 steps.
The second journey asks Quox to produce a signed security posture report. In 2 of 4 runs, Quox did not make a new report. It found the newest existing one and checked that instead. The answer read well, and a person skimming it might have accepted it. But you asked for a report and did not get one.
This time the test was right and the product was wrong. Our rule for that case is simple: never rerun until it goes green. A run that passes one time in two has not passed. We fixed the product so that asking for a report now makes one, and the next run passed all 5 steps.
That run counts from the fix forward, not from the earlier attempts. It is day 1 of 2: a case only earns the word verified on the Evidence Explorer when it passes again on a second day, so this one is not verified yet.
The third failure looked like a regression in a journey that files mail. The cause was in the test organisation, not in Quox: it had gained a second mailbox holding identical mail.
Faced with two accounts and the same messages in both, Quox asked which account to file from instead of guessing. That is the behaviour we want. An assistant that silently picks one of two plausible mailboxes is the one that files your mail in the wrong place.
So the product stays as it is. The thing that gets fixed is the fixture, the sample organisation the journey runs in, so that it once again describes the world the journey is meant to test.
Three red checks, three different answers: fix the check, fix the product, fix the fixture. Making the light green would have been easy in every case, by widening the check, rerunning until it passed, or accepting whatever Quox did. Each of those would have produced a guide you could not trust.
The goal is a guide that tells you what Quox will really do, backed by a run you can inspect. Every published case, with its runs and receipts, is on the Evidence Explorer.