How we test

Use the incumbent as the oracle.

When an established implementation has a behavioral suite, use it. Do not replace somebody else's contract with a demo written to flatter the new architecture.

Every failure gets three questions.

01

Is our implementation wrong?

Fix it. Add a regression test. Re-run the full gate.

02

Did we misunderstand the contract?

Correct the model before correcting the code.

03

Should the new model decline this behavior?

Write it down. Do not turn a deliberate semantic difference into a fake pass.

Never turn question three into question one just to make the suite green.

A bug report is not the end of the analysis.

Known defects are carried with their engineering disposition so candor is not mistaken for ignorance.

StatusWhat it means
UnknownThe mechanism has not been investigated far enough to make a stronger statement.
Root cause knownThe failure mechanism has been inspected and understood.
Solution designedA concrete correction is understood, but it is not represented as running code.
Deliberately deferredThe solution is known, but present requirements do not justify paying the implementation cost ahead of work that currently constrains the experiments.
ImplementedThe correction exists in code.
QualifiedThe implemented behavior has been tested against an explicit contract and condition.

That distinction matters in a research implementation. A five-machine development network does not need a roughly 59,500-node allocator merely to prove that the replacement can be designed. Deferring it preserves code velocity; it does not turn a solved design problem back into an unknown one.

Qualification is not characterization.

QuestionMeaning
QualificationDoes the required contract hold under the stated condition?
CharacterizationWhat operating envelope did this implementation, machine and path actually demonstrate?
PolicyWhat administrative limit do we choose inside that measured envelope?
PlanningDoes a proposed workload fit the measured and policy-constrained system?

Freeze the evidence.

Eight application classes. One vocabulary.

Psychedelic FrogTorch, Redis-on-Ribbit, Communicator, Monitor, Sourri, residency/swapper, FrogNet discovery/resolution and physical control do not share a business domain or timing model. The question is whether independently addressable truths, relationships, participants and held reads remain useful across all of them.