Variability is not a detail
Probabilistic models may produce different readings of the same material. For creativity that can be useful. For audit, legal, risk or investigation it can be unacceptable when nobody preserves the evidence behind each claim.
Bound the reasoning
CERBERUS separates claims, evidence, contradictions, alternatives and missing information. The goal is not forcing a model to repeat identical wording, but making a defensible conclusion depend on verifiable material rather than the statistical mood of one run.
Reproduce what matters
A system can use probabilistic models while keeping contracts, state, evidence and promotion rules deterministic. Reproducibility belongs wherever a difference can change a decision.
Separate linguistic variation from factual variation
Two answers may use different words while supporting exactly the same conclusion; they may also sound similar while relying on incompatible facts. Useful reproducibility does not require freezing every sentence. It requires stabilizing which sources were used, which claims were extracted, which contradictions remained open and which rule allowed a conclusion.
Extraction may be probabilistic; evidence should not be
A model can help locate entities, events or relationships in difficult documents. Those extractions can then become verifiable claims linked to their origin. This preserves probabilistic flexibility without allowing later paraphrase to erase the connection to the material that produced it.
Repeatability is not truth
A perfectly deterministic procedure can repeat the same mistake forever. Stability and validity should therefore be measured separately. Good architecture aims for the same evidence set to produce a reconstructable record while preserving mechanisms capable of discovering that evidence or interpretation was wrong.
Versioned context changes the conversation
If a source, policy or engine version changed, a different result may be legitimate. The system should explain what changed between runs. Without versioning, every difference looks arbitrary; with version and provenance, variation becomes diagnostic information.
The goal is reconstruction
Months later it should be possible to revisit a conclusion and answer: what was known, what was missing, which sources were current and which steps transformed those inputs into the result. That capability matters more than forcing a model to reproduce an earlier output word for word.
A useful test
One concrete way to put this idea under pressure is to run the same case several times and compare not wording but the sources, claims, contradictions and decision rules supporting the result. The test should not ask only whether an answer appears, but which state remains, what evidence is preserved and whether another operator can understand why the system behaved that way. This turns an editorial principle into an observable property and exposes places where architecture still depends on invisible assumptions.
What this note does not claim
Reproducibility does not mean every inference must be deterministic or that a repeated conclusion is true; it means material differences can be located and explained. This distinction matters because a good practice stops being useful when it becomes a universal promise. The goal is to make one design boundary explicit so it can be discussed, tested and adapted to the domain while facts, inferences, permissions and decisions remain separate.