Correctness
Invariants hold; decisions match evidence; the record reflects what actually occurred
Rules that must hold do hold at commitment time — and every committed fact can be traced to the authority, evidence, and moment that produced it.
Correctness is the aim that the system’s committed state is actually valid — not merely internally consistent, but aligned with the invariants, evidence, and authority the domain requires.
The aim
Correctness distinguishes “the database agrees with itself” from “the database agrees with reality and policy.” A fraud score stored as fact may be consistent across replicas and still wrong epistemically. A committee vote recorded without a vote occurred may be consistent in the minutes and still incorrect as organisational fact.
Pursuing correctness means enforcing invariants locally inside single writers, committing only at boundaries with evidence, and treating uncorrectable inference as derivation — not ground truth.
In software systems
Correct implementations enforce invariants inside the writer, reject invalid transitions before commit, and emit L0 events that carry enough structure to replay and audit. Correctness is verifiable: given the log, an independent reader can determine whether each commitment was valid under the rules at that moment.
In human organisations
Correct organisations assign accountability coextensive with authority. A decision recorded without the decider having reviewed sufficient evidence is incorrect — regardless of workflow completion. Rubber-stamp approval is a correctness failure, not a speed optimisation.
In socio-technical systems
Correct socio-technical systems align automated checks with human commitment at severity-appropriate boundaries. Model output may inform; it does not substitute for named authority where the domain requires it.
What threatens this aim
Failure modes: Inference as fact and rubber-stamp approval produce structurally incorrect records that appear valid. Bundled events hide which part of a composite commitment failed validation.
Anti-patterns: Epistemic tier collapse mislabels interpretation as observation. Diffuse write authority makes invariant enforcement impossible to verify locally.
Principled support: Commitment boundary and single writer principle advance correctness by making invalid states unwritable and violations traceable.