← All Research Notes
Design & ValidationEngineering note

Why every CAI value needs an authority record

Two CAI values for the same sequence were both mathematically correct—and still misleading when shown without their references.

2 min readMetric-provenance observation

The apparent discrepancy

The same output sequence received different CAI values in two evaluation paths. The formulas were not broken. Each calculation used a different codon reference.

One path evaluated against an expression-oriented reference. Another evaluated against the working table used during generation. Without a label, both numbers appeared under the same field name and invited a comparison they could not support.

From a number to a record

The corrective action was additive: preserve the numerical result and attach its authority. A useful authority record answers at least four questions:

  • Which reference was used?
  • What role did that reference play?
  • Was it distinct from the generation reference?
  • Was a fallback used?

An initial two-state model—preferred reference or fallback—was still incomplete. It could not represent the valid case where a caller intentionally selects one reference for both generation and evaluation. The final relationship therefore needed three states: distinct, intentionally the same, and fallback to the generation reference.

Why labels are part of the result

Metric provenance is not administrative decoration. A CAI value without its reference cannot be reproduced or interpreted reliably, and two unlabeled CAI values may measure different relationships while looking identical in a table.

The same principle applies to model-derived scores and validation summaries. A result should identify the authority that produced it and the relationship between that authority and the candidate-generation path. Otherwise, later readers are forced to reconstruct meaning from implementation details.