Criterio CFO · The finance functionStatistical analysis of the tests

On each entity, Clara, Tomás and Áurea run their setup and processing —the commands that process the documents and write the records—, with a token cap set before starting, on material whose key is kept apart and never seen. The other commands calculate on those records with the engine, which is tested on every change to the repository. Signals are compared by name and key; extraction, field by field, against what the written documents say. Control entities carry no seeded defect: any signal in them is a failure.

Coverage: of the seeded signals, how many were found. Precision: of those raised, how many were right. 95% margin: Wilson confidence interval over the cases.

A round that does not run every entity is partial: the first, small one, before the complete one, or a verification round, which repeats the runs a fixed defect left pending.

2 failures in 45 runs, each with the defect that explains it: its origin says whether the agent or the test failed. The figures are those of the original grading: 1 of those failures came from the test —the harness or the key—, not from the agent, and with the test fixed they grade as right.

12 defects logged, each with its origin, its fix and the run that verified it: 12 verified and 0 to verify. 6 came from the material, the key or the test harness, not from the agents.

The 12 rounds used 238.2 million tokens: 227.5 million in the graded runs and 10.7 in the ones repeated after fixing a defect. On average, 5.1 million per entity.

The material is synthetic. Whether a version is stable or laboratory is said by this evidence, not its number: stable, with the floor of the 95% margin at 95% or more for signals and data over each entity's most recent run, and no signal in the controls (decision 0020).