Jev for Engineers

Real calls, with unusually honest limits. Every example was executed against the live API and there are no mocked or recorded responses anywhere; the only fixtures in the test suite are copied verbatim from TypeSafe's published examples. What is not established is scale, since the inputs are ten synthetic ECRs, eight BOM pairs and one solver log, which the README calls a demonstration rather than a calibration.

Not structurally, for extraction. In example 05 the model can only return a span cut from the document, so a fabricated dimension is unrepresentable rather than merely unlikely, and the pipeline is regex finding candidates with Jev selecting among them. That guarantee covers spans only, not the judgement wrapped around them or the thresholds fitted on top.

Three defects, all the author's, documented in place. Example 05 rejected the correct tolerance because the guard question was wrong, not the threshold. Example 08 asked a judgement model to do arithmetic, got confidence between 0.04 and 0.28 with nouls near 0.5, and the gates still said BLOCK, which looked like success. Example 07's thresholds were guesses and an obvious substitution missed by 0.01.

Not directly, because a noul carries no confidence field at all, so a gate that thresholds confidence on a noul can never fire. The repo derives one with jev.confidence_of(), where 0.5 means maximal ignorance and either pole is a firm answer, and pins all three points in the test suite because a guard that cannot fire is worse than no guard at all.

Python 3.10 or newer with no dependencies; jev.py is about a hundred lines of stdlib. Every example runs without a key and prints the exact request it would send, and `python -m unittest discover -v` runs 28 offline tests. Export TYPESAFE_API_KEY to make real calls.

noul is a yes/no that returns the probability the answer is yes and carries no confidence field; choice picks one from a closed set and returns the winner, the full distribution and a confidence; score rates ordered levels and returns a probability-weighted float plus legend, distribution and confidence.

MIT. All sample inputs under data/ are entirely synthetic: the manufacturer names are real companies, but the part numbers, prices and descriptions were invented, so do not read an engineering conclusion out of them or use the BOM file as a cross-reference table.