Native2 Training EN

Turning the two observed requests into training data

After checking how similar functions behave differently, two existing draft requests were qualified for development training. Five candidates have two positive and three negative finite labels. Repeated inputs do not increase request weights; each candidate has unit weight.

Unexercised caption clauses were removed. Nil-output success, empty-hook error and a trailing line feed after each failure message are explicit. Every new text passed the existing Go limits of 512 bytes and 32 normalized words without truncation. Function names remain identifiers and are absent from model-visible descriptions.

The requests are development-only. The old 79-request corpus and its evaluation/calibration parts remain unchanged; no new model has been fitted or evaluated. Qualified data in this round expands toward 30/60 before a fixed-recipe comparison. Protected evaluation of 2,400 requests per claimed domain and existing utility gates remain unchanged.

The next request concerns rounding. Six inputs and three candidates prepare 18 observations, with separate mathematical oracle/Want review. Execution count remains zero pending the original-file boundary, compatible controller, compiler-source and resource checks.

Qualified data and limitations · Original behavior observation · 한국어