Silent Unless Exact: Clean Synthetic Training Data from Any Model's Examples

A model-free program synthesizer that ships only programs its independent solvers agree on turns a few examples written by any language model into exact training data, and stays silent on the rest. High-quality human data is running out, and more of the data that trains the next models will be written by models; unchecked, that data carries the writer's mistakes forward. We describe multi-PLUMB, a small, local synthesizer that compiles input/output examples into a program in its own fixed machine language, ships it only when its four solvers agree, and otherwise stays silent. Every program it ships reproduces its examples exactly, so it is exact data for its own rule, and it labels any number of new inputs at no marginal cost. In sealed, pre-registered tests, all 95 programs it compiled from examples written by Claude Opus 5.5 reproduced their examples and produced 90,355 new labels with no errors against their own programs and no variation between runs. The model that writes the examples sets only the yield: 57 of 60 tasks compiled from Opus 5.5's examples and 3 of 60 from a 4-billion-parameter open model whose examples were 63% wrong, none of whose errors became data. Where the data must also mean what the writer intended, requiring the writer's own code to agree removed every look-alike program (55 of 55 kept programs matched the true rule). A frontier model writing labels itself cost $1,606 per million labels; the factory cost about $27 per million at 1,000 labels per task. On fresh machine-made tasks multi-PLUMB alone scored 86 against Opus 5.5's 56 (right minus wrong). We report equally where it loses: on rules written by people, Opus 5.5 was perfect while multi-PLUMB, after learning from 200 such rules, answered half of them. Implementation, experiments and text were produced with Claude Code (Anthropic) under the author's direction. PLUMB's core and gate are not released.