Pyx

Pyx: Every lesson stands trial. Pyx is a gate for agent memory. Agents that learn from experience usually remember whatever they produce, including wrong, lucky, and poisoned lessons. Pyx makes every candidate lesson earn admission to a shared bank of earned lessons: graded against a hidden rubric by two independent Meta judges, screened for planted poison, tested by causal replay (must fix its target criterion in at least 2 of 3 repetitions and break nothing), and approved by a human reading its exact text. In the counted run (frozen protocol, Nov 2 to 3, 2026), 11 candidate lessons stood trial, 10 were rejected on evidence, and exactly 1 was earned. The earned lessons arm scored a 0.625 criterion pass rate versus 0.5208 for all lessons unfiltered, 0.4375 for no memory, and 0.4375 for raw traces, with paired bootstrap 95% confidence intervals excluding zero against every arm. The whole run reruns keyless from its own call log, reproducing decisions and metrics byte for byte, and the bank is served read only over MCP. Total counted run cost: $1.29.