The Weismann Machine: Outer-Loop Self-Improvement behind a Germline Barrier, with Time as the Grader
Year: 2026
Type: preprint
Abstract: Self-improving AI systems now edit their own code, archive their variants and draw parents from the archive largely by benchmark score. The agent being improved writes its own successor from its own traces, and each child inherits its parent's code, edits included. In 1892 August Weismann argued that the germ line is set apart from the body, so that what an organism acquires during its life is not passed on. This paper turns Weismann's separation into a design rule and calls a system built to it a Weismann machine. Its inherited genome is compact and specifies how a learner develops and learns, and whatever a lifetime learns is discarded at its end. A germline barrier confines each lifetime to private storage and a single typed output, and lets the operators that vary the genome see genomes and a statistic of their fitness and nothing else, so that variation and selection happen only in an outer loop. The grader is time. Learners are scored with strictly proper scoring rules on events that resolve after a sealed cutoff, across rolling cutoffs, with a live tier of forecasts committed before their questions resolve. Under these conditions the barrier excludes tampering with the grader and with what it reads, prevents the inheritance of lifetime state, and bounds the information about the selection data that can reach the genome. The paper places the scheme against self-improving agents, learned learning rules and outcome-trained forecasters, and specifies a test on a language model with a 1913 knowledge cutoff, with nine arms, a budget of about 37,000 H100-hours and a decision rule fixed in advance. Its comparison arms separate the effect of scoring after a sealed cutoff from those of scoring format, search, recalibration and added machinery.
Author: Sadineni, Abhay Chakra
Institution: University of California, Berkeley
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236144342
