Evidence ledger · 2026-09-06 · revision 1
This page records named local campaigns and their limits. A passing campaign proves the behavior it exercised on that version. It is not an external audit, a production certification, or proof of every claim in the broader research archive.
Counts are exact for the recorded run; scope and version travel with the count.
Reproduced locally
Shows: exact round trips and matching finite compiler outputs across the exercised transforms.
Does not show: general semantic understanding, completed control flow, or a finished 18-language toolchain. Ten Rosetta runners remain unfinished.
Inspect the public toolchainPassed locally
Shows: local command correlation, capability propagation, structured denial, bounded output, and descendant timeout behavior in the tested harness.
Does not show: universal CLI superiority, production penetration-test coverage, broad hosted-provider quality, or multi-machine coordination.
Inspect the Agent Bus surfaceExact control
Shows: exact, invertible, compositional arithmetic under the tested prime-power encoding.
Does not show: a task-level advantage over a direct exponent vector. The representation is classical unique-factorization machinery used as a coordinate sidecar.
Experimental
Shows: a concrete experimental program for testing topology-guided routing and representation ideas.
Does not show: that geometry caused a model gain. Current results need stronger held-out evaluation, controls, and independent replication.
Read an underpowered resultMultiple seeds, a verified evaluator, dispersion large enough to judge the effect, and wins against both a size-matched control and the no-intervention baseline.