Abstract

This systems paper separates experimental work, evidence of a measured change and the authority to alter a running system. It retains the accepted-work and transition predicates of the original edition, and relates them to the completed October evaluations without treating an implementation event as an empirical improvement.

Mathematical definitions

T_eff = ∑ⱼ aⱼ τⱼ
Accepted scheduling time over completed work units.
A = I_exact ∧ R_complete ∧ Q_pass ∧ B_safe ∧ H_authorized
Identity, complete records, quality decision, safe boundary and authorization.

Experimental identity

A result is associated with a specific input policy, declared procedure and completed observation set. A proposed change, an executed experiment, a measured result and a deployed state are separate objects. Their records should not be replaced by a single task-completion flag.

Accepted-work accounting

Effective work counts only completed units accepted under their original identity. Let aⱼ indicate acceptance and τⱼ the elapsed assignment-to-acceptance time of unit j. Summing aⱼτⱼ gives a scheduling measure. Overlapping independent worker assignments may contribute simultaneously; this is not wall-clock duration, hardware occupancy or energy use.

Transition predicates

A transition record checks input identity, evidence completeness, the applicable quality decision, a safe boundary and explicit authority. The conjunction describes admission of an action; it is not a learned objective or a probability of success. The numerical decision criterion remains attached to each experiment rather than being replaced by a generic score.

Application to the revised comparisons

The 15 October comparisons completed all declared pairs and passed independent identity, role-mirroring and utility-conservation checks. The utility intervals nevertheless include zero. Complete execution therefore does not by itself satisfy a claim of measured improvement. A later operational decision must retain the experiment result rather than rewrite it as a victory.

Limits of verification

Software checks can establish properties of the tested execution path. They do not establish strategic quality, recovery under every failure or performance in a new environment. Counts of passing checks from the original August edition remain historical verification, not a new October experiment.

References and data

Rhythmlag Lab. Revised paired evaluations, 15 October 2026: E20261015-A and E20261015-B. The data page records units, denominators, completion counts and result digests.

Data and measurement definitions