Abstract

We distinguish worker scheduling time, represented-state coverage, action agreement and paired utility. This revision preserves the measurement definitions of the original systems note and connects them to the October evaluation records. Resource quantities and empirical performance are reported independently.

Mathematical definitions

 = (1/N) ∑ᵢ 1[aᵢ = aᵢ,reference]
Exact-action agreement on N evaluated decisions.
r_completion = N_completion / N_decisions
Completion frequency uses observed decisions as its denominator.

Measurement classes

Installed capacity is an inventory quantity. Worker time is accumulated elapsed assignment time. Indexed-state counts describe a representation. Action agreement compares observations on a fixed corpus. Paired utility compares outcomes under a declared evaluation procedure. None of these units can replace another without changing the scientific question.

Historical computational context

The August manuscript reported approximately 8.21 thousand cumulative worker-hours over its M1–M10 campaign. This is an archived scheduling quantity, not the current total. It includes accepted assignment intervals and is not a measurement of full-core occupancy. No new hardware inventory or current accumulated training time is inferred here.

Agreement on a fixed corpus

For N evaluated decisions, exact agreement is the fraction whose actions match the recorded reference. Category agreement applies the original grouping before comparison. The historical studies contain 15,169 and 11,000 decisions. Their exact agreement is 65.76% and 51.29%; category agreement is 73.87% and 62.15%. The extract does not contain record-level observations for estimating additional uncertainty.

October evaluation accounting

Each completed October comparison contains 60,000 scenario pairs, not 120,000 independent observations. Action counts are larger and depend on the policies’ visited histories. Passive completion counts are reported with their decision denominators. They describe missing-policy handling, not an independent performance metric or the fraction of the full state space covered.

Reporting requirements

A numerical result should retain its date, unit, denominator and reference version. Historical and current records stay separate. The public data page gives the completed October comparisons and the earlier agreement extract in separate tables and CSV files. Neither infrastructure capacity nor a larger sample count is presented as evidence of strategic superiority.

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