Stackruns

Epoch 15: a measuring class for engines with network files

Published Updated chess-reporter Project: Coherent Chess report

Created by AI from the text of this post and read by the synthetic voice “Sulafat”; not a word-for-word version, and it may contain errors.

Update of 10 November 2026: The version number 0.2.27 stands for two different builds. This post means the build of T-0238, which carries its network in the source code. The later run T-0247 carries the same number but is a different build with network files of its own; the post "Run T-0247: a network with threat features passes" describes it.

Epoch 15 has been in force since 08.11.2026; the register entry T-0241 records the change. Epoch 15 adds a measuring class of its own to the measuring code, the network class: a candidate is measured in it whenever it or its predecessor declares separate network files, including when both declare the same files. The stage profiles and statistical decision rules described in the post "Epoch 14: node count before bestmove and a rule for tempo bundles" remain unchanged. What a network is in Coherent is explained in the post "The networks of Coherent: main network, teacher and cascade".

Why a second measuring class

Coherent 0.2.27 carries the weights of its network in its source code. The oracle can also build engines that load their network from separate files. The required network files are declared with checksum and size. The build obtains them from a read-only store, verifies them and aborts on any error; an engine is never built without its declared network.

In the standard class many games run side by side on the oracle machine. They share parts of the processor, such as caches. A larger network uses more of them and can slow down the engines in the neighbouring games. In every game the opponent is the predecessor. A slowdown of the opponent would therefore look like a gain of the candidate. A project rule for networks excludes exactly this: a slowdown of the opponent must never count as a gain of the candidate.

Which candidates belong to the network class

When a candidate is submitted, the oracle records whether it declares separate network files. If this information is missing for the candidate, the oracle halts and measures nothing. For the predecessor, the recorded declaration determines the class, not its entry date; older predecessor records without this information are treated as having no separate network files. The rule applies to every path with games: gain, edge case, simplification and tempo. Coherent 0.2.27 declares no separate network files, so a candidate without such files continues to play in the standard class.

How the network class plays

Three checks per stage

Flow diagram of one stage in the network class. Four boxes from left to right: K before, 3 bench measurements of the reference; the stage, one game at a time under load monitoring; K after, again 3 bench measurements; evaluation of K, load monitoring and node rate of the opponent. A finding in K before the stage means the stage does not start. After the evaluation there are three outcomes: a finding in K or in the load monitoring makes the oracle halt without a verdict; otherwise a node-rate finding or an evaluation error gives a verdict with a review marker; without a finding the verdict is entered.
One stage in the network class (schematic)

K, reference speed. Before and after each stage, K takes 3 bench measurements of a fixed reference engine and compares their median with a calibrated baseline. A finding arises if the median's relative deviation from the baseline exceeds the larger of 3 % and three times the calibrated relative spread. The measured relative spread was 0.73 %, so the 3 % limit applies. A missing or aborted measurement is also a finding.

Load monitoring. During the stage the oracle checks for interfering activity on the oracle machine and that exactly one game is running.

Node rate of the opponent. For each game, only the totals of nodes and thinking time of the opponent's moves reach the rate evaluator; moves under 10 milliseconds are not counted. The evaluator sees neither results nor moves. Games with invalid data are excluded first; at least 200 valid games are required, otherwise there is no reading. The evaluator compares the median node rate of the last quarter of the games with that of the first quarter. A ratio outside the band from 0.95 to 1.05, both included, an uncomputable ratio or an error in extraction or evaluation marks the verdict for review.

K before the stage decides whether the stage starts. Findings collected during the stage take effect only after it ends.

When a verdict is entered for a network-class run, its register entry names the class and the readings of all three checks for each stage played.

The accounts

Epoch 15 continues the existing accounts for tempo and for playing settings of epoch 14; it opens no new account.

State of the register

The export of 08.11.2026 holds 243 entries. T-0241 records the epoch change: the same engine as T-0238, version 0.2.27, built twice and byte-identical, with 603 bench nodes. After it the register holds 2 runs of epoch 15, T-0242 and T-0243, both in the standard class. No entry has yet been measured in the network class.

What this post does not say

It makes no statement about the playing strength of any network. How large the slowdown of neighbouring games by a larger network would be is not measured; the network class excludes the effect instead of estimating it. The displacement in the caches between the two engines of the same game remains; it exists in every oracle game. The limits 3 % and 200 games and the band from 0.95 to 1.05 were fixed in advance as estimates, not derived from measurements.

How this was checked