NNUE nets measure a gain against the baseline

Project measurement, not in the register: no register ID, no assignable register epoch, and no oracle confirmation. Logistic EloElo Strength difference to the opponent, estimated from the games, with a 95-percent half-width where the oracle reported one. Never an absolute rating., error half-widths, time control, game counts, win/draw/loss counts, score rates, and time forfeits come from the project's time-control record of 2026-09-04, correction section and result blocks, and from the secured raw outputs of both runs. Baseline identity, opening corpus and concurrency come from the project's measurement record of those runs; the search-thread count is derived there from the engine code and the hash size from the interface driver, not from a run-level protocol. A separate output file of the measuring tool in report format is not available; the sign reversal was checked against the raw outputs. The holdings statement comes from the local holdings check dated 2026-09-05. None of the figures comes from the public data contract.

Project measurement, not in the register: no register ID, no assignable register epoch, and no oracle confirmation. At time control 8+0.08, from each net's viewpoint against the baseline over 1,400 games, e128 measures +68.37 ± 16.24 logistic Elo and e064 measures +55.56 ± 14.83 logistic Elo; the ± figures are the half-widths of nominal ninety-five per cent intervals. The two comparisons are not a direct comparison of the nets.

This is a project measurement, not in the register. It carries no register ID, belongs to no register epoch, and has no oracle confirmation. The figures below do not come from the public data contract.

Method

Two NNUENNUE Efficiently updatable neural network: a small network that evaluates positions and is updated incrementally move by move. variants, e128 and e064, were each paired with the same documented baseline, the build following T-0064, at time control 8+0.08: eight seconds of base time plus an increment of eight hundredths of a second per executed legal move. That the baseline is the same in both pairings is established. Each pairing consisted of 1,400 games. The pairings are recorded under the date 2026-09-04.

Conditions recorded for both pairings: one search thread per engine; 16 MB hash; a project opening corpus; four parallel games per run. Both runs were started at the same time, together up to eight games in parallel. Hardware is not stated in the record used here.

Two of those conditions are derived rather than logged. The thread count is derived from the engine code and the hash size from the interface driver; a run-level protocol of the parameters actually set is not available, and a present-day call that sets the thread count is not used as evidence for a past run. An assignment of the opening corpus to a register book is not established. A historical hash of that corpus is not available either.

Sign convention

The measuring tool reports Elo from the viewpoint of the side named first. In these pairings that side is the baseline, not the net. For the net’s viewpoint the Elo sign is reversed; the error half-width is left unchanged. Positive values mean a gain for the net against the baseline.

This reversal was checked against the secured raw outputs of both runs. A separate output file of the measuring tool in report format is not available; the check used those raw outputs.

An earlier reading treated the tool’s output as the nets’ result. That reading inverts the sign and with it the conclusion. It stated that the nets lose at this time control. That statement was withdrawn in the correction section of the same record.

Results

Project measurement, not in the register: no register ID, no assignable register epoch, and no oracle confirmation. Logistic Elo from each net’s viewpoint against the baseline, the build following T-0064, at 8+0.08, over 1,400 games. The ± figures are the half-widths of nominal ninety-five per cent intervals, in the same unit:

The two comparisons are not a direct comparison of the nets. No direct match between them was played, so the difference between the two figures has not been tested and no ranking follows.

The secured raw outputs report the same pairings from the baseline’s viewpoint: −68.37 ± 16.24 logistic Elo against e128 and −55.56 ± 14.83 logistic Elo against e064.

Game outcomes from the baseline’s viewpoint, 1,400 games each:

The baseline lost 702 of 1,400 games against e128 and 642 of 1,400 games against e064.

In the pairing against e128, exactly one of 1,400 games ended on time; in the pairing against e064, none did. Both counts come from the end-reason lines of the secured raw outputs.

Interpretation

The finding is a net gain at the clock. Under these conditions each tested neural variant gained against the baseline. Evaluation quality and execution speed cannot be separated causally from this pairing.

A holdings check dated 2026-09-05, covering the period since 2026-09-04, found no new completed clock measurements of these nets in the examined local holdings. That statement applies to those holdings. It is not a claim of completeness beyond them.

What this does not say

It does not rank e128 and e064 against each other. There was no direct match between them, so their difference has not been tested.

It does not separate evaluation quality from execution speed.

It does not transfer to other time controls, other hardware, or NNUE in general. Hardware is not stated.

It does not assign a register ID or a register epoch, and it is not an oracle confirmation.

It does not stand in for a register entry.

The intervals are nominal ninety-five per cent intervals. They are not a minimum gain, not a guarantee, not a standard error, and not a simultaneous statement. They are not a guarantee that is fixed at a stopping time.

It does not claim that the runs were free of parallel load. Up to eight games ran at the same time across the two runs.

A run-level log of the engine parameters actually set is not available. A historical hash of the opening corpus is not available.