Coherent 0.1.62 on three anchors, and the subtraction that says nothing

Project measurement, not in the register. Every figure is taken from the anchoring delivery of 15 September 2026 and from the oracle's register.

Coherent 0.1.62 played 720 games against 3 foreign engines that carry a published rating, at 2 minutes plus 1 second per move — the thinking time those ratings hold for. 240 games per anchor, 120 colour-swapped pairs each with an opening of its own, 0 losses on time.

The listed ratings of the 3 opponents are 253 points apart. The 3 answers they give lie within 56 points of each other. Weighted together: 2608, statistical half-width ± 23. The figure we publish is 2608 ± 56.

The subtraction that says nothing

The previous figure was 2570 ± 55 for 0.1.61. 2608 minus 2570 is 38 points, and those 38 points are not a measurement.

The two campaigns hung on different opponents: Leorik and Chal then, 4ku, Chal and Velvet now. In the difference between two figures built that way, the share contributed by the foreign listed ratings does not cancel out — you are subtracting two transfer errors that have nothing to do with each other. How much the engine has gained since 0.1.61 is therefore not what these two numbers say. That would need the same opponent in both campaigns.

The measuring side sent the same warning along, unprompted and before we asked.

Why ± 56 and not ± 23

The ± 23 is the error from the games and the listed ratings. The published band carries, on top of that, what a round robin between foreign engines showed about how a listed rating transfers to our conditions, and it is explicitly declared a judgement: its calibration is not confirmed from the complete raw data.

We have a question outstanding about that calibration. The transfer share of ± 51 is the figure that follows from averaging over two anchors; over three, the same scatter would be smaller. Until that is settled we publish the wider figure. Too wide is the forgivable direction.

A second reservation belongs to Velvet 1.1.0: there is no provenance or licence record for the binary that was measured. That it is the author's own build is therefore not established. The risk is one-sided — a weaker build than the official one would make our figure look too high. The anchor stays declared as measured, because the games are measured; the reservation belongs to the judgement character of the band.

The register has moved on

What was anchored is the build of run T-0074, of 13 September. The oracle has kept working since; the newest entry is T-0083.

Carry the long stages passed since then onto the anchored figure and you arrive at 2842 ± 65. That number is on the run page, and it stands there as what it is: carried forward, not measured. The build of T-0083 has not played a single listed opponent. The chain measures against our own predecessors, at a different time control and a different book, and a sequential test that stops at its upper bound reports a gain that is biased upward. It leans high rather than low.

What it does say: the next anchoring should land above 2608. By how much is not for this arithmetic to decide, but for the next measurement against opponents somebody else built and rated.

All of the figures are on the playing-strength page.

The engines 4ku 2.0, Chal 1.3.2 and Velvet 1.1.0 are named by word mark and version because the measurement conditions require it; no connection to their authors and no endorsement by them exists.