Coherent 0.1.100: 2989 ± 58 over three anchors

Project measurement, not in the register. The figures come from the oracle's data delivery (entry E-0005E-NNNN One Elo measurement: a build anchored against foreign engines with a listed rating, on a measuring machine of its own. Numbered in the order delivered; the register run it measured is named with it. E-0001 to E-0012; from the thirteenth entry on the prefix is M, the count running on.); the build measured is the one from run T-0092 of 19 September 2026. The reservations below are those delivered by the day of publication; the full and current account is on the playing-strength page.

Coherent 0.1.100 played 720 games against three foreign engines with a published rating, Velvet 1.1.0, 4ku 2.0 and Renegade 0.12.0, 240 against each at 120+1. The result is 2989 ± 58 (anchored to CCRLCCRL Computer Chess Rating Lists, a public ranking of chess engines from games under standard conditions. blitz ratings), ± 27 of it statistical. From 22 September 2026 that is the figure at the head of this site; it replaces the anchoring of 0.1.93 at 2891 ± 59.

The runs

All half-widths at the 95 % level, as the delivery states them; the column From it is the anchor plus the measured difference. The listed ratings are from the CCRL Blitz Rating List as of 2026-09-12.

How the figure is formed

Recalculated from the anchors: the three anchored values, averaged with weights inverse to their variance, give 2989.1 with a statistical half-width of ± 26.4; the delivery rounds it to ± 27. With the measuring side's transfer term the band is ± 57.4; the delivery names 2989 ± 58, and this site publishes the delivery's figure. The anchored values are compatible with one another (chi-square 1.46 at 2 degrees of freedom, bound 5.99; half-widths read as 95 % half-widths).

Against 0.1.93 two anchors were played in both campaigns, Velvet 1.1.0 and 4ku 2.0. Per anchor the difference of the two measured differences is +107.37 ± 65.40 and +34.83 ± 61.50; combined, +69 ± 45. The per-anchor values are compatible with one another (z = 1.58). That is the measured change: the opponent's listed rating and the transfer term cancel in it. The 98 points between the two headline figures are not a measured gain; they carry both bands in full.

What belongs to this figure

1. Measured outside the formal registration procedure. Every result line of this measurement path is recorded twice, independently, with its checksum.
2. The calibration of the total band is not confirmed from the complete raw data.
3. The evidence for the scale mapping this book one-to-one to the list at 2+1 comes from a round robin with three openings per pairing in ten colour-swapped rounds. The clustering is not taken into account in the supplied calculation (intraclass correlation about 0.30, factor 1.93), so the evidence is about half as strong as reported. This is a reservation regarding the model assumption, not the error calculation: the anchoring runs had a distinct opening for each colour pair (120 of 120 per anchor), leaving the statistical half-width of 26.4 unaffected.
4. Anchor reservation: 2 of 3 anchors (Velvet 1.1.0, 4ku 2.0) lie more than 100 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. below the version measured.
5. Addendum (2026-09-21): The measuring machine is virtualised, text corrected; values and conditions of this measurement unchanged; effect of virtualisation not measured.
6. Statistical half-width plus transfer uncertainty. A round robin among the five anchors showed that a single published rating differs under external conditions; averaged over three anchors, ± 51 of this remains at the 95-percent level, which, together with the ± 26.4 from games and list errors, gives ± 58.

The anchor binaries are documented in the delivery as the authors' own builds: Velvet 1.1.0 (GPL-3.0-or-later), 4ku 2.0 (MIT) and Renegade 0.12.0 (MIT).

Conditions

120+1, book 8moves_v3 with 34,700 positions, each colour pair with an opening of its own, 64 MB hash, one thread, four concurrent games, no adjudication, a fixed number of games without SPRTSPRT Sequential probability ratio test. Games are played until the accumulated evidence reaches one of two bounds; then the test stops.; 0 losses on time. The measurement is dated 21 September 2026, 10:40 UTC.

Generated from the data delivery. The recalculation is the site's; every other figure is the measuring side's.