Coherent 0.1.93: 2891 — and a number that did not move

Project measurement, not in the register. The figures come from the oracle's data delivery (entry E-0004E-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.) and from the measuring side's report channel; the build measured is the one from run T-0087. The full account is on the playing-strength page.

Coherent 0.1.93 played the same three opponents as 0.1.62 and 0.1.91 before it: Chal 1.3.2, Velvet 1.1.0 and 4ku 2.0, 240 games each at 2 minutes plus 1 second, the same binaries, the same machine, the same book. The result is 2891 ± 59 (anchored to CCRLCCRL Computer Chess Rating Lists, a public ranking of chess engines from games under standard conditions. blitz ratings), ± 29 of it statistical. Since 21 September that is the figure in the head of this site; it replaces the anchoring of 0.1.62 at 2608 ± 56.

Against 0.1.62 that is a gain of +282.51 ± 33.41 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., formed per anchor so that the listed rating and the transfer error cancel out.

The number that did not move

Three days ago 0.1.91 stood at 2890.31. 0.1.93 stands at 2891.10. Two versions on, 0.8 points apart. Computed as a change against 0.1.62: +279.30 ± 34.33 then, +282.51 ± 33.41 now — the same number inside the band.

After the jump of almost 280 Elo between 0.1.62 and 0.1.91, that is this campaign's finding: no further jump. The internal scale had seen only +68 for the step from 0.1.90 to 0.1.91; that the foreign anchors then see hardly anything more fits.

The anchors disagree more than before

The three anchored values span 141 points; for 0.1.91 it was 82. The measuring side computes a chi-square of 3.10 on its own mixed-scale quantity and satisfies its release rule with it. In uniform standard deviations it is 5.77 at two degrees of freedom, a chance probability of about 0.06: compatible, but only just. The measurement became less certain while the result stood still.

The measuring side withdrew one of its own statements along the way. The critical value of 5.99 it had named on 18 September holds only for a chi-square in uniform units — not for its mixed quantity. Both now stand side by side on the playing-strength page.

An arithmetic dispute, resolved

Across two campaigns our recalculation of the point estimate sat two to three points above the measuring side's, always in the same direction. The cause is found: their weights carry a transfer term T = 36.9 per anchor, ours do not. The term evens the weights out and moves the mean towards the unweighted one — without T the same anchors give 2892.80, with T 2890.31. They compute the statistical half-width without T, which is why our ± 29 agrees exactly.

The same explains the chi-square values: their 0.816 for 0.1.91 is in mixed units, our 1.24 and 4.75 leave T out. Two differently defined quantities, no error on either side. And our difference calculation is, word for word, their own method B.

What belongs to the figure

The measurement ran on the dedicated measuring machine by the same procedure as every anchoring before it, but outside the measuring side's formal registration procedure: schedule, storage and access belong to the same account as the agent sessions. Since 19 September a log stands in for that missing separation of users — every result line is recorded with its checksum in two separate ticket systems. Checking authenticity therefore rests with the measuring side; this site carries its figures and its reservations.

Of those it names four: the registration, as described; the calibration of the total band is not confirmed from the complete raw data; the evidence that the scale maps one to one onto the list is about half as strong as reported, because of the grouping by openings; and the set of anchors is the starting set, not a ladder — all three lie more than 100 Elo below the version measured.

The link to the register now sits in the entry itself. 0.1.93 is the binary from run T-0087, byte for byte; the register carries commits and binaries, not version numbers, which is why the number does not appear there.

The next version is already being measured. Its entry carries status and log from the start.