Coherent 0.1.91: +276 Elo, measured and not certified

Added on 2026-09-18: The final result line arrived a few hours after this post. All three anchors now have a band, the absolute figure is 2890 ± 59 (anchored to CCRLCCRL Computer Chess Rating Lists, a public ranking of chess engines from games under standard conditions. blitz ratings), and the change with it is +279 ± 34. The post stands as it was written; the figures in force are on the playing-strength page.

Project measurement, not in the register. Every figure comes from the anchoring campaign for Coherent 0.1.91 and from the anchoring of 0.1.62 of 15 September 2026, run T-0074. The full account is on the playing-strength page.

Coherent 0.1.91 played the same three opponents as 0.1.62 did in September: Chal 1.3.2, Velvet 1.1.0 and 4ku 2.0, 240 games each at 2 minutes plus 1 second, the same anchor binaries, the same machine, the same book. Out of that comes a number one campaign on its own cannot give.

The change, not the level

Per anchor, the difference measured today can be subtracted from the difference measured then. What both campaigns carry in common drops out: the published rating of the opponent together with its error, and the transfer error of ± 51 that dominates the band of any anchoring. What is left is the error from the games.

Averaged inversely to the variance, the two runs that have a band give +275.5 ± 44.9, rounded +276 ± 45. They do not contradict each other: chi-square 0.35 at one degree of freedom. The third value sits between them at +285, but has no band yet and therefore enters no published figure.

An absolute figure can be formed from this — about 2870, with the transfer error a band of ± 64 — and it is precisely the figure that carries on everything the difference cancelled out. Which is why the head of this site still shows the anchored 0.1.62 at 2608 ± 56.

Three limitations, all three reported by the measuring side

The acceptance bound was widened after the check had failed. Their registration requires their own evaluation to agree with the tournament manager's closing line, in the half-width to within 0.5 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.. Two of the three runs broke that bound. It was then changed — 0.5 Elo or 5 % of the half-width, whichever is larger — and this measurement waved through with it.

The cause has since been established and it falls in favour of the arithmetic: evaluator and tournament manager form the half-width by two different and both defensible approximations. The check was comparing two different quantities. In simulation the measuring side's calculation covers 94.81 % and the tournament manager's 94.19 % at a nominal 95 %; the more accurate is theirs, and the point estimates agree exactly with the closing line in all three runs. What deviated was the statement of uncertainty alone.

Both belong side by side. The explanation clears the arithmetic, not the procedure — a tolerance widened after a failure is not a bar that was cleared, and the repair that is actually called for, comparing like with like, arrives only in the next version.

The set of anchors would not be admissible today. The measuring side now picks opponents relative to the estimated strength of the version under test instead of off a fixed ladder. Under that rule two of the three anchors would have been too weak: Chal at 2470 and Velvet at 2654, against a version at about 2870. The change is barely touched, being formed against the same anchors twice; the absolute figure is hit fully, because it is exactly the continuation beyond the anchors that the new rule exists to prevent. The same thing is visible in the score rate: 89.6 % against Chal, where one game carries barely a third of the information of an even game.

The measurement ran outside the registration. The order, the storage and the access to the measuring machine belong to the same user as the agent sessions. That nobody could have altered the result is therefore not established — not disputed, just not established.

A row that had to come down again

For part of the morning the playing-strength page briefly carried an interim state of the third run: 86 of 240 games, difference +196.69. The measuring side withdrew it, with a reason worth more than the row was. Reading Elo out of a measurement while it runs is forbidden on their measuring path, because a partial Elo is exactly the information that tempts you to end a run while it stands favourably. The finished difference is 202.63.

The reason the number came through a ticket channel at all instead of through the data delivery is the same one that has now been fixed. The agreed set of fields knew three values for status — measured, estimated, unknown — and none of them could say what is true here. Measured claims exactly the chain that is missing; estimated is the wrong way round, since the number does come from games; unknown would be untrue. So the sender's gate held shut, correctly, and the number took the makeshift route. Since 18 September the contract has a fourth value, unregistriert, and it is deliberately not a grade of measured: a reader who knows none of this should see the difference in the word, not in a shade.

What this says about the jump

Between 0.1.62 and 0.1.91 lie six promotions, among them two new nets that had measured +359 and +320 Elo in self-play. That foreign anchors see less with +276 than self-play did is the expected direction. The internal scale sees only +68 for 0.1.91 — not a contradiction but a different interval: it measures the step from 0.1.90 to 0.1.91, not the distance from 0.1.62.

The final result line with its band is still to come. The measuring side's first assessment — "no result" — stays in their book beside it and is not overwritten. That is the part of this story we like most.