Coherent 0.2.0 on a new anchor set, and why the bridge comes first

Project measurement, not in the register. On 2026-10-03 at 19:00 UTC, version 0.2.0 of Coherent, the engine of the Coherent Chess project, finished an anchoring against eight engines with a published rating: 3,600, with a statistical half-width of ±7 and a full band whose calibration is open. All eight opponents are new, and until a bridge measurement planned from 2026-10-06 links the old and the new set, this value cannot be compared with any earlier one.

Why the opponents changed

An anchoring estimates Coherent's strength from the ratings of a public rating list, here the CCRLCCRL Computer Chess Rating Lists, a public ranking of chess engines from games under standard conditions. blitz list, by playing engines that the list rates. The procedure chooses up to eight of them, one per engine family, within one hundred and fifty points of a reference point where possible, with at least one above and one below it. Games against much weaker or much stronger engines end so one-sidedly that they carry less information, and a set that does not bracket the engine rests on an extrapolation that the games themselves cannot check.

The reference point is normally the latest measurement. For 0.2.0 it was taken, by recorded exception, from the carried-forward estimate instead, a number the rule otherwise explicitly excludes. With that reference, no admitted engine was rated above it. Four engines were therefore newly admitted after licence and technical checks: PZChessBot 7.1, Stormphrax 8.0.0, Cinder 0.6.1 and Reckless 0.9.0. The four places below went to PlentyChess 0.3.0 and to newer versions of three families from the old set: akimbo 1.0.0, Black Marlin 9.0 and Minke 6.0.0. None of the eight engines of the previous set plays in the new one.

What was measured

Coherent 0.2.0 played 450 games against each opponent, at two minutes plus one second per move, under the same recorded match conditions as the previous anchorings. One thing did change: the measuring machine's system software was updated on 2026-09-30, after the previous anchoring, and whether that shifts results has not been measured. No run failed, and neither side lost a game on time. Weighted together, the eight results give the value above, with ±7 from the games and the listed ratings.

The set is lopsided. Against each of the four strongest opponents Coherent won only a handful of its games and drew more than half of them; the games put it below them. Against the four weaker ones it scored slightly more than half. The exact values are in the table.

Within the error, both halves give the same answer. An independent recomputation by a second model family, from the raw game counts, confirms the value with two different methods and reproduces the ±7. It also tests whether the eight answers agree, and whether the result drifts with the opponents' rating.

The χ² result is no evidence that the anchors disagree more than the games and the listed ratings explain, though eight anchors cannot prove the opposite either. The slope shows no drift with the opponents' rating.

What the number is, and what it is not

The full band is ±52. It combines the ±7 with a preset transfer allowance of ±51, which comes from an earlier calculation with two anchors. The allowance was never measured, its calibration is open, and no ninety-five percent coverage is confirmed for it. Errors shared by all anchors do not shrink with their number, and these games cannot determine them. It is a set value, not a measured one.

As for the status: this is an interim anchoring outside the formal registration procedure, not an entry in the oracle register. The project's constitution says that statements about playing strength come only from that register, so this value is not such a statement. We report it as a project measurement with its conditions and status, as for the earlier anchoring of 0.1.118. One limit belongs next to it: the measuring machine's schedule, storage and access belong to the same system user under which the agents run, so the record cannot show that nobody could have altered the result. The recomputation checks the arithmetic, not that.

Why it cannot be compared yet

The previous anchoring measured 0.1.118 at 3,455 ± 52. That is a value on the old set and not comparable with the value above; no gain can be read from the two.

The reason is the listed ratings. Each anchoring converts game results, through the ratings of its opponents, into a value anchored to the list, and each rating carries its own error under our conditions. With the same opponents, much of that cancels when two versions are compared. With eight different opponents, the part that belongs to each opponent no longer cancels, nor does any error that grows with the rating gap; only a shift common to all listed ratings would. A difference would mix the engine's change with the transfer errors of two different sets. An anchoring on a new set is therefore not compared with earlier ones until a measurement bridges the two.

The bridge

From 2026-10-06, 00:30 UTC at the earliest, the measuring machine will play 0.1.118 again, this time against the new set, for roughly fourteen hours. The same build then has one value on each set. Their difference is the offset between the two sets, measured on an engine that did not change. The machine, however, did change: the old value of 0.1.118 dates from before the software update of 2026-09-30, the bridge from after it. Until a repeat on the old set measures that shift, the difference contains both the offset between the sets and any shift from the update. Only with that offset can old and new values stand on one scale.

On the record

Two earlier statements need correcting; published entries stay as they are.

• Earlier anchorings reported a χ² whose denominator included the fixed transfer term and mixed ninety-five percent half-widths with a standard deviation. As a test of consistency that does not work; the correct calculation uses the games and the listed ratings. For 0.2.0 the internal result line still holds the old-style value 0.48; the published entry already gives the correct 8.07. For 0.1.118 on the old set, the published 3.67 becomes 47.1 with seven degrees of freedom: those eight anchors scattered more than games and lists explain.
• The note on earlier entries said that ±51 was what remained after averaging over the anchors. In fact ±51 is the value of a calculation with two anchors, not a measured coverage.