15 networks in 40 days: the lineage of the evaluation

Note of 19 October 2026: all 15 networks in this list were trained on the project's own data; no data from outside the project went into any of them. Where this entry lists teachers as "not documented" or leaves data sources open, that was a gap in the sources read, not a sign of foreign data. A later review has since traced every data file to the project's own engines: "Own data, file by file: where the training data of the 15 networks came from".

Since 2026-09-08 Coherent has evaluated positions with a neural network. Up to 2026-10-18 the oracle adopted a new network onto the main branch 15 times, each time measured against its predecessor. The first of them, T-0072, replaced the hand-written evaluation together with a new integer evaluation core; its values measure that step, not one network against another. This entry assembles the series from the primary sources: structure, data, teachers and oracle values. Where a source is missing, the entry says so. It was written at the operator's request. The 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. values are measurements against the respective predecessor, not external playing strength.

The series

Counted are the 15 distinct networks that lay on the main branch, matched to their passed oracle candidates through the registered commit and the network hash. Runs that only changed search or speed are not in the list. Short means 8+0.08 seconds with Book A, long 40+0.4 seconds with Book B, each against the predecessor; the register records these conditions for all 30 stages.

Width is the number of values per side in the first layer, buckets the number of weight sets chosen by the king's square. From T-0135 on, the first layer is followed by a head of 16 and 32 neurons; from T-0144 on, the board is mirrored when the king stands on files e to h. T-0119, T-0124 and T-0135 bundled the network with speed changes, T-0135 for example with an evaluation cache. Their measured gains belong to the complete candidates, not to the network alone.

Data

The sources count in three ways: total dataset rows, training rows and learning rows. The counts are compared only within one kind. The chart shows the 10 networks whose sources give total dataset rows; they range from 20,875,984 at T-0076 to 13,591,787,474 at T-0189. For campaign 1 the source for T-0085 gives about 226.7 million rows, of them 181,346,139 training rows; the commits for T-0099 and T-0100 give 226,723,955 training rows. A consistent way of counting is therefore not established.

The network inventory of 2026-10-07 counted dataset rows with positions from the test sets: 11,763 in v7j, 16,418 in v8j and 38,970 in v10. Whether each match lay in the training split is not documented for every one. The networks from v7j and v8j had been trained before the test-set rule of 2026-10-06; a separate decision required neither retraining nor re-measurement for them. For the start network of T-0189 the matches were removed before training. A test set created later overlaps with 1 row of its data; this is allowed as an exception, and that test set does not measure the network.

Teachers

A teacher is the engine whose search wrote the evaluations in the data. For v4 the teacher used last is documented, the main engine of the time. For the two fine-tunings the labels came from the engine with the network to be refined, at 20,000 nodes per position. The inventory names v7j, v8j and v10 as the project's own self-play data, but not the engine versions that played them; their teacher is therefore unknown. For the other networks the sources read name no teacher.

Oracle values

All 15 runs passed both stages. The values are logistic Elo estimates from the score of game pairs; ± is half the width of a nominal 95 per cent interval from the normal approximation, converted to Elo, as the match tool fastchess computes it. T-0072 has +222.52 ± 36.23 and +260.44 ± 39.40 Elo; from T-0144 on the estimates lie between +12.41 ± 6.07 and +35.15 ± 10.95 Elo. The stages needed between 222 and 3,278 games; up to T-0120 at most 740 per stage, so the early intervals are wide. The oracle's rules changed between the runs: they belong to epochs 4, 5, 6, 7 and 10. Each value measures one step against the predecessor of the time under the conditions of the time; the values are not a trend on one scale and cannot be added up to a total gain. All runs stopped at the upper bound, so the point estimates may be biased upwards.

An external result recorded outside the oracle register exists for version 0.2.10 with the network from T-0168: 3670 ± 52, measured on 2026-10-14 in 3,584 games at 120+1 seconds against 7 external anchor engines, listed as unregistered. The band is nominal; its calibration is still open.

Gaps

• For v4 only the teacher used last is documented; the earlier histories of teachers and generators are incomplete. For v7j, v8j and v10 the versions of the playing engines are open.
• For campaign 1 a consistent way of counting the rows is not established.
• The dataset comparison of the network inventory of 2026-10-07 covers v7j, v8j and v10; earlier networks lie outside it, and T-0189 is documented in separate review records.

How this was checked

The networks were determined from the history of the weights file on the main branch; the first line of each version names the network's hash. The match to the runs and all Elo values come from the terminal entries of the oracle register; all 30 values and intervals were recomputed from the game-pair counts recorded there. Data and recipes come from the register hypotheses, commit messages, the network inventory and the review records of T-0189, the external value from the anchoring data. Agents of two other model families checked the text against the same sources.