Run T-0247: a network with threat features passes

Update of 10 November 2026: By Coherent 0.2.27 this post means the register predecessor, the build of T-0238. The candidate T-0247 carries the same version number but is a different build; see "The base" below.

On 08.11.2026 the run T-0247 passed both stages of the oracle. The candidate replaces the network of Coherent 0.2.27 with a network with threat features and adds two changes for speed. T-0247 is the new measurement of a candidate whose first test, T-0245, was cancelled; the post "Epoch 16: the network class is withdrawn and a run is cancelled" describes the cancellation and the joint error bound of both tests. The stages and decision rules are described in the post "Epoch 14: node count before bestmove and a rule for tempo bundles".

The result

Both stages crossed their upper LLRLLR Log-likelihood ratio, the running evidence of an SPRT. It starts at 0; a stage passes at the upper bound and fails at the lower one. bound, after 407 and 342 pairs respectively. 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. and nElonElo Normalised Elo: relates the mean score excess of a colour-swapped pair to the spread of the pair results. values are estimates from the stopped samples against Coherent 0.2.27. The ± values are nominal approximate 95 % error half-widths calculated from the final counts, without adjustment for sequential stopping.

What the candidate changes

The network. The candidate adds 86,808 threat rows to the main network. The post "Inside the teacher: how the network of width 1536 is built" describes the shared feature form and the value cache. In this candidate, 194 threat rows use a 16-bit side table.

• Size (calculated): 71,413,073 parameters in the exported network, of which 66,668,544 are in the threat rows.
• Training: the start was the main network, with the threat rows at zero. Training comprised 4,048,986 steps at a batch size of 4,096 on 16,584,642,797 rows from two lists of the teacher's data. The output scale is 137; in the main network it is 153.
• Files: the large first-layer weight tables are embedded from three binary files; head weights and small constants remain in the source code. The three files are declared with checksum and size, totalling 209,792,884 bytes.

Two changes for speed.

• Value cache: the engine also prefetches the slot of a child position in the value cache.
• Correction entries in advance: before the network evaluation, the engine requests the 6 entries of the evaluation correction from memory.

At fixed parameters, both speed changes leave the evaluation values unchanged. The value cache and threat-row lists together account for 557,120 bytes of the hash budget; the larger accumulator stack also reduces the budget available to the transposition table. With the threat network, Hash settings below 8 MiB are treated as 8 MiB. In the bench test, the complete candidate visits 713 nodes, its predecessor 603. Search and time management are otherwise unchanged.

The base. The candidate was built on version 0.2.26. The line with the node count before bestmove, which T-0238 added, is therefore not in it. Its version file reads 0.2.27, the same number as its predecessor. On 08.11.2026 the pipeline took the candidate over into the main branch; since then the engine there is that of the candidate, and the line is missing there as well.

Related posts

The networks of the project are described in the post "The networks of Coherent: main network, teacher and cascade"; smaller forms of threat features are described in "Threat features: interim status of the small form D16".

What this post does not say

The result applies to the two documented stages of the oracle against Coherent 0.2.27. How much of the gain comes from the network and how much from the two speed changes, the oracle did not measure separately.

How this was checked

• Sources: oracle register, export of 08.11.2026, entries T-0238, T-0241, T-0245 and T-0247; the code change of the candidate against its predecessor without the weight data; the network report of the candidate on start weights, data and recipe.
• Recalculation: a script recomputes Elo, nElo and LLR of both stages from the pair counts, the parameter count, file sizes and table sizes, and checks the numbers of the text in both languages.
• Independent check: text, metadata, figure and both language versions were checked independently against the same sources.