Threat features: interim status of the small form D16

Threat features tell the evaluation network which pieces are currently attacked. This post presents the interim status: a small form called D16, its speed measurement and the state of the preparation. The status is that of 2026-10-28, about 08:00 UTC. As of that date there was neither a trained D16 network nor a playing-strength result for D16.

Why earlier forms failed

A new form must pass a speed gate before training: it may slow the search down by at most a share fixed in advance. Such limits applied to the three earlier forms and to D16. Three earlier forms of threat features failed this gate on 2026-10-27.

• Speed loss means: the number of search nodes per second falls by this percentage compared with the build without the feature.
• Another, larger threat kernel was rebuilt with AVX-512 instructions after a review report. It still costs 42.2 %, down from 88 %.

What D16 is

The small form D16 was selected on 2026-10-28 after a consultation in which three model families independently proposed the same direction: only attackers without line action, a narrow additional store instead of wide threat rows, and the calculation only at evaluation time, from the bitboards.

• What counts: own knights and bishops count when an enemy pawn attacks them. Own rooks and queens count when an enemy pawn or knight attacks them. Attacks along lines (bishop, rook, queen) do not count. Pieces of both colours are recorded; a piece attacked several times counts once.
• Size: 2 colours × 4 piece types × 64 squares = 512 feature rows, each with 16 channels per side.
• Connection to the network: the 16 channels of both sides enter the second layer of the network as 32 additional inputs.
• Start of training: only the weights with which the new inputs act in the second layer start at zero. At step 0 the network therefore computes exactly like the control without D16. A check on the CPU showed a non-zero gradient for these weights at step 0. After the connection into the second layer was activated by hand, the D16 input weights also received a gradient.

Gate 0: the speed measurement

It was measured on 2026-10-28 from 06:05 to 06:14 UTC on a measurement machine of our own, with one core, 64 positions of 100,000 nodes each and 15 pairs per comparison in alternating order. Four builds of the same engine were compared.

• Limit: the upper 95 % limit of the loss from K to B could be at most 5 %. Measured were 2.93 %, upper limit 3.22 %. Gate 0 is passed.
• Reading: this is a pure run-time measurement, not a statement about playing strength. The 2.93 % is not a promised value for later, trained builds.

State of the preparation

• Trainer: finished and checked. It was implemented in four rounds; another model family reviewed it after round 1 and round 3, and rounds 2 and 4 implement its requirements. The acceptor reproduced the trainer in a copy of its own: acceptance without deviation.
• Start packages for the training machines:
  • Round 1 was incomplete: the pilot rejected the real control files. The finding is documented and fixed in the trainer.
  • Round 2 is finished and reproduced. The review blocked it because plan and package by then named different machines.
  • Round 3 was in progress at the time of this post.
• Plan: the advance plan exists in version 3 and is not yet sealed. Two model families reviewed it in two rounds: one released it with conditions both times, the other blocked it twice but did not ask for a third round in round 2. All text conditions of both rounds have been adopted.
• Probe builds: probe builds of the engine with and without D16 on the main version are available. According to the delivery report they match an integer reference on three calculation paths without deviation. These are probe builds.

What does not exist yet

• No D16 network has been trained; no training has started.
• There is no result on the playing strength of D16. Whether D16 makes the engine stronger is open.

How this was checked

• Sources: the decision of the consultation, the advance plan in version 3, the measurement records of gate 0 and of the earlier speed gate, the advance plan of the earlier forms with their limits, the reports and reviews of trainer, start packages and probe builds, and the interim status of the acceptor.
• Recalculation: a script re-reads the speed losses and upper limits from the measurement records and checks feature, size, limits and review states against the plans and reports.
• Independent review: an agent of another model family checked text, tables and chart in 2 rounds against the same sources and recalculated the speed losses and upper limits itself from the individual measurement pairs. It confirmed all numbers and identified corrections, among them the description of the measurement build B and the evidence on the start of learning; they have been worked in.