Skip to content

Keying and replay experiments: no gain found, one effect measured

Five preregistered experiments on remembering and keying dead ends found no gain that matters for a full solution. Three were killed (difference-map bands, cut-keyed replay, keying with five clues) and two were kept as measurements (quota keying, and the quota curve with five clues). The kept quota result is a rate effect of 53 times, about 5.7 bits of rate, which is not a smaller tree.

by Raphaël Anjou

experimentmeasurementmeasuredBacktrackingReduce the searchUpdated 2026-11-14
Reproduceseeded: reproduces with the given seedreruns the searchBudget: 5.67 thread-hours for the five-clue quota experiment (5.06 test, 0.52 calibration, 0.09 gates); 6.81 core-hours for the difference-map test grid; 16.1 thread-hours for quota keying; 20.6 for five-clue keying; 2.13 for cut-keyed replay
Hardware & run
Native runCPU only
Cores
12
Threads
12
CPU
Single-threaded runs spread over a 14-core low-power x86 laptop (12 threads), an 8-core Apple M1 (7 threads) and an Apple laptop (4 threads)
Machine
three laptops, runs pinned to one thread each
Budget
5.67 thread-hours for the five-clue quota experiment (5.06 test, 0.52 calibration, 0.09 gates); 6.81 core-hours for the difference-map test grid; 16.1 thread-hours for quota keying; 20.6 for five-clue keying; 2.13 for cut-keyed replay
Start
seeded and deterministic per row; wall-clock seconds vary by host
Reproduce this result

seeded: reproduces with the given seed

just experiments five-clue-quota test 4 local

Runs the local share (openings 514 to 516) of the kept five-clue quota experiment, 1e9 nodes per run; the full 16 openings need the three seed groups, and every row reproduces node for node, only seconds differ.

Code & data on GitHub

Keep exploring

Referenced by

Page sourceView as Markdown