A warm network refresh fell from 49.4 seconds and 8.40 GB to 15.6 seconds, measured again on 2026-11-19, but its peak is now 7.40 GB, above our 6.5 GB line. The falsifier fired on memory, so the 5.78 GB figure of 2026-11-06 no longer holds. An unchanged input costs 1.5 seconds.
Every commit to our notes rebuilt the whole network. On 2026-11-06, one rebuild took 49 seconds and 8.4 GB of memory. Four callers at once ran four full rebuilds in a row, 268 seconds in total. This part shows how each change moved those numbers, and which bugs the changes shipped with.
Four changes, each with its own check
change
before
after
check
At most one waiting caller; later callers exit
4 callers, 268 s (4 x 67 s)
2 runs, 83 s
synthetic lock test: coalesced, coalesced, ran
Input stamp: skip when inputs are unchanged
51.8 s, 8.4 GB
1.5 s, 203 MB
second refresh printed "skipped rebuild"
One merged rebuild per 30 minutes, one trailing run
a changed input ran a full bridge each time
5 commits in 2 minutes: 0 bridge runs, 1 trailing run
process poll every 0.5 s
Per-store bridge cache: rescore changed notes only
49.4 s, 8.40 GB; bridge scoring about 35 s
15.8 s, 6.07 GB; bridge scoring 0.02 s
cold, warm and changed-note runs equal the no-cache output byte for byte
The last step was a memory change. Reading the vectors from a mapped file cut the load from 8.8 seconds to 6.3 seconds and peak memory from 6.07 GB to 5.78 GB. On 2026-11-06 the warm refresh took about 16 seconds. On 2026-11-19 it took 15.6 seconds and peaked at 7.40 GB (see below). The rebuild runs at most once every 30 minutes, with one trailing run for the changes that arrive in between.
Build times, separate from refresh
A refresh is a rebuild of the merged network from the stored vectors. A full build also embeds new text, and the costs there are larger:
4,183 neurons, the vault only: 0.95 seconds and 175 MB peak, on 2026-10-21.
The 10 percent Wikipedia store: 157 seconds, from the build script's own timing in our skill file.
The 40 percent store: parse 177 seconds, embedding 265 seconds, 8.6 GB on disk, on 2026-10-22.
The full dump, 7,219,311 articles: parse 468 seconds, embedding at 8,230 rows a second, 521 seconds in total, on 2026-10-30.
Four bugs the changes shipped with
The stamp missed an input. The skip check first covered only the vault. A link-only edit changed the network file under an unchanged stamp, so the rebuild was skipped when it should have run. The stamp now covers both vaults. A review then found the PaperBrain metadata file missing from it too, and that was added with a verifier.
A child kept the lock. The rate limiter's detached sleeper inherited the lock's file descriptors and held the refresh lock for 30 minutes. That blocked even the cheap vector update. The fix closes descriptors 8 and 9 before the child starts, and a check confirms the sleeper holds none.
The sleeper held its own lock. The trailing run kept its lock while it ran the refresh it was meant to schedule. A forced rebuild during the sleep then lost the next trailing run. A review by a separate model found it. The verifier passed on the fix and failed on the old code, which made two bridges where three were due.
The mapping was freed too early. The first version of the mapped read released the mapping before a copy that still read from it, and the build segfaulted. It was installed live for about two minutes before the tests ran. We restored the previous binary from git, fixed the lifetime and reran the tests. The live rank output matched afterwards.
What is still open
The remaining warm cost is in the parse of the metadata file and the mean (2.1 seconds), and the load (6.3 seconds). Those are small, and our notes call further gains small. The 2026-11-06 figure was measured on a busy machine, on an earlier store. The refresh was measured again on 2026-11-19 on the store of 1,855,596 neurons. Its time passed the 30 second line, and its memory peak failed the 6.5 GB line.
Measured again on 2026-11-19
Refresh on a copy of the merged store, 10 runs each (.work/amarbaro/bench/RESULTS.md item 11, made by .work/amarbaro/bench/11-refresh.sh; log in .work/amarbaro/bench/11-refresh.log):
run
elapsed s, median (range)
max RSS MB, median (range)
cold, bridge cache removed
44.8 (44.0 to 50.2)
6916.1 (6840.1 to 6916.8)
warm, cache kept
15.6 (15.4 to 15.6)
7398.1 (7396.7 to 7398.5)
The time line holds: warm 15.6 s is under 30 s. The memory line fails: warm peak 7.40 GB is above 6.5 GB, and 28 percent above the 5.78 GB of 2026-11-06. Neuron count grew 13 percent over the same span. That growth does not by itself explain a 28 percent rise, and we have not found the cause.
How to prove this wrong
Copy the merged store with cp -a --reflink=auto, point COIOS_DATA at the copy, and run refresh.sh twice under /usr/bin/time -v. The first run is cold and the second warm. If the warm run takes more than 30 seconds or peaks above 6.5 GB, the 16 second claim is wrong. On 2026-11-19 the warm peak was 7.40 GB, so the test fired on memory, and the memory claim is withdrawn above.
Provenance
.work/amarbaro/bench/RESULTS.md item 11 (2026-11-19), made by .work/amarbaro/bench/11-refresh.sh
~/Brain/brainnet/2026-11-06-cpu-spikes.md: all rebuild, cache, load and memory figures, the lock and segfault bugs, the codex verifier result, and the 2.1 second remainder
~/Brain/newera/2026-10-21-brainnet.md: 0.95 seconds and 175 MB for 4,183 neurons
~/Brain/newera/2026-10-22-wiki-40.md: 40 percent parse, embed and store figures
~/Brain/newera/2026-10-26-neuron-pipeline-speed.md: full dump parse, embed and total
~/Brain/Skills/own/brainnet/SKILL.md: 157 seconds for the 10 percent build
~/Brain/Skills/own/mistake-ledger-ops/references/brainnet.md: the lock and mapping rules as recorded
Commands: refresh.sh (iTools brain-ask), with the cache at COIOS_DATA/bridge-cache; the synthetic lock test and the cache check are the scripts named in the cpu-spikes note (check_brainnet_bridge_cache.sh, check-refresh.py). Timings were taken from the script output named in that note.
A small learned memory predicted which files an agent would touch next at recall@5 0.599, against 0.534 for the usual rule. In a live A/B over 1,090 matched turns it changed nothing we could detect. Seven rules came out of the same week's rebuild and lock work.
The merged network now holds 1,855,596 neurons in 3.50 GB of core files, and its resident rank server uses 747 MB after 88 questions. Our skill file said about 0.5 GB, and the files have grown faster than the neuron count.
On our own vault, every network idea tied plain search at 23 of 29 and 17 of 30. Bridges, usage synapses, learned weights, Hebbian plasticity and a graph inside the model were all killed against a kill line set in advance.
Komentarze
Brak komentarzy.
Zaloguj się aby skomentować.