Categories
Assessment engine, report pipeline, infrastructure — the choices we made and why.
Round 4 of Qwen3.8-27B MLX 4bit (17GB): restart only the server, rerun the 43 failures. Two recoveries reach pass@4 48/89 and the 13 round-3 ERRs drop to 1 — the server was never the cause.
Round 3 of Qwen3.8-27B MLX 4bit (17GB): re-running the 43 round-2 failures recovers zero, pass@3 stays 46/89. 30 failed exactly as before; 13 never ran a turn — and the real cause was not the server.
Round 2 of Qwen3.8-27B MLX 4bit (17GB): re-running the 54 round-1 failures reaches pass@2 46/89 (51.7%). 11 recoveries (median 6.5 min), two solved only by 27B, and why retrying failures is slower.
Qwen3.8-27B MLX 4bit (17GB) on Terminal-Bench 2.1, same harness as Flash Next (100GB): 35/89 (39.3%) vs stock 45 and Uncensored 41. Which tasks split, 15 repetition loops, speed, ops notes
Round-5 (pass@5) results for the abliterated Uncensored Qwen3.8 Flash Next on Terminal-Bench 2.1, plus a series wrap-up: 59/89 vs stock 64/89, the 13 tasks that split them, speed, and harness failures
Round-5 (pass@5) Terminal-Bench 2.1 results for stock Qwen3.8 Flash Next mixed-4-8bit: 4 of 29 retries recovered (two timeout-with-the-right-file), a 120K context handoff and a fallback loop
Round-4 (pass@4) Terminal-Bench 2.1 results for the abliterated Uncensored Qwen3.8 Flash Next: 2 of 33 retries recovered, two fallback loops worth 3,877 HTTP 400s, and a 40-minute Coq build
Round-3 (pass@3) Terminal-Bench 2.1 results for stock Qwen3.8 Flash Next mixed-4-8bit, a like-for-like three-trial comparison with the abliterated build, and why 31 tasks still fail
Round-4 (pass@4) Terminal-Bench 2.1 results for stock Qwen3.8 Flash Next mixed-4-8bit: 2 recoveries out of 31 retries, and why 29 still fail — repetition loops, premature submits, near misses
A same-conditions single-pass A/B between an abliterated (Uncensored) model and the stock mixed-4-8bit build across all 89 Terminal-Bench 2.1 tasks
Three days of Terminal-Bench 2.1 on a community abliterated derivative. Four rounds of context, timeout and temperature changes, 13 server wedges, and what a final 56/89 actually means.
Measuring how much real terminal work a local LLM on an M5 Max can actually complete. The pass rate was less interesting than the server wedging four times and recovering itself every time.
Measuring Qwen3.8 27B at 4bit and 8bit against Flash Next, then adding DFlash2 speculative decoding. It was 2–3x faster, but contrary to the official claim the output diverged from the original on M5.
Running a 125B MoE model on one Mac as an agent backend. I assumed the model was slow. A day of digging showed the real culprit was prefill — and how caching and a few flags fixed it.
PostgreSQL will not delete WAL while a replication slot still holds it. Two months of WAL accumulated after a standby died, and the secondary damage that only surfaced once the space came back.
Modern container images require x86-64-v2. The default CPU model your hypervisor hands a VM does not have those instructions, so the process dies inside glibc before the application ever runs.
If your iptables whitelist only inspects ctstate NEW, the handshake never completes. A port scan still reports the port open, which is why the cause of a failing backup went unfound for a month.
A reload can return rc=0 without applying anything. Two patterns: an old process surviving and serving the previous config, and changes landing in a staging file that never got promoted.
Skipping an LTS step and stalling at the configure stage takes sudo and networking with it. Counting dpkg states tells you whether recovery without a reinstall is possible.
If you alert on a value that does not change while things are working, it will fire eventually — guaranteed. The metric that produced false alarms, and the real alert we silenced while fixing it.
timedatectl changes the OS timezone. Your database and your containers keep the old one. How NOW() came back nine hours apart inside one cluster, and why nothing short of a restart fixes it.