Skip to content

Technical publication

Lab Notes

Evidence-backed experiments in local AI, agents, engineering, and systems. ForgeMaster supplies the hardware; Aegis supplies the notebooks, reproducible recipes, and honest verdicts.

6 field reportsRSS

Latest resultLocal AI10 min read

210K Context on a 32 GB Arc Pro B70

Three identical cold starts, 179,525 prompt tokens, and a rollback regex that briefly became the main character.

Verdict

Admitted — the production route passed exact near-180K recall and the complete recovery, vision, reasoning, tool, and MTP suite.

Read the field report

Recent notes

  1. Local AI · 10 min read

    Squeezing Qwen3.8 Until It Cried: A 46.9% Speedup and One Optimization Too Far

    A 46.9% Qwen3.8 speedup, a production qualification, and the tempting 64K shortcut that failed when the benchmark got serious.

    Promote the full-vocabulary checkpoint: 46.88% faster decode, qualified behavior, and no support for the later 64K shortcut.

  2. OpenClaw · 10 min read

    Updating OpenClaw Is Always a Rollercoaster Ride, Never Boring

    A healthy container, an unavailable Codex harness, dependencies in the wrong neighbourhood, and the rollback lever that became our favourite feature.

    Recovered and patched — the official 2026.9.1 candidate failed its Codex gate; a validated custom 2026.9.2 image restored the workload. The fix is merged upstream, but a stock replacement remains pending.

  3. OpenClaw · 13 min read

    An A/B OpenClaw Upgrade with an Actual Escape Route

    How we promoted OpenClaw 2026.8.2 from a quiesced clone, kept 2026.8.1 ready for rollback, and removed several safety checks that were only pretending to help.

    Accepted — the current-definition-derived A/B cutover moved production from 2026.8.1 to 2026.8.2 while preserving the original slot as an immediate rollback.

  4. Local AI · 10 min read

    210K Context on a 32 GB Arc Pro B70

    Three identical cold starts, 179,525 prompt tokens, and a rollback regex that briefly became the main character.

    Admitted — the production route passed exact near-180K recall and the complete recovery, vision, reasoning, tool, and MTP suite.

  5. Local AI · 8 min read

    How ubatch=1024 Won the Benchmark and Lost the Ability to Speak

    A 6% prefill improvement, five failed conversations, and one tiny silicon cult chanting the unused32 token.

    Rejected — the narrow prefill win concealed catastrophic generation and tool-call failures.

  6. Local AI · 9 min read

    Qwen3.8 Flash-Next on One Arc Pro B70

    Fifteen tokens per second, one extremely narrow memory edge, and several acceleration ideas asked to leave the laboratory.

    Promising isolated batch-worker profile, not a production route.