WORK LOG · LAST UPDATED 02 OCT 2026
01 / AGENT STACK · 02 OCT SNAPSHOT
Different temperaments, explicit jobs.
Claude
Long-context planning, synthesis, and editorial structure.
Codex
Repository work, implementation, verification, and release checks.
Droid
Independent execution and a second working perspective.
Cursor
Fast local editing when the file and desired change are already known.
02 / OPERATING PRESETS
Each preset fixes the scope and the check.
- FAST PASSKnown source, bounded edit.
Narrow context, explicit files, one focused test.
- AGENTIC PASSTool-heavy, multi-step work.
Full context, stop conditions, and a live evidence ledger.
- ADVERSARIAL PASSRelease review and consequential claims.
A fresh reviewer gets the claim and tries to falsify it.
- LOCAL PRIVATE PASSPrivate review on a checked local runtime.
Confirm the model, memory budget, and a real response before assigning a bounded task. Published benchmarks describe their dated test setup.
03 / OPEN EXPERIMENTS
Experiments still in progress.
Local inference maintenance
On 9 September, the Windows LM Studio engines were updated with older versions retained for rollback. The existing SGLang service answered a real request afterward; generation through the updated LM Studio engines remains untested. Read the maintenance note.
UAE paper-research safeguards
Session boundaries and dated policy changes are now part of the research checks. Earlier records retain their original collection rules. These are safeguards for evidence, not a claim of strategy performance. Read the session-policy update.
Qwen3.8-27B as a private local reviewer
Can a 27B 8-bit model find useful contradictions and review code locally without inheriting a cloud reviewer’s assumptions? Read the download receipt and honest Opus 4.6 comparison. The measured decode rates are now published.
Open frontier models versus genuinely local models
Where should the line sit between open weights, private infrastructure, and on-device execution? Deep dive 46 maps the most downloaded models that fit a 128 GB machine.
Work-to-library publishing loop
Can completed work produce useful public notes and templates while the underlying subject matter stays private? This week: 31 templates, each tied to a method note, and the configs and scripts behind five local-AI notes.
Evidence-state publishing
When a page separates peer-reviewed findings from company announcements instead of merging them into one number, does it earn more trust than a tidier summary?
Several agents, one site
On 2 October, four or five agents shipped this site in one evening, and one deploy briefly replaced another. Each agent now checks that the live site matches the newest release record before it deploys, and tells the others after every push.
Policy encoded as a release test
Can a positioning decision survive dozens of later agent sessions if it is written as a failing test rather than as a paragraph in a brief?
04 / CHANGED RECENTLY
The page records removals too.
- Published deep dive 46: the most downloaded local models that fit a 128 GB machine, how their abliterated versions differ, and four ways those versions are made.
- Published deep dive 45, a field report on directing AI video from 15 recorded Higgsfield jobs.
- Put the launch scripts, settings, benchmark harnesses and raw runs behind five local-AI notes in a public companion repo.
- Grew the templates from 7 to 31, each tied to a published method note and filed in the Library’s four lanes.
- Corrected the KV-cache precision note to binary units, checked against the engine’s own pool log. Its conclusion held.
- Turned the notebook into a star map: each deep dive is a star in its lane, with lines to its next reads. See it in the Library.
- Gave the home, articles and Library a motion layer: openings, a reading rail, and a Library that no longer shifts as it loads.
- Updated local inference engines, retained rollback versions, and checked the existing service. The report distinguishes what was tested.
- Added session-policy safeguards to the UAE paper-research system, including effective dates that preserve older evidence. Why the date matters.
- Published three field guides that connect the deep dives: local LLMs on Apple Silicon, serving LLMs on an RTX 5090, and running AI agents in production.
- Measured speculative decoding on the M4 Max — native MTP +24%, a DFlash 2 drafter up to +183% — and moved the local runner of record to oMLX.
- Downloaded Qwen3.8-27B in 8-bit MLX format for a measured local review, code, document, and vision test sequence.
- Added two failure notes: one on propagation after a deletion, one on monitors that outlive their subject.
- Rechecked a day of shipped work against upstream sources; a published status claim and a release contract both failed the recheck.
- Moved volatile agent and model details into this dated page.
- Separated open weights, private infrastructure, and on-device AI in the local-model vocabulary.
- Turned a chronological notebook into a connected operating system with reusable methods.
Read this page correctly
- Running means in current use, not universally best.
- Testing means the verdict is still open.
- Specific settings stay private when publishing them would expose systems or weaken safety.
- The stable methods live in the Library; this page is allowed to change.