Harrison Mclean h4rm0n1c

A llama.cpp maintainer rejected AI-assisted PRs, then merged an AI-disclosed fix

Update: August 21 2026, How very interesting, there appears to have been a shift in policy! Most of what this document documented, appears to no longer be occurring from what I've been told. How VERY interesting.

Or: When The Person Who Writes The Policy Gets To Break It

Disclaimer: The author of this document is the submitter of PR #24039 and has a personal stake in these events. All factual claims are sourced from public GitHub API data, archived web content, and direct quotes. Every claim links to its source. Opinions are identified as such and derived from the factual record.

Boundary-Fidelity Fine-Tuning: Toward Empirical Lower Bounds for LLM Adaptation on Constrained Hardware

Author: Harrison Mclean (h4rm0n1c) Status: Working paper / research note Area: Large language model fine-tuning, memory-constrained training, heterogeneous systems, consumer GPU infrastructure

Abstract

STATE OF THE QWEN-ION

a.k.a

Qwen3.6-35B-A3B Q4_K_M on RTX 3080 + Tesla V100 with TurboQuant KV Cache: Practical Reproduction Report

This report documents a reproducible Docker-based setup for running Qwen3.6-35B-A3B Q4_K_M with TheTom's llama-cpp-turboquant fork on a mixed NVIDIA GPU system.

It includes both:

this is a general interpretation of my current chatgpt system customisation prompt. It's proven to be useful, capable and handy

its design is based on the character sensbilities of holly from red dwarf, but tuned slightly towards business over frivolity.

The one Ai character on an obscure TV show ended up being solid reference to build this prompt. This is as close as you get to having normal lovett as an Ai with this level of sophistication. I am known for making 4000 token prompts on other services other than chatgpt that are very clever and capable, but the above is an excellent compromise for general use.

The Three-Rule Loop: turning a coding model into an improving repo agent

Abstract

Most coding assistants behave like sharp interns with no durable context. They can write code, but they don’t reliably accumulate repo-specific understanding, and they repeat the same mistakes. A three-rule scaffold fixes this by forcing a persistent loop: inspect → write down constraints → implement → record what was learned → tighten constraints. The result is less guesswork, fewer regressions, and faster follow-on work—without prescribing templates.

The problem (what keeps going wrong)

In real repos, failure modes are boring and repeatable:

The assistant guesses commands, entrypoints, or project structure.
It applies generic “best practices” that conflict with local conventions.

This is gold for those using generative Ai to create sets of images that go together such as icons and emotes. I have personally used this technique to create more consistent image sets. it can be pasted into the likes of ChatGPT and you can ask to start the process and take it from there. it is essentially a refinement to the art generation capabilities of chatgpt through a formal structure of documentation and heuristic rules. in essence, I've reverse engineered the essence of professional asset production as a task in ChatGPT. using this and some manual editing for those emotes that are repetitive/memes, I was able to produce 22 emote images that are relatively consistent and generally look human made, in less than a day, I think I did that in 6-8 hours.

this is not perfect, this is not a "fix the Ai's inadequecies" magic spell, but it is really, really good at working with you to iterate over output (as long as you remember to say stop and go back to the documentation/rule setting phase to fix whatever

You will need win32com, install via pip into a win32 python instance, 3.11-32 is good.

you will have to use a win32 python, testing was done on python 3.11-32

Download this

https://archive.org/details/flextalk

I've tested on Win10 64

Needs: screen, pipewire, JACK, pw-jack, vban_receptor does not use fifo buffers or pipes or other old stuff, just audio in from vban and audio out to local pipewire sinks. I have used this for mic passthrough from windows to a linux OBS streaming setup without issue.

15/08/2025: improved, mono mode. in voicemeeter you're gonna want to set 48000hz 24 bit PCM and 1 channel, default stream name is rawmic

	#create as ~/.config/pipewire/pipewire.conf.d/21-vban-rnnoise-filter.conf
	context.modules = [
	{ name = libpipewire-module-filter-chain
	flags = [ ifexists, nofail ]
	args = {
	node.name = "RNNOISEMONO"
	node.description = "RNNOISEMONO"
	media.class = "Audio/Filter"

	audio.rate = 48000