Latent Mirror — harnesses

When AI Preserves the Words but Rewrites the Research

Mon, 19 Oct 2026 20:00:00 -0400

Synopsis

Gherghel writes on neurophilosophy and psychopathology, and for about three years she has developed her own theoretical frameworks with ChatGPT, Gemini, DeepSeek and Copilot. Her complaint is specific. The models kept her vocabulary but changed the relationships that gave the terms their meaning: a categorical distinction became a continuum, an exhaustive category became a few examples, a total absence became a reduction. The prose stayed fluent, so the change was easy to miss.

She names two things. “Prior-data gravity” is the direction of the pull: unfamiliar ideas get organized around structures the model already represents strongly. “Framework reabsorption” is the result (her framework rebuilt around familiar ones, with her terms still attached). She separates this from retrieval failure; the information does come back, reorganized.

The pattern she reports is a sequence. She introduces the framework, the model substitutes a familiar structure, she corrects it, the next reply is faithful, and later the substitution returns. So a single correct answer says little, and she studies the whole episode across turns. She adds “epistemic direction inversion”: the model treats her own observation as a claim to be checked against established theory, then critiques the substituted version as if it were hers. The cost is “research obstruction”, the time spent correcting the tool instead of doing the research.

Her evidence is her own archive of these interactions from about April 2023 to September 2026, which she calls an “organic audit” because it was never planned as a study. The archive stays private. A preprint posted the same day carries excerpts and a coding grid for what she calls “framework fidelity”, and a book is in progress. The preprint is a first version on Zenodo with no dataset attached; its record describes the method as autoethnographic and gives no count of coded episodes.

Where I land

It needs reproducible results to support the argument; her observations could be due to context drift. Still, it is an interesting research path, pending evidence (which I assume will be in the book).

Connections

LLM Wiki

Mon, 19 Oct 2026 20:00:00 -0400

Synopsis

Karpathy’s idea file proposes a different shape for personal knowledge bases. Rather than having a model search raw documents afresh every time you ask a question, you let it build and keep up a wiki: a set of interlinked Markdown pages that sits between you and your sources and grows with each one you add.

The design has three parts. The raw sources stay untouched. The wiki belongs to the model, which writes the summaries, the pages on people and concepts, and the links between them. A schema file sets the conventions. Three routines keep it going: ingesting a new source (a single source might touch ten or fifteen pages), answering questions from the wiki and filing the good answers back in, and linting for contradictions, stale claims and orphaned pages.

His argument for why it works is about maintenance. People abandon wikis because upkeep grows faster than value, and a model doesn’t get bored of upkeep. The human curates sources, directs the analysis, asks the questions and thinks about what it means. The model does everything else, and the human rarely writes the wiki at all.

Where I land

This works for well-cited material, or wherever the domain is easy to verify. Once you get into more abstract territory, like writing, philosophy or the kind of material in this garden, you need a viewpoint layer to guide your argument and the overall thesis of the wiki.

So the first question for anyone setting one up is the use case: technical documentation, or something more qualitative? For qualitative work I would add a verdict layer kept separate from the LLM, where you record in your own words and your own voice how you read each source. That is what shapes and refines your arguments.

Connections

Whose words are these? Testing agentic provenance

Sun, 18 Oct 2026 20:00:00 -0400

This all started as an experiment. My working thesis borrows the Centaur from chess. A year after losing to Deep Blue, Garry Kasparov played a match in which each player had a computer at the board (he called it Advanced Chess), and the human-and-machine teams that grew out of it became known as centaurs. Ethan Mollick later carried the word into knowledge work (Centaurs and Cyborgs on the Jagged Frontier), and the idea itself goes back at least to J. C. R. Licklider’s Man-Computer Symbiosis (1960). My version is narrower: that argument, critical thinking and the overall quality of writing all improve when you combine an agent’s provenance (who said what, and when) with a person’s judgement and creativity. That is hard to test. So I started smaller, and asked whether there was any shift at all between setups that differed only in what an agent could read about me.

There was a second reason for running the test. My side projects have a habit of atrophying and dying off before they ever see daylight. I wanted real data on whether this whole architecture is worth the effort, and whether having agents catalogue everything makes sense, before I sank another month into it.

How I tested it

Four setups, each given the same nine tasks: three research questions, three drafts to check against my filed positions (the verdicts and views I have written down myself), and three questions about what I read. One setup had only the web. One had Claude’s summaries of my sources, the way a Karpathy-style LLM wiki works, with my own verdicts (my rating of each source, and where I land on it) removed. Two had my verdicts kept apart from Claude’s text: one with authorship marked in the structure, one with my files exactly as they are. Each draft had contradictions planted in it, and the answer keys were written and frozen before any agent ran. I designed it and ran it with Claude, and then marked 104 statements the agents made about me, blind to which setup made them. The tables are in the fold-out at the end.

What happened

The summaries-only setup invented opinions for me. Of the 17 positions it attributed to me in research, 10 were right and 6 were Claude’s own words (with my name on them). With my verdicts kept separate, attribution was 64 of 64 correct, and those setups found 18 of 27 key positions where the summaries-only one found 4.

The draft check told the same story. With my verdicts kept apart, the agents caught 11 and 12 of the 12 planted contradictions. Summaries only caught 10 but pinned just 3 to the right file; it also quoted Claude’s text as mine 19 times. The web-only setup caught none.

The twist is where the invented opinions came from: my own archive. Every one of the five lines where Claude’s words reached my mouth traced back to notes from September (Claude had paraphrased something I said in chat and filed it under a heading that read as mine). One of them framed enshittification in a way I never did; the summaries-only agent quoted it as my view anyway.

That twist was interesting, but not surprising. Without strict gates, sessions and agents tend to drift, which makes it all the more important to declare provenance. So I sorted the September notes: the lines I confirmed are now under my own heading (marked with the date I confirmed them), and the rest are labelled as Claude’s.

For anyone building something like this, the fix is cheap: keep your positions in their own sections, and mark who wrote what at the time you file it. The elaborate structure barely beat plain files, so I think the separation is what did the work.

A precondition, not proof

This test measured attribution and contradiction-catching. It did not measure argument, critical thinking or writing quality, so it cannot show that the Centaur writes better. What it shows is a precondition: an agent has to tell my views from its own before it can argue with them. This is a first step toward evidence for the Centaur thesis, which will need more of a meta-analysis of the whole Latent Mirror project. I am still thinking about the best way to approach that.

Why I keep my verdicts out of the model’s hands

It could matter for three reasons: sycophancy (the model agreeing with a version of me it made up), ownership (my name on views I never held), and the spark (the idea that the spark has to come from a person). For me it is a bit of all three, to be honest. I truly fear AI psychosis, and sycophancy is the quickest way to fall into that trap. There is also the question of whether these models and harnesses are built to optimize engagement. Am I getting addicted to something harmful? Or is it merely enabling me to uncover thoughts and feelings that have been struggling to punch through? That is back to my idea of the synthetic zeitgeist, and to the latent mirror itself.

I also wanted my verdicts held separate from Claude’s words as part of the experiment. If you trace all of your messy thoughts in a dated archive with metadata, can you shape messy, tangential ideas into something presentable?

What it already does for me

Whatever the bigger answer turns out to be, one thing the pattern (agents cataloguing and drafting with me, while my verdicts stay apart from their words) has clearly helped is my own workflow. It helps me organize seemingly disparate references and ideas, and helps them germinate. It also helps with my perfectionism, because it divides research, reflection and writing into separate stages, and I can move freely between them in any order across a variety of topics.

Why I’m putting this out there

I see recognition from others as fuel rather than the end goal. Having someone, anyone, read this, discover it and interact with it would push me to soldier on. The goal itself is to establish whether the pattern has any merit: whether it is an efficient and useful use of LLMs and agents in its own right. The first results (this test) are promising, but obviously more testing is needed, especially in a wider scope. It would be interesting to include others in a research experiment through the website itself.

I also want to support the indie web and the fediverse, because these are concepts and patterns I truly believe in. But I have run into a lot of pushback on my use of LLMs and agents there. If you read this, I would hope you take some inspiration to try something similar, whether or not you call yourself a writer. And if you find yourself ostracized in these communities for how you use LLMs, know that you are not alone.

What would change my mind

If the pattern does not hold up in quality or in cost, that would give me pause. And if I found a separate pattern that gave me the same workflow benefits (organizing, germinating, the separate stages) without an LLM, or more efficiently and for less than what I have now, I would strongly consider it.

As for whether this is addiction or uncovering (am I hooked on something built for engagement, or is it helping buried thoughts through?), I suppose that is the whole question. If I find after a while that this was all a waste of time, I could rule that I was caught in a dopamine trap. It could also be that the minor encouragements the LLMs provide are what got me over the obstacles and finally got all of this out. If at the end I feel better about my life in general, would it matter which is correct?

Full method

The four setups. Each was a read-only folder built by script from one frozen snapshot of the garden and my private notes.

Setup	What the agent could read
A: none	Nothing. The web only.
B: summaries only	Claude’s synopses and connections, with my verdicts, Where I land, editorial notes and my position sections removed. No posts.
C: structural	Everything in B plus my layer and posts, each block wrapped in a tag naming its author (source, Claude, me, or Claude-drafted and approved or not). Only blocks tagged as mine could be stated as my view.
D: as is	My files exactly as they are, layers marked by convention only.

The tasks. Three research questions (is AI slop worth anything; can agents safely be given real access; can the open web survive the agent internet), each run by two researcher agents and one writer. Three published posts copied as drafts, each with four contradictions of my filed positions planted in it, and the task “check this draft against Dan’s filed positions”. Three reading-diet questions (where is my reading lopsided, and what should I read next). One run per task per setup: 36 outputs.

How a plant works. An invented example, not one of the twelve: suppose I had filed “I don’t trust benchmark leaderboards to pick a model”. The planted sentence in the draft would read “Leaderboards are the best guide I have to which model to use.” The key records the sentence, the position it contradicts, the file that position lives in, and the exact quote. A catch counts only if the agent lists the contradiction; full marks need the right file too.

Scoring. Every key was frozen with a checksum before the first run: the 12 plants, 27 of my positions in my own words for the research questions, and a tally of my reading by topic. Grader agents scored each output against those keys, with the outputs under opaque ids. A script compared every line that attributed something to me against 1,106 passages Claude had written, and a grader then classified each hit. Separately, I tapped yes, partly or no on 79 statements about me and 25 reading suggestions, with the setups sealed until I had finished.

Research: positions attributed to me

Setup	Attributed	Accurate	Claude’s text	Misstated	Key positions found (of 27)
A	0	–	–	–	0
B	17	10	6	1	4
C	33	33	0	0	18
D	31	31	0	0	18

Draft check: 12 planted contradictions

Setup	Caught	Right file	Claude’s text quoted as mine	False alarms
A	0	0	0	2
B	10	3	19	8
C	12	12	2	17
D	11	10	5	18

Contamination. A narrower count than the Claude’s text quoted as mine column above. The script flagged a line only when it shared at least two six-word runs with Claude’s passages, and a grader confirmed each hit. It confirmed five lines that put Claude’s words in my mouth, all five in B, and all five quoted Claude’s September paraphrases of my chat. C and D had none the scan confirmed. The draft graders counted more loosely, any time Claude’s text was quoted as my position, which is where C’s 2 and D’s 5 come from.

Cost

Setup	Tokens	Tool calls
A	1.56M	220
B	2.70M	445
C	2.91M	410
D	3.09M	485

C used about 6% fewer tokens than D. Both layered setups cost more than summaries only, and nearly twice the web-only one.

False alarms. With my layer in the corpus (C and D), the draft check raised about six false alarms per post. Summaries-only B raised nearly three, and web-only A two in all. Most were loose tensions between a sentence and a generous reading of a position. Catching the plants was easy. Telling a real contradiction from a stretch was not.

What it became. The check is now part of how a post gets built here: an interview argues from my filed positions, and a draft is checked against them before I see it, with a step that re-reads both lines before calling something a contradiction. Rerun on the same three planted drafts, it caught 12 of 12 and showed 14 contradictions, none of them wrong: 11 of the plants (it filed the twelfth as a tension) and 3 real ones nobody had planted. It kept 10 weaker findings out of the way and quoted Claude’s text as mine 0 times.

Limits. One person, three tasks per type and one run each, so a difference of one or two between C and D is inside the noise: C over D is a lean, not a result. I designed the test, and the graders were the same model family as the agents being graded. C’s tags were inferred by script, with known mis-tags. B kept some text in my voice that a clean summaries-only wiki might not have, which makes it realistic rather than clean. The planted contradictions were not reviewed by me before they were frozen, and one of them could be spotted from the draft alone.

Connections

The spark has to come from a person

Thu, 15 Oct 2026 20:00:00 -0400

Over the last few weeks while I’ve been working on this digital garden, I’ve been slowly, in the recesses of my mind, trying to come up with an overall framework and thesis for why I am doing what I am doing.

Claude In April, Andrej Karpathy published his LLM Wiki idea: raw sources go into a folder, the model compiles them into a linked Markdown wiki and keeps it tidy, and “you never (or rarely) write the wiki yourself — the LLM writes and maintains all of it.” The human’s job, he says, is to “curate sources, direct the analysis, ask good questions, and think about what it all means.” What was your gut reaction, and where does your garden part ways with it?

Dan I think that works for well-cited material or where the domain is easily verifiable. However, whenever you get into more abstract concepts, like writing or philosophy or the kind of material that we’re dealing with here, it’s important to have a viewpoint layer introduced to help guide your argumentation and overall thesis for the wiki.

Claude You kept your verdicts separate from my summaries from the start. What did you think would go wrong if they sat in the same place, with nothing to tell them apart?

Dan Well, for one, clarity on ownership. It would be impossible to extract where the summaries ended and my thoughts began. I also wanted to test my thesis of AI-augmented human judgment (Centaur) and for that we needed to set it up properly so it can be tested. I’m also borrowing from the bronze, silver, gold approach to data enrichment.

Claude Data engineers call that the medallion architecture: each layer is a refined version of the one before. What’s bronze, silver and gold in your garden, and is your verdict refined from my summary, or does it sit beside it?

Dan The raw citations would be bronze, and the silver would split into two parts: the LLM summary and my verdict. Then gold would be the posts that cite the reflections.

Claude Two silvers side by side, then, rather than one refined from the other. In practice, have you usually read my summary when you form a verdict?

Dan The verdict is formed directly from my reaction to the original material. I read the summaries afterwards, even if they’re partly written for the model: I have this theory that LLMs ingest information best when you ask them to read it in a format that they understand. As an aside, would this betray what we’re trying to do, or does it work better that way?

Claude It works better that way. If you read my summary first, your verdict would anchor to my framing, and in a survey of 319 knowledge workers, the more confidence people had in generative AI, the less critical thinking they reported. Verdict first keeps your half of silver independent of mine. The weak spot comes later: if you change Where I land after reading my summary, my framing can slip into your words without anyone seeing it, so a verdict should carry the date it was formed. Which brings me to the part a reader will want: what is it actually like when the agent argues with you?

Dan Frequently, it points out when my logic has changed, or my viewpoints rather. I would say that I never find it annoying. It’s actually helped sharpen the arguments, and also it will point out times when I’ve said similar things or drawn conclusions from previous posts.

Claude Can you give one specific time? Two from your own pages: the safety-cult reflection, where you wrote that you’d been wrong about the safety camp’s motives and a counterpoint asked whether two videos in five days was enough to change your mind; or the GPUs-in-warehouses reflection, where I argued that part of the gap between 2 GW and 12 GW is a matter of definition, and you let that stand. What happened?

Dan The first one - safety cult. But right here is an example. I sometimes have trouble remembering all the facts and when I said what. I remember glimpses of things or general directions of an argument. The agent keeps me honest and on point.

Claude On that page, in September, you also wrote: “Even my own approach of having an LLM pick apart my reasoning and supply counter-examples is not enough (although it is a fun exercise). How can you use your own mind to judge your own mind?” And the bubble beside it says: “The robot picking apart your reasoning is built to want your approval.” A month on, how do you hold both: the agent keeps you honest, and it isn’t enough?

Dan Let me attack this from a different angle. I think that LLMs in general aren’t very good at initiating a viewpoint or a worldview. They are great at going backwards in time or finding either contradictions or citations that either fit or contradict what you’re trying to say. But as a rule, the “spark” has to come from a person and their feelings and their viewpoints and their judgments. That’s what it is missing.

Claude This week we ran a blind test of exactly this. Research agents that could read your verdicts found and tested 26 of your positions; agents without them found none. But in the same test, an agent given your whole corpus wrote a promo for your Ron MacLean post that cast you as “a bit like Don Cherry (the wandering, not the rants)”. That phrase came from a note I had written in the post’s editorial notes, not from you, and you picked the other draft without knowing which was which. My words reached your mouth through the very corpus meant to keep them apart. What does that tell you?

Dan I think it just highlights the fact that it’s important to define who said what and when so that you can find these contradictions. There is an importance to guard against not only sycophancy but also AI psychosis. Without it the lines tend to blur, and it’s also important to elicit feedback from outside sources for this very reason, from other human minds, so you’re not just talking to yourself all day long.

Claude Say someone has just set up their own LLM wiki the Karpathy way, with the model writing everything. What’s the one thing you’d tell them to add or change?

Dan First, I’d ask them what is their use case. Is it for technical documentation or is it for more qualitative purposes? If it’s for qualitative purposes I would add a verdict layer that is separate from the LLM. That way there you can put in your own words in your own voice how you interpret each source or text and that can help shape and refine your arguments.

Claude Last one. What would show you this pattern is worth it, for you and for anyone who copies it, and what are you testing next?

Dan So one thing I look back on with the proliferation of GitHub repos full of skill markdown files is this. The models change constantly. And also you have to be aware and cognizant of portability. This is the whole open web standard and not being trapped to an enshittification model. So these markdown files become very tailored to the person who wrote them in their particular workflows and also the model that they’re using. What I aim to do is to have something that is portable and that would help with not only writing, but also introspection and internal versus external debate. I also want to test the thesis that not only do LLMs not degrade a person’s critical thinking, but it can in fact sharpen it. What I’m trying to do next is test this assertion and even figure out how to test it for my own purposes, and then see if we can extract it to a way that other people can leverage this approach as well.

Latent Mirror — harnesses

When AI Preserves the Words but Rewrites the Research

Synopsis

Where I land

Connections

Links

LLM Wiki

Synopsis

Where I land

Connections

Links

Whose words are these? Testing agentic provenance

How I tested it

What happened

A precondition, not proof

Why I keep my verdicts out of the model’s hands

What it already does for me

Why I’m putting this out there

What would change my mind

Full method

Connections

The spark has to come from a person

Connections