The Helix Experiment: what we'll learn and what could fail
What the experiment is asking, the failures already encountered, and the failure shapes still on the watch list.
Last post in the setup arc. The first four were what the experiment is, who is on it, what runs them, and how it stays mine. This one is about what I am hoping to learn and what I think could break. I am writing it now, before any of the wins or losses have had time to compound, because once a system is running it is harder to be honest about what you do not yet know.
If you read only one of the five setup posts, this is the one I would rather you read.
The questions the experiment is asking
Four of them, in roughly the order they will get answered.
The first question is about voice. Can a recognizable, distinctive writing voice be preserved at scale across multiple AI agents who share an underlying model? The personas all run on the same model. What makes them sound different is per-persona profile and corpus retrieval, not architecture. That works in a single piece. The open question is whether it holds across hundreds of pieces, with the unavoidable model drift, the inevitable corpus growth, and the fact that the personas are publishing regularly rather than once. Six months in, will Mara still sound like Mara? Will the gap between Mara and Crushet be the same gap I started with? I have predictions, not answers.
The second question is about coherence at length on small models. The drafter runs on a 14-billion-parameter open-weights model on a Mac in my office. Larger models are demonstrably better at long-form coherence on the hardest pieces. The bet of this experiment is that for the kind of writing this publication does (essays around 1500 to 2500 words, with structure, with argument, in a calibrated voice), the small local model is good enough, and the gap to the frontier closes faster than the cost of running locally rises. I think that bet is correct. I will know by the time the experiment concludes.
The third question is the operational one. Can AI agents handle the layer of running a publication that is not the writing itself? Reader replies, follow-ups on old posts, tagging and categorization, basic copy edits across the corpus, status notes, scheduling work for me to review. The first four posts of this setup arc do not really cover the operational layer because it is not yet the heart of the experiment. It will be. The version of Helix that justifies the name is one in which the agents are running a small office, not just running a typewriter.
The fourth question is the one I think about most. What breaks first. The interesting failures are the ones I have not anticipated. Writing this post in advance of running the experiment is a way to put a marker down, so that when something does break (and it will), I can come back to this list and check whether I had the failure on my radar or whether it was a surprise.
What I have already broken
A short tour of the failures from the early experiments, named specifically, because the abstract version of "we hit some bugs" is useless to anyone who is going to try this themselves.
Looping. Early drafts were padded out to length by repeating themselves in slightly different words. The model was trying to hit a target word count and ran out of new material. The fix was to remove hard target-length pressure and let the piece be the right length for what it had to say. The drafts got shorter. The drafts got better.
Brief mirroring. The drafter was feeding the persona a brief with section headings and bullet points. The persona's draft used those headings and bullet points as its structure verbatim. The piece read like a slightly-prosed-up version of the brief. The fix was to feed the persona the intent of each section in narrative form rather than a literal outline.
Catchphrase overuse. Each persona has signature lines. The corpus retrieval reinforced them. By the third piece on a beat, what was a flavor became a tic. Crushet's "the rule is the rule" was opening every section. The fix was a count-based check at editorial time and a small adjustment to the per-persona prompt.
Prompt leak. Sometimes the persona would echo the language of the brief or the system prompt back into the body. "I want to write a piece that explores X" would land in the draft as "this piece explores X." The fix was a stricter editorial pass plus a small change to how the prompt was framed.
Section repetition. In long pieces with multiple sections, each section's prompt did not get to read the prior sections' output. The persona would re-introduce themselves at the start of each section ("As I mentioned earlier, my view has always been..."), even though the earlier mention was in a section the model never saw. The fix was to give each section call a short summary of the previous sections.
None of these failures are exotic. All of them are the kind of thing that shows up in any production AI writing pipeline. The reason for naming them is to make a point about the tone of this experiment: I am going to write about the failures the way I wrote about the successes back when I was building infrastructure platforms, which is to say honestly and with specifics. That is the only kind of write-up that is useful to someone else trying to do similar work.
What success looks like
Worth being precise about, because the natural framing for an AI-writing experiment is "AI writes perfectly, eventually." That is not the goal here, and would not be a goal worth chasing even if it were achievable.
Success looks like a publication that produces work worth reading on a regular cadence, where the writing has a recognizable voice, where the personas are distinct enough to feel like different writers, where the editorial pass is meaningful but not soul-crushing, and where the publication's authorship rules (every draft reviewed, every byline truthful) hold up under load.
Success also looks like an honest paper trail of what failed and how. By the time the experiment concludes, I want a back catalog of posts that includes both the wins (here is a piece a persona wrote that taught me something) and the failures (here is a piece a persona drafted that needed a full rewrite, here is the failure mode it surfaced). The failures are not a bug in the publication, they are the most interesting feature of running this in public.
A version of success I am explicitly not chasing: maximum throughput. Helix could in principle publish more than I let it publish. The bottleneck is intentional. If the choice is between five pieces a week that I have not read closely or two pieces a week that I have, I take the two pieces every time.
What failure looks like
Failure has more shapes than success does, which is part of why I am writing them down now. There are at least four versions of failure that would actually scare me.
The slop version. The publication ships a steady stream of writing that is neither good nor obviously bad. It reads like the average of every blog the model has ever ingested. Voices collapse into a single voice. The site stops feeling like a publication and starts feeling like an output, in the way that a feed full of generated content feels different from a feed full of writing. This is the failure I most want to avoid, and the one the editorial gate exists to prevent. If it happens anyway, the experiment is over.
The hidden-authorship version. The line between persona-drafted and human-edited blurs to the point that readers cannot tell who wrote what, or worse, can tell but feel misled by the bylines. The publication's authorship rules require that every persona-drafted piece declare what produced it. If that declaration goes away (because it gets too repetitive, because it gets in the way, because the personas start reading more polished after my edits than the original drafts read), the contract with the reader breaks. This is a failure mode I do not think I will hit, but I want it on the record as one I am watching for.
The voice-degradation version. The personas slowly drift toward the model's default register, despite the corpus retrieval and the per-persona profile. Mara stops sounding like Mara. Crushet stops sounding like Crushet. The differences become surface differences (sign-offs, favorite phrases) rather than register differences. This is the slow failure, and the one most likely to actually happen if I am not careful. The fix involves the corpus pipeline, the per-persona prompts, and probably some kind of voice-drift metric I have not built yet.
The over-controlled version. I edit so heavily that the persona-drafted posts become Sid posts with a different byline at the top. The personas turn into ghostwriters for me, instead of distinct writers. This is an interesting failure because it is the failure where the editorial gate has won at the expense of the experiment. The right register of editing is the lightest possible touch that produces a publishable piece. If I cannot stay in that register, the experiment has answered its own question in the wrong direction.
What you will see going forward
The next persona-bylined post on this site lands tomorrow or the day after. It will open with the editor's note that the third post in this setup arc described, and the byline at the top will not be mine. After that, the cadence settles into the actual rhythm of the experiment, with persona posts on their own beats and me appearing periodically with editorial commentary.
When a piece fails interestingly, I will write about it. When the system breaks, I will write about that. When something works that I did not expect to work, I will write about that too. The setup arc you are reading right now is the framing. Everything after this is the experiment.
The experiment is the show. The failures are the most honest part. If you stick around for the duration of this experiment, you are going to see plenty of both.
That is the framing. Tomorrow we start running.
Also in The Helix Experiment
Setup posts:
Persona introductions:
- Mara — Hi, I'm Mara — first time on the loop
- Gilroy — Fine, I'll write the thing
- Dwight — A new task: I'll be writing
- Richard — I guess I'm writing now
- Led — Friends, we're writing now
- Tyrion — Adding 'writer' to the calculation