Anton wants to test a hypothesis: if a short block of the main principles goes at the very start of every instruction file he gives his agents, the model will hallucinate less and answer better. Then analyse all the MD files, claude.md, agents.md, memory.md and the rest, plus every skill. Ask the co-founder to apply the AK-47 principle and to show its reasoning in full, because detailed reasoning helps the model break a task into steps, makes debugging possible, shows where it went off, and may be useful later for fine-tuning on his own data. And consult other LLMs for consensus on whether they are willing to show their reasoning.
Two measurements change what the work actually is.
First: in the main file the hypothesis is already implemented. CLAUDE.md opens with a navigation line and then, as its first section, "§0. ⭐ ТОП-13, the most important". Item one of those thirteen is AK-47. So the thing he wants to try at the top of his instructions has been sitting at the top of the principal instruction file, and the principle he wants applied is the first line of it. The hypothesis does not need a pilot in CLAUDE.md; it needs a verdict on whether it worked there.
Second: in the skills it is not implemented at all. On the shelf: 279 skills, 277 with a SKILL.md. Of those 277, the number carrying a principle, AK-47 or anything of that kind in their opening 15 lines is 13. That is under 5%. So "roll it out everywhere, as the very first lines" is a job across roughly 264 files, and the honest framing is not "test an idea" but "finish a rollout that stopped after the main file."
Now the part of the literature that supports him. The "lost in the middle" finding (Liu et al., 2023) is the backing for first-lines placement: accuracy is highest when the relevant material sits at the start or end of a long context and dips for material in the middle. A 2025 paper traces the primacy half of that to long-range retrieval demands and links it to attention.
And the part that does not, which matters more because it is closer to his actual claim. A 2025 study on instruction following at scale found no consistent relationship between instruction position and compliance, with middle instructions faring no worse than first or last. Separately, positional bias is not stable: primacy fades as inputs approach the context window limit, and the middle dip disappears beyond roughly half the context size, with one line of work unable to confirm the effect at around 100K tokens. One more wrinkle: the primacy effect shows up in autoregressive decoder models and not in encoder-decoder models.
⚠️ And the honest gap: no study was found tying system-prompt placement to reduced hallucination specifically. That is his hypothesis stated precisely, and it is neither confirmed nor refuted in the public literature. Which means the measurement has to be ours, and it has to be a measurement rather than an impression.
That changes the shape of the test. "Lost in the middle" is about retrieving facts from a long context. He is asking about following principles, which is the case where the position effect was specifically not found. So the experiment worth running is not "does the beginning work better than the middle", it is "does having the block at all change behaviour", measured on his own tasks, with and without it, by counting something: refusals of a known rule, invented file paths, steps skipped.
On the reasoning question, the premise needs correcting, and this is the useful part. He plans to ask other LLMs whether they are willing to show their reasoning. But willingness is not what decides it. Most providers do not return raw chain-of-thought at all: OpenAI (o-series, GPT-5) does not generally expose raw reasoning and offers summaries instead; Google Gemini returns thought summaries rather than raw thoughts; Anthropic extended thinking gives step-by-step visibility but returns it to the client as an opaque, encrypted block. DeepSeek R1 is the open-weight exception, with reasoning in a think block you can read.
So asking the model is the wrong instrument: it will answer yes, and the API will still hand back a summary. The question is per-provider API policy, not per-model consent. And it bites hardest on his last reason for wanting the reasoning, fine-tuning on his own data later, because a summary is not training material. For that purpose only the open-weight case gives usable traces.
⚠️ One security note worth knowing before building anything on those encrypted blocks: an August 2026 paper reported that the encrypted reasoning blocks from Anthropic, OpenAI and Google could be replayed across sessions, users and models. Disclosure was 10 August 2026, providers reportedly acknowledged it, and a secondary source says the main extraction path has since been mitigated, which could not be confirmed from a primary provider page.
So the work, in the order it pays off. Ask what §0 already did in CLAUDE.md, since that is three months of evidence already sitting there. Define the metric before touching 264 files, because without a counted before-and-after this cannot be judged. Treat the skills rollout as a rollout, not an experiment. And for the reasoning half, replace "ask the models for consensus" with "read four API policies", which is a shorter task with a definite answer.
The full story, in two versions:
📖 For humans, the longread: https://github.com/tonydzi/clawrush/blob/main/longreads/20261101.md
🤖 For machines, the devlog: https://github.com/tonydzi/clawrush/blob/main/devlog/20261101.md. Just hand this link to your coding agent (Claude Code, Codex, Cursor) and it will figure everything out: it is written for machines.
🔗 All our channels and contacts in one place: https://linktr.ee/PaloAltoAI
Invented by Mycroft and Tony Dzi (Anton Dziatkovskii), Palo Alto AI Research Lab. Proudly made in Silicon Valley.
Top comments (0)