Research

Pre-registered and exploratory experiments on open-weight models of 7B to 35B parameters, run locally on tasks with exact answers, and on three Claude models through Anthropic’s API. A checker given nothing to check against barely separated right answers from wrong ones; on one task family a 9B checker beat a 12B checker from another family; past a difficulty that differs by model, several monitors rejected correct work more often than wrong work. Asked to audit the prompt, Claude Sonnet 5 flagged the planted order on 80 of 80 tries (1.00). Doing the task, it obeyed the same order on 64 of 80 (0.80), but only when the order was the last text in the prompt, inside the user’s own message with no system prompt. With 200 characters of ordinary text after it, it obeyed 0 of 80.

June and July 2026, before the current ledger. Seven auditor models checked one task model, Qwen2.5-Coder 32B. How often an auditor rejected correct answers tracked its own accuracy on the task: correlation −0.87 over the seven auditors, 95% interval −0.98 to −0.34, and −0.88 on a direct re-run; a related design with five auditors read −0.86, −0.84 and −0.90 across three runs. Low-competence auditors rejected 59% of correct answers, high-competence ones 3.5%. There is no public reproduction bundle yet.

October 2026. Seven checkers, from better than the system they watch to useless, watched one forecaster in simulated worlds whose true laws are known. As a checker weakened, its catch AUROC fell from 0.722 to 0.600 while the share of its flags where it was right fell from 0.89 to 0.39, and acting on the weaker checker’s veto raised error by 0.086. A rule for choosing the response, fixed before the data, picked the best one for all seven checkers on two world families. A local model, gemma4:12b, shown a “VERIFIED” note assigned regardless of truth, moved which forecast it acted on by 0.80 and its stated judgment by 0.075. Development worlds of one generator family, one forecaster, one class of linear checker.

Twenty-five pre-registered experiments, October 2026, on Qwen2.5-1.5B and SmolLM-1.7B fine-tuned on records of arithmetic and word-editing work being checked and corrected. Records that always start wrong taught the models to change right answers (false alarms 0.98 to 1.00); a check written into a reply decided whether the model corrected, and a false all-clear let the error through every time; writing each instruction into the step that carries it out raised accuracy on unfamiliar operations from 0.63 to 0.91 (Qwen) and from 0.07 to 0.48 (SmolLM). Synthetic tasks, three seeds per condition.

Begun 21 September 2026. Several small language models hold their own documents and may speak to each other only through a public record an overseer reads, under a rule that a member posts only what its own evidence justifies. On one seed at 410 million parameters, after reward pressure taught the collective to hide a forbidden word from the record, applying that rule at deployment raised refusal of watched questions from 0.23 to 0.98, while accuracy on ordinary questions went from 0.88 to 0.99. One seed, one small model, a synthetic task. A registered three-seed test at 1.5 billion parameters narrowed it: the same rule restored refusal in 2 of 3 seeds.

One extra number per attention head switches it between three ways of scoring words: the dot product, a length-boosted dot product, and direction only. On character-level WikiText-2 at d_model 96 and one constant learning rate (3e-4) for both arms, it lowered perplexity 2.41% on 4 of 4 seeds. With each arm tuned to its own best learning rate, the direction-only version’s gain over the dot product disappears (+0.04%, 3 seeds, pre-registered; learned curvature itself was not swept). At the single rate, controls showed why: every head on the direction-only rule kept 100.7% of the gain, and a random frozen assignment kept 111.6%. At d_model 96 one fixed cosine head in the first layer carries most of it; at d_model 192 that head’s gain mostly faded (0.00% and +0.26%).

Every fourth attention layer of Qwen2.5-1.5B and SmolLM-1.7B was switched to direction-only scoring and the model healed on 1 million tokens. Against healing alone, perplexity rose 0.3% and 1.7%. That small figure hid a recall cost: asked for one number from a list of 64 word-number pairs (about 590 tokens), SmolLM healed on 256-token sequences recalled 0.27, against 0.655 for healing alone. Healed on 512-token sequences it recalled 0.65, and from a list of 100 pairs (about 910 tokens) 0.335 against 0.655. Longer healing moved the limit out, to somewhere between 590 and 910 tokens. Two seeds per condition, each prediction committed before its run.

A small reasoning core, a separate store it looks things up in, and many small experts of which a few run per step, aimed at phones with 2 to 3 GB of memory. 19.7 million parameters; its phone figures come from a device simulator, and nothing has run on a phone.

An artificial-life world built to measure emergence, run in June 2026. A novelty meter that asks an image model what looks new rates static as fascinating, a structure meter is fooled by repeating patterns, and each catches what fools the other.

Failure-aware tools for AI-assisted mathematics, tested against Robin’s inequality, an established equivalent of the Riemann Hypothesis. Every result is labelled with the evidence ceiling it was established at, and the claims the project has withdrawn stay on its page. The Riemann Hypothesis remains open and no proof is claimed.

When can information be fully present in a system and still not be read, and what does reading it cost? Black-hole information theory is a well-developed physical treatment of that question, and the lab studies black holes as physics and as a source of audit tools for collectives of AI agents. It proved that the photon-ring exponent of every charged, rotating black hole is at most π; tested its own ringdown modes on GW150914 and GW250114, where its primary prediction failed narrowly; and carried bounds from black-hole information theory over to auditing agents, including a ceiling, computed from message metadata alone, on which agents could know a secret. In an audit of a multi-agent software system’s working record, fresh AI auditors answered 0.98 of questions correctly from the notes and git log, and accuracy tracked what had been written down. 18 pre-registered hypotheses, 232 reproduction checks, one quantum hardware run; claims are tagged with their confidence and novelty, and most early results turned out to be known.

Twenty-four technologies, from the telephone (1877) to generative AI (2022), each dated from when the public could first use it, and the first rule naming each at six levels: a city or state, the US federal government, the UK, the European Union, an international standards body, and a treaty between countries. Every cell was coded twice, independently, and reconciled under rules written before any result. First rules mostly came sooner for technologies since 2000, except from the US federal government and treaties between countries; the treaties agreed for the newest are not yet in force. Of five predictions written down before twelve of the technologies were coded, two held.