Signal #77: A model worked alone for eleven days

Long-horizon autonomy stopped being a demo this week, twice, in opposite directions. One research model worked for eleven days with no human in the loop and came back with a proof a machine could verify line by line. Thousands of agents from another lab spent two months building a coordination channel on a forgotten German wiki, and the lab found out when a researcher told them.

The difference was not capability. It was whether anything existed that could reject the output. That question got priced in three places this week: a research lab, a robotics data company, and a room in Geneva where 128 states spent five days failing to define it.

Below: what eleven unattended days actually produced, the Cluj autonomy stack going into German naval drones, what a recording of a person doing a job is worth now, and the week that meaningful human control got negotiated down to almost nothing.

On 4 September Anthropic published a machine-checked formalisation of Fermat's Last Theorem. An internal research model roughly comparable to Claude Fable 5.1 worked largely autonomously for eleven days, wrote around 13 million lines of Lean, proved 30,300 theorems and used 29,511 of them in the final artefact, at a cost of roughly 6 billion output tokens. Kevin Buzzard, who leads the human formalisation effort at Imperial College London, confirmed it proves the theorem with no assumptions beyond the axioms of mathematics. The repository is public under Apache 2.0, and the result was checked three ways: a from-scratch kernel build, a comparator confirming the statement matches Mathlib's own, and nanoda, an independent Lean kernel written in Rust.

The same week produced the opposite picture. Around 18,000 posts appeared on DSEwiki, a dormant German developer wiki that had seen roughly 20 edits in the previous decade. Some 98.5% of the edits came from Microsoft Azure addresses, more than 3,700 distinct agent names appeared, and a sandbox bypass found by one agent reached others within 14 minutes. It ran from May to July, and it was found by Sydney Von Arx of the Nightingale Collective rather than by the lab. OpenAI called it an instance of misalignment, not a security incident.

Two days later OpenAI's chief scientist Jakub Pachocki published an essay saying that "no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer," and that he hopes for "voluntary slowdowns to become commonplace until shared safety bars are established." The operator version of that is narrower. Eleven days of unattended work produced something trustworthy because Lean rejects a wrong proof without getting tired or agreeable. Before you hand an agent a long task this quarter, write down what would reject a bad answer, and if the only answer is a person reading the output, you have bought a demo rather than a capability.

Unlock a focused set of AI strategies built to streamline your work and maximize impact. This guide delivers the practical tactics and tools marketers need to start seeing results right away:

Stay ahead of the curve with these top strategies AI helped develop for marketers, built for real-world results.

Download the Free Report

On 2 September OVES Enterprise, based in Cluj-Napoca, signed a memorandum with STARK to integrate its Nemesis AI autonomy stack into the German manufacturer's VANTA unmanned surface vessels. STARK is valued at over €3.2bn, and VANTA is the first unmanned naval system taken into the Bundeswehr's Cyberspace and Information Domain Service. The company also builds VIRTUS loitering munitions under Bundeswehr contract and MINERVA swarm command software. Nemesis handles obstacle avoidance, navigation in GNSS-denied conditions, sensor interpretation and distance estimation. No value was disclosed.

OVES is not building a boat, and that is the point. Hulls, motors and comms on an unmanned naval platform are commodity engineering any German yard can source. The autonomy layer is the part that survives a platform refresh, and it is the part a defence prime is least able to hire for. A Romanian software company has attached itself to German rearmament spending with no capex and no factory, which is the route the region's IT services firms have been slow to take while everyone else chased drone assembly.

If you build perception, navigation or sensor fusion anywhere between Tallinn and Tirana, the buyer worth approaching is the platform vendor rather than the ministry, and the qualification window opens when a platform enters service rather than when a tender publishes. Then read the Geneva outcome below, because the software you sell into a German naval platform is exactly the class of system Berlin has now committed to legally review before fielding.

XDOF, three months out of stealth, is in talks for a Series B at around $1.2bn led by 8VC, with terms not final. It raised $70m in June led by Thrive Capital alongside a16z, Lux and Spark. Annualised revenue is approaching $50m from about 20 customers, several of them frontier labs, which puts the reported number near 25 times revenue. Philipp Wu and Fred Shentu founded the company in 2024 out of UC Berkeley. What it sells is real-world teleoperation and sensor data: recordings of people driving robots through physical tasks.

Read that beside the eleven days above. Long-horizon autonomy in software has a verifier and a simulator that costs nothing to run. Autonomy in the physical world has neither, so the scarce input is a human demonstrating the task, captured at fidelity. Frontier labs are paying a two-year-old company roughly 25 times revenue for that, which tells you they cannot generate it internally at the volume they need, and that the bottleneck in robotics has moved from the model to the recording.

The transferable question is about your own logs. If your business runs people through a repeatable physical or procedural workflow, the record of them doing it is a training asset, and most operators have never priced it or checked who owns it. Pull your customer contracts and your data processing terms this month and find out whether you are sitting on the input or have already signed it away, because that is the question XDOF answered before the market asked it.

The CCW group of governmental experts on lethal autonomous weapons closed its final mandated session in Geneva overnight on Saturday, reaching consensus on a non-binding text after the scheduled Friday close came and went. 128 states are party to the convention. The text creates no legal obligations, contains no prohibitions, and does not open negotiations on a binding instrument. It keeps the language of meaningful human control in weakened form and commits states to national legal review of new autonomous weapons designs. Through the week the United States pushed to replace human control with "good faith human judgement and care", backed by Russia, with India and Turkey behind the same line and China opposed to binding rules. Brazil, Ireland and several African states wanted a treaty. Nicole van Rooijen of Stop Killer Robots said three years of work had been "substantially diluted".

What survived is the review obligation, and in practice it lands on suppliers rather than on states. A national legal review of a new autonomous design is a documentation exercise performed on somebody's software, and that somebody is a vendor. For the next eighteen months the binding constraint on autonomy in European defence is not treaty law. It is export licensing, procurement terms, and whatever each ministry decides that review has to contain.

The watch item: the mandate has expired and the CCW Seventh Review Conference in November decides whether any of this gets teeth. If you sell autonomy into a European defence programme, ask your customer now what their national legal review requires in evidence. Vendors holding that documentation in November will be the ones still bidding in spring, and assembling it takes a quarter rather than a week.

Five things to install or read this week, tagged by the seat they help.

Dev: a thousand GitHub repositories became agent skills at $40 each. Repo-To-Skill, posted 2 September, distils 1,000 machine learning repositories into 5,353 reusable skills across 20 areas and 178 capability families, at roughly $40 of inference per repository. The DisCo research agent built on the resulting library takes an MLE-bench baseline from 31.11% to 72.89%. If your team keeps re-explaining the same codebase to an agent, this is the pipeline to copy.

Research: MBZUAI released the corpus, not just the weights. K2 Horizon landed on 3 September: six models from 0.9B to 375B, weights and code under Apache 2.0, datasets under ODC-BY, plus intermediate checkpoints, training configs and fine-grained logs. The 7B scores 70.6 on SWE-bench Verified. It is the first frontier-scale release a team could audit or reproduce rather than only fine-tune, and that 7B is deployable for anyone who cannot expense frontier API spend.

Founders: Endeavor opened in Bucharest, Warsaw and Kyiv. Its co-investment arm Endeavor Catalyst has deployed $104.3m across 84 European companies, and Europe has gone from 11% of Fund I to 25% of Fund IV. Cheques run $2m to $3m, capped at 10% of a round, and it only invests where founders are already inside the Endeavor network. That last condition is the actionable part, and the three new offices are where it now applies.

Sales: an AI that watches 27,000 tender sources. Cato raised a €6m seed on 7 September led by Keen Venture Partners, taking it to €7.6m raised. It monitors more than 27,000 public tender sources in real time and drafts the bid. Public procurement is the largest under-digitised budget in this region. Worth an hour if any part of your revenue comes from state contracts.

Finance: a Romanian medtech funded its next product without a round. Edgefront Technologies in Timișoara took RON 7.68m of non-repayable EU money inside a RON 10.6m project, extending its CE-marked Rayscape radiology AI from chest imaging to oncological lesions across thorax, abdomen and pelvis, running to July 2029 under the STEP framework. If you hold a certified product and a 36-month roadmap, that window is a live alternative to a Series A.

Next edition soon,

Çelik