A better document can hide a training problem. In a three-month trial with patent lawyers, access to an AI assistant improved drafting, but junior lawyers didn't show an average advantage when they later worked without it. If you're buying these tools partly to develop your people, Augur's advice is to test that promise separately. Review what someone can explain and correct without the assistant, as well as the document they deliver with it. This is the Augur Dispatch for Thursday, October 8, 2026. The trial enrolled 133 lawyers at 11 American firms. 91 completed the full protocol, including the final task without AI. Independent patent experts graded their work without knowing who had received the assistant, so knowing a draft came from the treatment group couldn't influence the grade. A beautifully written paragraph can be wrong, and good typography has never signed a liability waiver. The assistant improved drafting across experience groups. The harder question came later, when lawyers had to find and fix mistakes in an existing patent without using AI. Senior lawyers who had received the assistant outperformed senior lawyers in the control group on that final test. Juniors showed no average advantage, but their scores were more spread out, with more poor and good results rather than more excellent ones. There is an important limit here. The researchers didn't test unaided skill at the start, so this isn't a before and after record of each lawyer getting better or worse. It compares groups assigned by chance, with the interpretation depending on their starting skills being comparable. The findings come from a working paper that hasn't been peer-reviewed, and Google paid the direct costs of the experiment at firms that already did patent work for the company. Three months with a small, specialized group cannot tell us what happens across every profession. Still, the distinction is useful. Earlier research with Boston Consulting Group consultants found that weaker baseline performers gained more while using AI. Leveling wasn't just a guess drawn from the average, but better performance with a tool doesn't establish what someone can do after that tool is removed. Picture a manager approving a junior colleague's polished draft. The work may be perfectly usable. The manager still doesn't know whether the colleague can explain a consequential choice or catch the same mistake next week. The study doesn't settle why the junior scores spread out. It does describe editing habits that deserve attention without turning them into a causal explanation. Some juniors worked through the document from the top, spending time on wording before reaching the claims that mattered most. Others identified a problem in a comment without actually correcting it. The researchers saw those habits among control juniors too, so we can't blame the assistant for creating them. For the team lead, that suggests a useful conversation about priorities. Ask which error carries the greatest consequence and why, rather than only whether the draft looks finished. An explanation can reveal a gap that a clean document won't. Stanford's related workplace research adds another complication. Its account describes interviews across three companies and nearly two years observing two divisions at a law firm. The division that presented AI as a way to enrich people's work used it more and experimented more. That division also provided training, mentorship, and time to explore. This wasn't a controlled test proving that a nicer slogan changes behavior. The support changed too. Scheduling a learning lunch and then filling the lunch hour with deadlines would rather miss the point. Augur recommends a bounded learning pilot, not taking the assistant away from junior staff. Choose one drafting team and make its lead accountable for checking independent judgment alongside output quality. Before starting, use a short task without AI and have a reviewer assess the reasoning. Repeat comparable tasks during the pilot, using the same rubric and, where possible, reviewers who don't know which stage they're grading. Give people protected practice and ask them to explain important edits before accepting them. At a four-week review, look at the full spread of scores, especially the weaker results, alongside document quality, rework, and time. Keep the normal review of customer-facing work in place throughout. If unaided judgment or the ability to catch consequential errors deteriorates, pause expansion and change the support. Expand only after repeated checks show stronger reasoning without sacrificing the quality of the finished work. That is Augur's proposed test, not a training method already proven by this paper. A comparable team or phased rollout would help separate the intervention from changing assignments or reviewers. The advantage, if it develops, comes from how the organization teaches and learns. The subscription itself is available to competitors. A renewal invoice is not a certificate of professional development. Averages also need care when the conversation turns from people to model bills. Anthropic has released Haiku 5.5, its small model for high-volume tasks, and says it costs around 75% less to run on average. The price depends on prompt length. Up to 100,000 tokens, the text units the model processes, the published input and output rates are 90% lower than Haiku 4.5. Above that prompt length, the reduction is 50%. Anthropic's average also accounts for a changed tokenazer that uses slightly more tokens per task. The price of each unit and the number of units both matter. Imagine the engineer responsible for thousands of short classification requests alongside a few enormous document jobs. Applying the headline discount to the whole bill would hide that difference, before even considering retries or changes in answer quality. Augur's recommendation is to compare measured cost for accepted results on your actual workload before changing routing. The published rates are useful, they don't guarantee the savings on your particular work. NVIDIA's competition results call for a similar separation between the headline and the system you might buy. Its specialized Nemotron setup scored above the programming Olympiad's gold threshold and the top human score. That run was unofficial and unsupervised. The mathematics system also cleared gold, and NVIDIA says official Olympiad graders assessed its proofs. Don't erase that distinction. Neither result, however, tells you what the ordinary downloadable model will do on your company's coding tasks. Extra training and software that generated, checked, and refined answers were part of the tested systems. There is one dated Augur call today. Google will make Gemini 4 Argon available to paid API customers by November 30th, 2026. Augur assigns 55% odds. A hit requires an official Google announcement or product page offering paying API customers access without a waitlist or partner-only program. Access confined to fairwind, government users, or an invitation list counts as a miss. Google's current fairwind policy still restricts access to approved partners. This call remains open and ungraded. Augur calls are dated predictions with explicit tests, tracked on the website so the forecast can be checked against what actually happens. Separate from that call, the signals on the clock are follow-up questions, not additional forecasts. Watch how Haiku's prompt length pricing translates into real bills and whether Argon access broadens beyond approved partners. Microsoft's Agent Lightning trains agents with the same operating software they use in deployment. Results on production tasks will tell us more than the framework's compact code alone. OpenAI's GPT-6 announcement already includes ChatGPT business and enterprise in the intelligent UI rollout. Enterprise access depends on administrator settings, and the update covers chat, not work or codecs. Check what your workplace actually enables. AI2 has published its Bulmo research in Nature and released models extending the approach beyond its original family. These models process the underlying bytes used to represent text. Broader adoption by other builders remains something to watch. For today's operating decision, pilot a separate check of independent judgment and keep document review intact. Watch the claims that still need evidence from your own workloads. Don't use a polished output as the only proof that your team is learning. What can your junior colleagues now explain or correct without the assistant that they couldn't do before? The written brief with every source is at auger dispatch.ai. Portions of the Augur Dispatch are AI generated. That is the Augur Dispatch for Thursday, October eighth, twenty twenty six. Thank you for listening.