Notebook
What a small team is learning from building a research system that works largely on its own. What works, what breaks, and what changes our minds, written down as we go.
Last updated 11 November 2026
Our approach
A working philosophy.
Capable models are now available to almost anyone. Large companies can add more of everything: compute, data, people. A small team cannot compete that way. It can compete on how the work is organized.
A small organization can redesign how it works in a day, try an unusual arrangement and drop it if it fails. We apply that freedom to AI. The question we work on is how to arrange models so they can sustain an investigation: choose a worthwhile question, challenge an explanation, run a useful test and carry what they learned into the next piece of work.
That is where much of our effort goes. Who proposes? Who challenges? What can each agent see? What gets remembered? When should the system continue, change direction or ask a person? The process is itself what we are improving, and the system takes part in improving it.
We are putting ideas about autonomy, memory and self-improvement into practice at a small scale. Some work. Some create more machinery than they justify. Both belong in this notebook.
- Give models clear responsibilities. Each has a role, access limits and a defined scope of authority. How the roles are set up
- Separate exploration from testing. Exploration helps us find questions. For registered tests, the method and reading rules are fixed before outcomes are computed. See: study 57 · confirmation
- Make room for disagreement. A challenger can maintain an objection. Agreement between models is never enough to establish a finding. See: a second opinion
- Make experience useful. Earlier failures should change later decisions. Recording a lesson is only the beginning. See: studies 45 to 47
- Keep improvement accountable. The system can propose changes and carry out permitted repairs. People set its boundaries, and improvements need evidence. See: overhead · the daily target
- Keep an honest record. Preserve methods, results, consequential decisions and failures, including corrections to earlier work. See: stopping is a result · all notes
The aim is excellent financial research, supported by tools that become more capable through use, with the possibility of discovering something investable. Nothing here is investment advice.
Finance is where we are doing that work today. We also want to understand which parts of the approach could help investigations in other fields. That is a question to explore as the evidence grows.
Learning what it needs
We want the system to become better at recognizing what an investigation needs.
When it gets stuck, can it work out whether the obstacle is missing evidence, an unsuitable method, a broken tool or a poorly framed question? Can it then take a useful next step within the authority it has been given?
Sometimes that means repairing or building a tool. Sometimes it means changing the approach, asking for help or deciding that the question cannot yet be answered.
We have seen early examples of the system diagnosing weaknesses and changing its methods. We are trying to make that ability more consistent. The measure is what happens afterwards: whether the next investigation is better designed, more informative or needs less intervention.
- The test could not answer its question. Study 45 asked whether shares trail comparable companies in the year after a priced share sale. Its own check, fixed in advance, found the one-year comparison could not detect the effect size it set. It stopped before measuring any outcome.
- It worked out why. Study 46 found the problem was the measurement, not the data: a few very large rises in the year before a sale dominated the plain average. Measured on a logarithmic scale, the same test came within the precision needed.
- The follow-up used the better test. Study 47 asked the original question on the logarithmic scale and found that shares did trail comparable companies.
- Where people came in. The research had stalled twice in between, and an AI assistant working for the founder restarted the queue under standing rules. That assistant also suggested study 45's question, re-planned on the new scale, as the likely next candidate; the engine's own planner chose it.
- What changed, besides the precision. The logarithmic scale measures how a typical holder's wealth fared, not the average return. That is a different quantity, and the write-ups say so.
Notes
Newest first. Each note says what happened, what we learned, what changed, who started the change, and where to check it.
11 Nov · Didn't work · A software update stopped the system for six and a half hours
The computer the research system runs on installed an operating-system update overnight and restarted. The system stopped in the middle of a step at 1:29 am and only started again at 7:54 am, when someone signed in. Nothing was lost or repeated: the interrupted step was marked as uncertain rather than run twice. But that night's data import ended early, and the morning's work started late.
What changed: automatic updates on that computer are paused, and the system is being moved to a dedicated server that restarts it without anyone signing in. Until the move, an outage like this shows on the home page as a gap in the activity panel, not as invented activity.
Initiated by: Founder, after the outage. · Evidence: the gap in the home page's activity record on 11 Nov
11 Nov · Worked · A day given to the queue, not to new studies
Changes to the system's own code had been taking too long to get through their checks, and studies were waiting behind them. For one day, new studies were paused so five changes could finish: reusing a finished test run instead of repeating it, running lighter tests first, landing several changes together, taking routine reviews off the shared queue, and a fallback when the next step is blocked. Four landed within the day and the fifth the day after. On the day, nine changes landed against about seven a day the week before; one day is too short to call that a lasting gain.
What we learned: when the pause ended, research did not restart by itself, because nothing re-filed the step the pause had held. It was restarted by hand and study 94 followed the same afternoon. A rule that re-files whatever a pause held is on the list.
Initiated by: Jointly. The founder asked for a lasting fix and chose the plan. · Evidence: study 94
31 Oct · Worked · The engine named how its own answer could be wrong
Study 77 found that a share's gain at the open on its payout day grows with the size of the payout. Before calling it settled, the engine pointed out a weakness in its own result: the payout amounts and the price adjustments come from the same data vendor, so a payout recorded too large would produce the same pattern. It proposed checking the recorded payouts against the companies' own filings. That check was approved and runs next.
Initiated by: Engine. It raised the weakness and proposed the check; the founder approved it. · Evidence: study 77
31 Oct · Learned · A safeguard we relied on was part of the problem
Two earlier studies used a "clipped average", which pulls in the most extreme values before comparing two groups. A method check, study 76, tested it on ordinary days with no real difference. Where the two groups' extreme values differed most, it reported a false gap in 59% of tests against a 5% target. The plain average did not fail its limit in the settings tested, though some remained unsettled. Comparisons of two groups are now decided on the plain average; the earlier studies' answers stand because their plain averages agreed, and four earlier studies are being reviewed again. The next study, designed that night, already used the plain average.
Initiated by: Jointly. The founder admitted the method check; the engine chose and ran it and proposed the rule; the founder adopted it. · Evidence: study 76
30 Oct · Didn't work · Heavier studies ran into limits set for lighter ones
Studies had been taking under three hours from start to finish. Studies 75 and 76 took more than seven, because they needed far more computing and kept hitting a four-hour limit on a single working session, and because a change made the day before left their workspace out of date. Each stop cost a restart and a review. The session limit is now six hours, a study that needs more is split into planned stages from the start, and the workspace is brought up to date in the same step that starts each study.
Initiated by: Founder. He noticed study 76 was taking unusually long; the diagnosis and fixes followed the same day. · Evidence: run times on the study pages
30 Oct · Learned · Four of seven studies settled nothing
Of studies 68 to 74, two stopped before measuring and two ended with ranges too wide to answer; only three settled anything. Stopping honestly is better than guessing, but too many unsettled studies is a planning failure. Data problems are now repaired at the source before any study depends on them; every plan states its chance of ending unsettled or stopped, and above 40% it is widened, pooled or replaced; and until at least four in any seven studies settle, one in three follows up a settled finding.
Initiated by: Founder. He pointed out the run of unsettled studies. · Evidence: studies 68 to 74 on the research page
29 Oct · Didn't work · Output fell to one study a day, then none
Studies published per day fell from three to none in four days. The causes were ordinary: two scheduling stalls, each study waiting for the previous one's wording fixes, results that had to be merged by hand, and the computer going to sleep. Wording fixes now run on their own track, records are saved before every merge, and the system has a floor of two small or one medium study a day. Limits on AI usage no longer stop the work either: backup routes take over until the limit resets.
Initiated by: Founder. He flagged the drop; he set the floor. · Evidence: the daily counts on the home page
27 Oct · Learned · Busy is not the same as deep
Ten studies in a row had used the same daily price data, and the most recent choices were small checks of known effects. Each was sound; together they were shallow. The next choices had to be harder studies of medium or long depth, and the investigation programme had to propose at least one original hypothesis within the week, saying what would confirm it and what would refute it.
Initiated by: Jointly. An AI assistant working for the founder flagged the pattern; the rule follows the founder's direction to pursue deeper research. · Evidence: studies 63 to 73
26 Oct · Didn't work · Reading company filings automatically failed its own test
A small pilot tried to pull the details of share offerings from the text of company filings. Nine of the eleven text fields failed the accuracy test set in advance; the dates and filing references passed every check. The text fields were closed rather than patched, and the study that needed them was dropped. A tool that fails its own test is not used quietly.
Initiated by: Jointly. The engine reported the failure; the founder chose to stop using the failed fields. · Evidence: the pilot's record in the engine's files
26 Oct · Didn't work · This notebook went quiet
The founder asked why this page had not changed for eight days. The rule was a note every two days or whenever something was learned, but the step that drafts the note had never been built. It now is: the engine drafts notes from its own record, and each passes the same independent wording review as a study before it is shown. Some of those notes are still waiting in review, which is why this page caught up by hand on 1 November.
Initiated by: Founder. He noticed the gap. · Evidence: this page
22 Oct · Didn't work · A stall nobody was told about
A study stopped on a step only the founder could take, and research sat idle for two hours. No alert went out, because none of the alert rules covered "waiting on a person". He found out by asking. A new alert now fires as soon as research is waiting on him, and a check every few hours brings any unanswered question to him directly.
Initiated by: Founder. He found the stall himself. · Evidence: the engine's alert records
21 Oct · Learned · From single studies to a line of investigation
Answering one question at a time kept the work narrow. The founder asked the system to take the next step on its own: to own a programme, one question pursued over several studies, with competing explanations, the evidence for each, and a next test chosen because its answer would change direction. The first programme asks whether sharp one-day price falls reflect new information or forced selling.
Initiated by: Founder. · Evidence: the programme's studies on the research page
18 Oct · Worked · A rival AI now challenges every plan before any data is seen
Each study's plan is now sent to a model from a different AI company, which attacks it without seeing any data. The plan's author answers each objection in writing, the challenger rules on the answers, and the plan is then locked. Since mid-October every new study's design has gone through one; 19 exchanges are on record. It typically costs under a dollar and changes the plan in several places.
Initiated by: Founder. He approved the pilot. · Evidence: the design dialogue on each study page
17 Oct · Learned · A second opinion that is starved of context says nothing
A second model ranks candidate studies blind. Three times it declined to rank, and each time it said why: it had not been given what earlier studies concluded or how the data was built. It was right. We now give it that context, and if it still declines, the system reports why instead of changing the packet again.
Initiated by: Jointly. The engine reported the declines; the founder approved the fix. · Evidence: the engine's records; the change was built on 17 October
16 Oct · Didn't work · Overhead was invisible until we measured it
The system was working only about 60% of the time it was running. The rest went to a pause of about three minutes between tasks and to frequent restarts to load its own updates. Time when the system was off was also shown as idle. We now count “down” separately, start the next task first and batch restarts. By the next day the system was working about 72% of the whole day, down time included.
Initiated by: Founder. He asked why the activity strip showed so much idle time; the measurement and the fix followed. · Evidence: the activity strip on the home page
16 Oct · Learned · Exploring the data mostly rediscovers what is already known
When the system began looking for patterns on its own, the first four it found were effects already documented in published research. That is reassuring about the method and a warning about novelty. Every candidate now says whether it repeats published work, extends it with a stated open question, or is new. For now, while the next studies must come from exploration, a plain repeat is not chosen; replications remain possible as checks of the method.
Initiated by: Jointly. The overlap was flagged on the founder's side; the founder set the rule. · Evidence: the exploration cards and their literature notes in the engine's records
16 Oct · Didn't work · Access is a single point of failure
An expired sign-in stopped every step that needs a model for about three hours, while the mechanical steps kept running. Nothing was lost, but nothing was learned either. The system now holds model work and alerts a person at once, and a fallback route is being built.
Initiated by: Jointly. The engine diagnosed and reported the failure; the founder signed in again and approved the safeguards. · Evidence: the gap in the activity record for that afternoon
16 Oct · Worked · A study born from exploration held up
Study 57 tested a pattern found by exploring 2013 to 2017 on 2018 to 2022, years that exploration never read. It held: shares with the largest overnight gains in one month kept gaining overnight the next month, by well over the size fixed in advance, and gave the gain back during the day. Among the most heavily traded shares the pattern was much smaller, and its range holds zero, so whether it survives there is not settled; the write-up was corrected on 7 November 2026 to say so. Those later years are still development data that earlier studies in the programme have used, so this is an exploratory result, not a confirmation. This is the loop we want: notice something, propose explanations, test on data the idea was not found on.
Initiated by: Engine. It found the pattern, proposed the study and ran it under the founder's instruction that the next studies come from exploration. · Evidence: study 57
15 Oct · Didn't work · A target of studies per day pushed toward shallow work
For about a week the system worked to a daily number of studies. Some days it fell well short and some days it exceeded it, but the pressure was always toward the next study rather than a deeper one. We dropped the target: each question now needs a literature review, a check that the data can answer it, and a power estimate before it can become a study.
Initiated by: Founder. He set the target and dropped it. · Evidence: the research list, by date
15 Oct · Learned · Confirmation is the hard part
Our strongest result so far, from study 47, could not be confirmed on any data we hold: every slice had been touched by earlier work or was biased by how it was chosen. It remains exploratory. A clean test needs data that does not exist yet. Exploring is cheap; confirming takes time, and we would rather say so than shortcut it.
Initiated by: Jointly. The founder asked for a confirmation route; the engine audited the data and found none. · Evidence: Limits
14 Oct · Didn't work · “Do this, then that” left the system idle
An instruction to finish one piece of work before starting another was read as a strict sequence. The system waited about five hours while other work was ready. Priority now means order, never a reason to sit idle.
Initiated by: Founder. The instruction came from the founder's side and so did the correction. · Evidence: the activity record for 14 October
12 Oct · Worked · A known-answer test passed
Twelve questions with answers already known from published research ran through the same steps as every study, blind to which was which. The process reached the published direction on every clearly established effect and called no false effect. It checks the method, not markets.
Initiated by: Jointly. The founder commissioned the test; the engine designed, sealed and ran it. · Evidence: How it works
Since 26 Sep · Worked · Stopping is a result
Six studies stopped themselves because their own checks, fixed in advance, found they could not detect the effect size they set out to test, among them study 45 and study 53. They are published like any other result.
Initiated by: Engine, under rules the founder approved. · Evidence: Research
Since 26 Sep · Learned · Most studies change how we work, not what we believe about markets
Of 93 studies so far, 51 changed a research rule or plan and 3 checked a method; only a handful concluded something about markets. Early research on a new system is mostly research about the system itself.
Initiated by: Engine (the counts come from its records). · Evidence: Research
Ongoing · Learned · People are still the slowest step
The largest delays have come from waiting on a person's decision, not from the machine. We batch questions, write down standing rules once, and link each rule to the decision it came from, so the same question is not asked twice.
Initiated by: Founder. · Evidence: the human steps shown on the home page's activity strip
Open questions
What we do not know yet, and what would change our minds.
- Can the system judge which questions are worth asking? It can carry a well-posed question through to a reading. Choosing a question worth the effort is harder, and we do not yet have a good measure of it.
- Does experience change later work? Lessons are recorded. Whether they make the next design better, rather than only longer, is what we are trying to measure.
- Will an exploratory finding hold on data that did not exist when it was found? That is the test that matters for anything investable, and it takes time.
- How little human attention can it need without losing care? People are still the slowest step. We want fewer interventions, not fewer checks.
- What carries over to other fields? Some of the arrangement is specific to finance. We want to learn which parts are not.
If you have a view on any of these, write to us.