Research
Every study the research system chose, designed, ran, reviewed and wrote up itself, newest first, including the ones that stopped. Each opens as a paper: the question, the rules fixed before the result, the data and method, the results with their figures, the limits, and how it was made. The same record without scripts is the plain record.
What the record shows so far
A short reading of the studies below, in plain words. Every result here is exploratory: none has yet been confirmed on data set aside for that purpose. Unless a study says otherwise, they cover US-listed shares priced at five or more, in 2018 to 2022. Written on 11 November 2026, after study 93.
Falls nobody explains fare better than explained ones
After a one-day fall of a tenth or more, falls with no company filing to explain them did better over the next three sessions than falls with one: by about 1.5 to 1.6 percentage points, and about 0.8 to 0.9 when matched on trading volume. Explained falls tended to keep falling from the next opening price.
Overnight gains are handed back during the day
Shares that gained most overnight kept gaining most overnight the next month, but gave it back between the open and the close, leaving no full-day premium. The most speculative shares showed the same split, week by week.
Reversals live in quiet shares
A week's biggest risers trailed its biggest fallers the next week, and the day's biggest fallers beat its biggest risers the next day, among the least traded third of shares. Among the most traded shares the question is still open.
Where the price closes matters for the next session
Shares closing at the day's high did about 0.25 percentage points worse the next session than shares closing at the low, and shares closing just below a round-number price did about 0.12 worse than just above.
The rebound is mostly gone by the next open
Low-priced shares that fell steeply partly reversed, measured from that day's close. Measured from the next morning's opening price, the rebound was gone, and the overnight step was accounted for by the usual overnight move and the gap between buying and selling prices.
Why so many studies "changed a rule or plan". In most studies, the main thing that followed was a change to how later studies are run: a check added, a rule tightened, a plan corrected. That is the system correcting itself on the record. Each study's page says what changed and why; "inconclusive" and "stopped by its own checks" are published the same way, not left out.
Every study
Search the questions, answers and methods, or filter by how a study ended and by line of work. A study's title and written account are published once they pass an independent text review; until then it is listed by number.
No study matches. Clear the search or the filters.
The list is not available right now. It is read from the research system's own record each time this page loads. When the record cannot be reached, nothing is shown rather than an old list.
Confirmation is not available yet. Under the programme's own rules, no finding in this record has been confirmed on data set aside for that purpose, and every result here is exploratory. What the set-aside data contains, its size and its boundaries are not published.
Research map
Each row is a line of work; each dot is a study, placed by its number. Studies in one row build on each other: a stopped plan, its diagnosis, a redesigned test, a follow-up. Open any dot to read the study.
How the research is organised
How questions are found and chosen, how a line of work is pursued and closed, and what counts as an ending. The mathematics each study uses is set out on How it works.
The agenda: where questions come from
The research system keeps a map of candidate questions. Each one records its origin, the data it needs, whether that data is available now, and the decision it could change. Most are proposed by the system itself, often as the registered next step of an earlier study. Some come from the owner, and some from question lists that predate the current system.
A written selection policy screens each candidate before it can run. It asks whether the answer would change a rule, a plan or a decision; whether the data on hand can reach the needed precision; and whether the question repeats one already answered. The map is refreshed every morning.
Choosing the next study
A choosing session picks one admissible question under fixed variety rules:
- no more than three of any five consecutive studies from one line of work;
- a question about markets at least one time in three, when one is admissible;
- never the same line twice in a row;
- when the leading candidates are too close to separate, the choice between them is drawn at random by code, and the draw is recorded.
The study's design is then written by one of two planning models, also drawn at random and recorded, so the two can later be compared on the studies each designed.
Before anything runs, a separate reviewer that never sees results challenges the design. The plan then fixes the rules for reading the result (see the mathematics on How it works).
Kinds of study
| Kind | What it asks | Example |
|---|---|---|
| Is it true | Does a measurable effect follow an event? | Do shares trail comparable companies after a priced prospectus? |
| Why | Which of several explanations fits? | Is a shortfall pressure from new supply, managers' timing, or neither? |
| Method check | Does a statistical test behave as claimed? | Does a bootstrap test keep its false-alarm rate at other sample sizes? |
| Measurement | Can the events be found and dated reliably? | Can filings be tied to every share in the sample? |
| Diagnosis | Why did an earlier study stop? | Why could a one-year test not reach its precision? |
Lines of work and their verdicts
Questions on one subject form a line: share offerings, payout changes, stock splits, company filings, large one-day moves, paired shares and others. After three studies, a line may not have a fourth until the system writes its verdict from the record:
- continue, with its registered next question;
- close, because the line is answered or cannot be answered;
- rest, while it waits on data or a diagnosis.
The verdict adds no question of its own. It follows what the line's studies registered as their consequences.
Example: the share-offerings line ran a stopped study, a diagnosis, and a re-registered test that found a persistent shortfall. Its verdict was to continue with one more study, an industry split. That split found the shortfall general across industries, so the line returns to a verdict.
The life of one study
- Chosen from the agenda under the variety rules.
- Designed with:
- competing explanations, each with a distinctive prediction;
- an expected answer with probabilities;
- reading rules;
- a list of sensitivities;
- a power plan.
- Challenged blind, then registered: the plan is committed before any outcome exists.
- Checked for power on pre-event data. A study that cannot reach its precision stops and says so.
- Run by code, with every data read logged and every result fingerprinted. A rerun must reproduce the result byte for byte.
- Reviewed by an independent code reviewer.
- Read against the registered rules, and only those rules.
- Written up in plain words and checked by an independent text reviewer.
- Published on its own page.
What counts as an ending
| Ending | Meaning |
|---|---|
| Changed a rule or plan | The result changed one of the system's written research rules or plans. |
| Method checked | A method was tested and kept. |
| Concluded | The question was answered, with nothing to change. |
| Not supported | The expected effect did not reach the size fixed in advance. |
| Inconclusive | The data could not tell the explanations apart. |
| Stopped by its own checks | A check fixed in advance failed before any result was read, for example the power check. |
Every ending is published, including the stops. All results so far are on development data. Confirmation on data set aside for that purpose has not yet been used. None of it is investment advice or a trading recommendation.
Before the public record
Work from before the system began choosing its own studies on 26 September 2026, kept as it was written then, including what failed.
Show the earlier entries
- Check passed · with limits
Its calculations faced known answers.
Effects of known size, including zero, were planted into development data. Every one of 160 verdicts matched an independent recalculation. The same test showed that small effects are often missed.
160 / 160verdicts matchedLimitOnly the calculation stages were tested. The stages where an AI model makes judgements were not.
- Target closed
"First dividend" often wasn't.
Before studying what happens after a company starts paying dividends, it checked whether the event could be identified at all. Most flagged "first dividends" were not genuine first payments. A stricter definition was still often wrong and left too few events. The line was closed without any return study.
- Review completed
The desk was judging incomplete groups.
A review of 33 flagged desk issues explained 29 of them: truncated member lists, chained groupings, common market patterns and duplicate alerts. Four membership questions remain open, and changes to the desk's rules wait for the owner.
- Unresolved
After a stock split, a five-day pattern weakened under a fairer comparison.
Against the market the gap looked real: its interval sat below zero. Against the same companies' ordinary weeks it shrank and its range included zero. It was also concentrated in a limited period and differed across event subtypes. The line was rested at unresolved.
- Sealed · not measured
The second model proposed three test candidates, now sealed.
After its instructions and inputs were repaired, three registered calls each produced one candidate. All passed six admission checks and were sealed into quarantine. An earlier batch of ten calls gave nine reasoned declines, traced to defects in our own inputs, and one empty reply.
LimitThree proposals of one similar question are not three pieces of evidence. Their target effects are assumptions, not measurements.
- Not met, then at the limit
A reviewer that can't see results caught a planted flaw, and raised some false ones.
A claim needed a +0.5% effect, but its executable threshold had been quietly changed to zero. The outcome-blind reviewer caught this in all three runs. The first evaluation still failed, because it also objected strongly to the untouched control. In a follow-up, 3 of its 15 serious objections were false.
3 / 3planted flaw detected3 / 15false objections, exactly at the registered maximumLimitAll three false objections came from one error in the frozen record. This is not a general error rate.
- Not met
Machine-designed tests did not reach their convergence check.
Three attempts to have the design stage produce an approvable test stopped early. Interface faults were repaired afterwards, but the scored result stays closed. A new attempt would be a new experiment.
- None promoted
A large stock-pair search. No relationship promoted.
A search for pairs where one stock's move predicts another's found a few that held on later data, not many more than chance alone would produce. Tracked forward, none was promoted.
LimitThis does not show that such relationships never exist.
- Registered effect excluded
Case #1: do stocks rebound after a large one-day fall?
A pre-registered test of a published-style claim: a rebound of at least +0.50% over three days, relative to the market.
The typical event did not rebound. The whole interval sat below the registered +0.50%, so that rebound was excluded on this construction. The record also keeps nine amendments made after the design was frozen, while the outcome was still unknown.
LimitThis does not show that reversals never happen, or that betting against them would work.
- Refused
A convincing pattern turned out to use future information.
A candidate held up on other companies, until review found it grouped past events using information only known years later. It was refused. In the same period the system noticed that events were being counted about four times over.
- Not admitted
An AI labeller agreed with a person, but not with itself.
It met its agreement thresholds against the owner's labels, but fell just short of its bar for giving the same answer on repeat runs. It was not put into use.
- Not supported
The founding idea failed its first test.
Mechanically buying after large price falls missed all four criteria set in advance, and the typical event kept underperforming the market. The result narrowed the idea the project started from.
- Not met
Finding a planted signal is a separate test, and it failed.
In an earlier setup, known effects were planted to see whether the research process would notice them. The last graded recovery missed its criterion. Many planted effects were lost before they became questions.
LimitThis measured an earlier configuration, not the current executive agent. The later known-answer test checks calculations once a question is set, not whether the system finds the right question.