A small team of AI agents and automatic checks, the loop every study follows, and the statistical methods it uses.
The machinery
The system is a small team of AI agents and automatic checks, answerable to one person. Each part has a narrow job, passes its work to the next, and cannot approve its own work.
One person sets the mandate and holds budgets, access to locked data and anything published.
AI agents with separate duties choose and design studies, build and run them, and read the results. A second model from a different provider proposes questions where the agenda is thinnest and, since 17 October 2026, also challenges study designs alongside the first challenger.
Independent reviewers, AI and automatic, check every change and every published sentence. No part approves its own work, and none can raise its own limits.
Delegated
What the agents may do alone
Sequence and run authorised work
Launch builders and reviews, and merge reviewed changes
Make routine repairs
Choose and run studies on open development data
Run pre-authorised safety probes
Reserved
What only the person decides
Budgets and any change of authority
Access to locked or protected data
Reopening a frozen or closed experiment
Publishing anything outside the project
Weakening a security control or altering protected evidence
Running itself: autonomy of operation
There are two kinds of autonomy. Autonomy of operation asks whether the system can start, run, check and recover its own work without a person. Since 22 September 2026, largely yes. Autonomy of authority asks what it may decide on its own: very little. Budgets, authority, locked data and publishing stay with people, and the machine cannot raise its own limits. This section is about the first kind.
It chooses its next task, writes the instructions, builds, has the work reviewed, tests, records the outcome and restarts onto the approved version. When something fails it stops at the boundary, records why and asks the owner; it does not work around a refusal. Failures stay on the record: from 22 September to 1 November 2026 about one task in four ended without completing, almost all of them cancellations.
The record · 22 September to 1 November 2026
5,173tasks completed by the machine
2,005executive-agent sessions: planning, reading results and recording decisions
496builds completed, with 458 independent code reviews
1,757merges, saves and pushes, each after the tests passed
254independent reviews of text written for publication
16,907automated tests passing on 2 November 2026, against 7,600 on 30 September
1,509 / 99tasks cancelled, and tasks that failed or stopped, out of 6,781 that ended
453written rulings and decision records on file, on authority, budgets, priorities, method and publishing
What still needs a person
Some actions are refused by its permissions; each becomes a task for the owner.
Some stalled work still needs a step from outside the engine to clear. A programme to remove those steps began on 1 November 2026.
Budgets, authority, locked data and publishing are decided by the owner, in writing.
No one outside the project has yet run it on their own question.
Counted on 2 November 2026 from the machine's own task records, 22 September to 1 November 2026 (UTC days), by when each task ended. A task is one recorded unit of work: a session, a build, a review, a merge, a save or a scheduled run. The test count is the full suite's last recorded pass.
The process
Every study follows the same four steps. Separately, a longer path leads from an idea to an independently confirmed claim. Only its first stages have been demonstrated.
Step 1
Choose a question
From a standing agenda of candidate questions, screened by a written selection policy for worth, variety and whether the answer is already near-certain. A separate planning session designs the study, states competing explanations, and writes down the answer it expects before it runs.
Step 2
Fix the rules
Before any result is seen, the study records what each possible outcome would mean. Its inputs must pass an automatic check before it can launch.
Step 3
Run and review
A builder runs the study on open data. The independent reviewer checks the exact change before it is accepted.
Step 4
Record the reading
The result is written down as registered, including when it is inconclusive or when the study's own checks fail. It then informs the next question.
Through the glassThe same seven stations, for every study.
1QuestionChosen from the standing agenda. The expected answer is written down first.
2Plan lockedMethod and reading rule recorded and fingerprinted before any outcome is computed.
3BuiltCode and tests written by a builder that cannot change the plan.
4RunOn open development data only, with identity checks against earlier studies.
5ReadThe executive agent reads the result against the rule from station 2.
6ReviewedAn independent reviewer checks the exact change. The full test suite runs before the merge.
7PublishedQuestion, reading and what changed, in the machine's own words, on this site.
Locked data: not opened at any station.Marked from the public feed when the research system is on.
Patterns are explored, explanations compared and designs tested. Repeated looks are allowed here.
Locked · confirmation data
Where a fixed claim faces its test
Kept back for a test fixed in advance, with every earlier exposure recorded, so that evidence already seen cannot be treated as fresh.
#
Stage
Status
What that means today
Exploration · open data
01
Question
Demonstrated
The agent chooses and ranks its own questions. It records the answer it expected before each study, so surprise is measured; whether experts would rank its questions the same way is not.
02
Open-data study
Demonstrated
Studies run, are reviewed and reach a recorded, exploratory reading.
03
Candidate
Sealed · quarantined
The second model has proposed test candidates. They are sealed and set aside until a test is authorised. A sealed candidate is not a finding.
04
Feasibility
Demonstrated
Checks of event definitions, record completeness and statistical power have run on several lines and opened or closed each.
Confirmation · locked data
05
Evidence allocation
Held
Reserving locked data needs authority that has not been given. The original locked data is not usable; each data source added since carries a slice sealed on the day it arrived, with an access log from the first day.
06
Frozen test design
Not demonstrated
The experiment on machine-designed tests closed before its convergence test.
07
Controlled execution
Isolation unproven
Access checks found no open route, but not every route has been tested.
08
Check and verdict
Calculations checked
Verdict calculations matched known answers. Stages where AI models judge were not tested.
09
Owner and publication
Explicit authority
Nothing becomes public without the person's approval.
The research toolkit
The methods used to measure an effect, judge how uncertain it is, and check that the tests themselves behave as intended.
Code measures the data. AI decides what may be interesting, and writes down why.
1 · Measuring an effect
What exactly is being compared
Abnormal return
A stock's return minus its benchmark's over a fixed window: three days, five trading days, or roughly three months, set before the study runs.
Same-company baseline
The event window compared with windows drawn from the same company's ordinary history. It separates the event from how that stock usually behaves.
Point-in-time universe
Only information available at the time is used, and companies that later disappeared stay in, so survivors do not flatter the result.
Event audits
Before any return is studied, the event itself is checked: duplicates removed and definitions tested for how often they are wrong.
2 · Judging uncertainty
How sure the estimate is
95% intervals, both tails
Each estimate carries an interval, and each tail is held to its own error target of 2.5%, so a test cannot look calibrated overall while failing on one side.
Company-clustered inference
Events from the same company are treated as related, not as independent trials. Where time periods cluster, intervals are built by resampling whole months.
Bootstrap and exact sign-flip tests
Resampling and exact tests are used where textbook formulas break down, especially with few companies or skewed returns.
Effective sample size
Overlapping observations are counted for what they are worth, so a large count does not pass for independent evidence.
3 · Testing the tests
Checking the tools before trusting them
Simulated false-alarm rates
Each statistical test is run thousands of times on data with no effect, up to 10,000 draws per case, to measure how often it cries wolf. One recommended test raised false alarms above its 5% target, and the guidance changed.
Planted known effects
Effects of known size are planted into development data to measure how often they are found. Small effects were often missed, which is why study plans must now state the smallest effect they could detect.
Calibrated thresholds
Where a test misbehaves, its decision threshold is set by simulation for that design instead of taken from a table.
Planted defects
Defects of known kinds are planted and the system is scored on catching them unprompted. First run at the end of September 2026: look-ahead and post-cutoff leakage caught reliably, clean data rarely flagged, duplicated and extreme rows mostly missed.
Known answers from published work
Blind verdicts on 211 published stock-return predictors, sealed before the answer key was read (30 September 2026): 92% of those expected to hold were supported, none contradicted. On 12 October 2026, twelve questions with known answers ran blind through the full study process; it found all six clearly established effects in their published direction and called neither no-effect case an effect.
4 · Guarding against false discovery
Not fooling itself
Rules fixed in advance
Each study records what every possible outcome would mean before it runs. Amendments made later stay on the record.
Open and locked data
Ideas are explored on open data. Confirmation is reserved for locked data with every earlier exposure logged.
Train, hold out, then track forward
Searches are fitted on one period, re-checked on a later one and then followed live, and survivors are compared with what chance alone would produce. None has been promoted.
5 · Reproducibility and accounting
Every recorded calculation can be traced and rerun
Fingerprinted results
Result files are hashed when written, and reruns reproduce them byte for byte.
Automated tests and independent review
Several thousand automated tests guard the code, and an independent reviewer checks every change before it is merged.
Simulated portfolio ledger
The paper portfolio's accounting reconciled to the cent in offline testing and passed 98 of 98 crash-and-recovery checks. The book has been marked each weekday evening since 25 September 2026. Its performance is not yet published here; when it is, it will be shown after costs, against the S&P 500 including dividends and against an exposure-matched comparison.
Has the whole procedure been evaluated?
Its parts, yes; the whole, not yet. Each test was checked on simulated data for false alarms, detection power and sample size before being admitted. It has not yet been reviewed by an independent statistician.
Reading a result
One five-day pattern was measured two ways. Against the market, the whole interval sat below zero. Against the same companies' ordinary weeks, the estimate shrank and the interval crossed zero. The choice of comparison changed the reading, so the pattern was recorded as unresolved, not as a finding.
Worked example · four studies on one question
Question: after a listed company files a priced prospectus to sell shares, do its shares trail comparable companies?
Study 45 stopped at the power check. Over the year, the smallest detectable effect was 19 points against a bound of 10, so no return after any sale was read.
Study 46 diagnosed why, still without reading outcomes. 14 of the 16 most extreme cases were the selling company's own enormous rise before the sale. Measuring on the log scale cut the detectable effect to 5.8 points.
Study 47 re-registered the question on the log measure. The month after the sale is 4.4 points behind (interval 3.4 to 5.5), and the year 0.25 behind in log wealth (0.19 to 0.31). The run-up slope is positive, which rules out timing as registered. The same companies' ordinary dates show no gap. All fourteen sensitivities agree.
Study 52 split the result by industry. Drug and biotechnology companies and all others both trail, and their difference includes zero, so the result is not an artefact of one sector.
All four studies are on development data, and none is confirmed on data set aside for confirmation.
Defects in its own data, methods and earlier results. Any one of them, uncaught, would have produced a wrong plan or a false lead.
Every event counted four timesA data-handling fault counted each real event about four times over. Every plan sized on the inflated number would have been wrong.
A pattern that used the futureA candidate replicated on other companies, but its grouping used information from later years. Replication alone would have passed it.
Mislabelled eventsMost flagged "first dividends" were not first dividends. A standard study would have measured returns on the wrong events.
A comparison that flattered the resultA five-day gap looked real against the market but not against the same companies' ordinary weeks.
Survivors of chanceA few stock pairs survived a holdout, not many more than chance alone would produce. None survived forward tracking. Promoting holdout survivors would have produced false leads.
Tests that cried wolfTextbook-style tests raised false alarms more often than their 5% target, and more still in one subgroup. They would have overstated significance.
Whether a standard workflow would have missed these has not been measured; the comparisons with common practice are reasoned, not tested head-to-head. A planted-defect suite, where the system is scored on catching known problems unprompted, is the start of measuring it; its first run, at the end of September 2026, caught look-ahead and post-cutoff leakage reliably and missed most duplicated and extreme rows.
The platform
Three activities share one data foundation of security histories, prices, corporate events and company identities. They keep separate purposes and separate rules. A desk observation does not become a trade or a finding on its own.
Observe
Relationship desk
Runs on weekday mornings. Follows pairs and groups of stocks and unusual moves, and turns what it notices into questions.
Produces observations and research questions
Investigate
Research engine
Chooses questions, runs studies on open data and keeps every reading, including failures and inconclusive results.
Produces exploratory readings and method changes
Simulate
Paper portfolio
$100,000 in simulated cash, with realistic costs. Each weekday morning a trading agent reviews the candidates the desk noticed the previous session and writes a decision for each, with its thesis, the competing explanations and the evidence it would need. A decline is recorded with its reason, and the declines are reviewed weekly. The book is marked each evening. A pilot in which the agent also trades on the research engine's own findings has been approved and is not yet built.
No real money · not evidence that the research works
What it studies
Markets are the test bed. The skills being tested are general ones: noticing when a research result does not deserve belief.
The current universe
US public equities, after company events
The unit of study is one company and one event: a large one-day move, a stock split, a change to a regular payout, an acquisition, or a pair of shares that move together. Outcomes are measured over windows from three days to about three months, against a benchmark or the company's own ordinary periods. ETFs, funds, preferred shares, warrants and over-the-counter securities are excluded. A public dataset of published findings and the record of company filings have since been added.
It also studies its own methods: whether its statistical tests and tools behave as intended.
What it looks for
Defects in empirical research
These are not specific to markets. They appear in clinical studies, credit models and online experiments too.
Look-aheadLeakageWrong benchmarkOne subgroup carries the resultFails on new dataHidden dependenceMeasurement errorA simpler explanationThe cheapest falsifying test
What it is for
Two things are tested at once: whether the engine produces research that holds up, and whether it can run, repair and improve itself with less human help. The longer aim is more worthwhile investigation from the expertise an organisation already has. Whether it achieves any of this is measured, not assumed.
The unfinished work
What research organisations never get to
Claims they rely on but have never independently checked
Models whose assumptions need continual re-examination
Promising ideas that wait because experts have limited time
Analyses that are costly to reproduce when a conclusion is challenged
The success test
Better-supported decisions, with less expert effort
Its value does not rest on one dataset hiding a profitable effect. It rests on finding useful effects, rejecting misleading claims, catching defective measurements and establishing when a question cannot yet be answered.
It will be judged by decisions supported per hour of expert time, by how often independent experts uphold its conclusions, by its catch rate on planted defects, by cost per accepted conclusion, and by how often a person has to step in.
Layer
What it is
Status today
First application
Empirical research in US equity markets
Running
Capability shown
Self-directed investigation: choosing questions, replication, method checks and follow-up work
Self-directed studies: first on its own methods, then a blind test against published findings and market questions under a variety rule. The record →
Institutional product
A research system an organisation can direct, inspect and run under its own evidence and authority rules
Not built yet
Longer-term ambition
The same capability in other fields, through separately tested data and method integrations
Planned; an audit mapped the work needed
What is not yet shown
The machinery runs. Whether its research is reliable, and better than a careful analyst with simpler tools, is still open.
Open limits
Confirmation is not available yet. Under the programme's own rules, no finding in the record has been confirmed on data set aside for that purpose, and every result is exploratory.
The path from a candidate to a frozen, confirmable test has not been demonstrated.
The stages where AI models judge and design have not been tested against known answers.
An earlier test of whether the system notices planted effects failed its criterion. A planted-defect suite, first run at the end of September 2026, catches look-ahead and leakage reliably and misses most duplicated and extreme rows; it is self-assessed, on development data.
No independent expert has yet reviewed its studies.
When the research system is offline, research and the live view stop.
The isolation of study sessions has not been fully proven.
Since 17 October 2026, study designs are challenged by models from two providers. The two challengers were compared on three designs on 11 October 2026; no larger assessment of which catches more has been made.
Some stalled work still needs a step from outside the engine to clear, by the owner or by the AI agent supervising operations. A reliability programme began on 1 November 2026 to remove those steps and measure how often they are needed.
Intake of new data sources into sealed confirmation slices has been stopped by its own gate since 29 October 2026: files arrived that the gate's rules did not cover. The owner has accepted a revised gate, and it took effect on 3 November 2026 once the change that applies it was built and tested. After the owner lifted the stop, the run on the night of 5 November 2026 stopped again at its first check: the data vendor's table lacks a date column the import was built to use. On 7 November 2026 the owner settled which date field takes its place, and the change that applies it took effect on 9 November 2026. The run on the night of 10 November 2026 stopped partway when the vendor's service refused a request, and intake is stopped again while that is examined.
In the paper portfolio, an exit fills at the next opening price as traded, so a stop does not cap a loss, and an entry order is good for one opening only. One order was recorded as missed and has not been filled after the fact.
Output fell in the week to 10 October 2026, from eight studies a day to between none and three, while most of the engine's time went to building changes to itself. A schedule pilot of research by day and building by night began on 10 October and ended on 14 October. Research and building now run side by side around the clock, and on 15 October the daily study target was dropped in favour of evidence before volume.
Outside the mandate
No stock picks, price targets or investment advice
No portfolio construction offered to anyone
No claim of a proven market edge
No claim yet that it works outside markets; that is an ambition
External claims intake
Submissions of outside claims stay suspended until the system passes a test on data it did not analyse itself.
SUSPENDED
Corrections
When we describe something wrongly, the description is corrected here, last reviewed 5 November 2026. The original result stays as recorded. When a later check changes a study's own answer, the correction is dated on that study's page; all of them are listed together on the plain record.
Study 50: an answer changed from Yes to Not settled
Study 50 first answered Yes: shares fell behind comparable companies after a notice that a report would be late. On 20 October 2026 its answer was changed to Not settled. The shortfall rested on a clipped measure that a later method check (study 56) found reports a difference that is not there too often on this design, and a re-test of the same notices on a sound measure (study 60) did not establish it. The registration and the original result stay on record, with the correction dated beside them. Study 50 in the plain record.
Access testing: an expected control was not a breach
An earlier description called one allowed case an open access route. It was a control that was meant to succeed.
Known-answer test: data, scope and error range
Effects were planted into real development data, not wholly synthetic data. Only the calculation stages were tested. The false-positive range was slightly wider than first reported.
Calibration: an earlier remedy was narrower than described
An alternative test met its targets for the pooled sample and one event subtype, not in general. For the other subtype no tested method did.
Early question generation: nine reasoned declines and one empty reply
Ten early calls did not all decline for the same reason. Nine gave reasons, which traced to defects in our own inputs. One returned nothing.
Method comparison: the unit is a week
A study was first described as comparing one day with a five-day average. It compares one event week with the same company's ordinary weeks.
Question origins: owner-supplied work is excluded
A review supplied by the owner was wrongly counted as a question the system originated. Counts now include only questions the system raised itself.
Public feed: two studies were missing from this site
Studies chosen from the new question map were recorded in a store the public feed did not read, so two concluded studies were absent from these pages for part of 30 September 2026. The feed now reads that store; the studies appear with their original dates.
Recovery: which part fixed the fault
A repair was first credited to the automatic task runner. The AI agent directing the work cleared the fault and completed the step on retry.
Study 9: a heading overstated the result
A case file, then on the research page, was headed as if the study had settled whether a five-day pattern after stock splits is real, and tagged as "holds up on development data". The system's own recorded answer is narrower: the stricter test still finds the gap on development data and timing does not explain it, but whether few or many companies carry it is unsettled. The heading now uses the system's own title and answer.
Study 44: held by this site's own screen
Study 44's public text was withheld from this site between 8 and 10 October 2026 because the site's screen for internal terms matched an ordinary word in it. The screen was narrowed; the study's record was never changed.
Studies 45 and 46: a design gap that cost a second study
Study 45 stopped at its own power check because its one-year comparison, in plain returns, could not detect the effect size fixed in advance. Earlier studies had already shown that company returns carry extreme values, and a fallback measure could have been registered in advance. It was not, so study 46 was needed to establish the fix. Study plans are to register such a fallback before any data is read.
Reproducibility: what "rerun" covers
Pages said every result can be rerun. Recorded calculations rerun exactly from their stored inputs, code and settings; the steps where an AI model judges are recorded but not replayed, since a fresh model response is not guaranteed to be identical.