01 · RESEARCH
Market & strategy testing
Does the rule actually work? We define pass criteria in advance, test against a neutral baseline, and report the limits alongside the result.
Data & market research
We turn assumptions into evidence—and say where it stops.
Independent research on markets and business decisions. Rules written before the data is opened, a slice of history sealed until the end, and failures kept on the record.
pre-registeredheld-out samplesneutral baselineshuman × AI
EVIDENCE BEFORE VERDICTthe map is inspectable · the judgement remains human
01 · WRITE THE RULEBefore opening the result.
02 · SEAL A SAMPLEKeep one honest test untouched.
03 · KEEP THE SCARSFailures remain part of the record.
What we do
Five things,
done properly.
Different assignments, one discipline: make the question explicit, preserve the trail, and keep the limit beside the answer.
02 · BEHAVIOUR
What people do, not what gets reported
Which promotion truly lifted sales. Which signals precede a customer leaving. Whether field activity matches what was written down. Whether a long-standing rule still holds in recent data.
03 · DASHBOARDS
Live data dashboards
Working instruments, not screenshots: live feeds, automated checks, and interfaces built so a non-technical reader can act on them.
04 · AUTOMATION
Research automation with AI
Pipelines where AI does the labour and a human keeps the judgement—with the guardrails that stop speed from producing nonsense.
05 · AUDIT
Leakage & error detection
Money leaves a business quietly: approvals made outside policy, assets idle but still carried on the books, figures that quietly stop reconciling. We rebuild the numbers from source records and show where they diverge—with the trail attached.
How we work
Three rules that make a result worth reading.
Not ceremony. Three gates that stop a convenient story from dressing itself up as evidence.
GATE 01
Pre-registered
Definitions, pass criteria and measurement are written down before the data is opened. If the bar moves after seeing results, the test proves nothing.
GATE 02
Held-out sample
A slice of history is sealed during development and opened once, at the end. It is the only honest way to know whether a finding survives outside the room it was born in.
GATE 03
Failures stay on the record
Rejected ideas are kept with the reason they failed. A map of dead ends is an asset—it stops the same money being spent twice.
Speed is not the achievement. With modern tooling, testing is cheap—which means coincidences that look like discoveries are cheap too. The guardrails are the product.
Selected work
Built from real systems.
These are sanitized views of working research and operating systems—not invented demo data. Identities, private figures and proprietary logic are omitted from the public build.
FIELD STUDY · PUBLIC CASE
When an agent became “gun-shy”
A four-day behavioural shift in a persistent AI-assisted research system, reconstructed from timestamped logs. Reported as one field observation—not a causal result—with competing explanations left intact.
SYSTEM · REAL SNAPSHOT
A working memory that compounds
The live graph scans notes, research, tools and routines into one provenance-aware network. The counts shown are a dated system inventory; private node contents are not part of this page.
FUNDAMENTALS · PUBLIC SOURCE
Following each 100 units of revenue
A real annual-report chain, normalized so the business structure is visible without publishing a company identity. The private receipt maps every displayed index back to its filed source line.
OPERATING SYSTEM · ANONYMIZED
A private dashboard with receipts
One source of truth, explicit reconciliation and visible exceptions before a human call. The architecture is real; company identity and every operational figure remain outside the public source.
PUBLIC BUILD POLICY · SYSTEM-INVENTORY COUNTS AND PUBLIC-FILING INDICES ONLY · NO CLIENT IDENTITY · NO PRIVATE OPERATING FIGURE · NO TRADING DATA · NO PROPRIETARY SOURCE CODE
Human × AI
“AI does the labour. A human asks the question, makes the call, and carries the responsibility.”
Most setups invert this—asking a model for judgement while a person does the grunt work. Run the right way round, the gain is not just speed: it is being able to test the boring, unglamorous questions that nobody had hours for before.
Contact
Start with one question.
Send the question you want settled. If it cannot be answered honestly with the data available, we will say so before any work starts.