Record AR-EXP-000 · Research Index
Every experiment,
including the ones that failed.
A research record is a question, a method, a result and a status. This index holds every study the Foundry has run, each one carrying the lifecycle state it has actually earned rather than the one that would read best. Failed records stay in the index permanently, because a negative result that is deleted has to be discovered again by someone else.
01Archive DefinitionAR-EXP-000·01
A result without a status
is an opinion with a chart.
Most trading research is published at exactly one moment: when it looks good. There is no record of the version that did not, no record of the assumptions that were changed to reach the published version, and no way to tell whether a conclusion has been stressed or merely stated.
The archive is structured to make that impossible here. Every study carries a lifecycle status drawn from a fixed eight-state vocabulary, and a status can only advance when a specific, published condition has been satisfied. A record whose conclusion is interesting but untested sits at Simulated indefinitely, next to records that reached Validated, and the difference is visible at a glance.
The other structural decision is retention. Failed is a terminal state, not a deletion. AR-EXP-097 attempted to classify market regimes from price alone, could not do so accurately at the transitions, and remains in the index with that finding written plainly. It is one of the more useful records here precisely because it closes a line of enquiry that looks promising from the outside.
What the archive is not: a performance report. No record contains a return, an account curve or a claim about profitability. Every study is a statement about the relationship between decisions and behaviour inside a documented model, and every one of them is simulated.
Anatomy of a research record
Every record in this archive is simulated
All fourteen studies were produced inside the Foundry's deterministic preview engine, which reads no market data, no broker history and no trade journal. They are original to this laboratory and they describe the behaviour of a model. Their value is in the direction of the relationships they report, not in the magnitude of any number.
02Record IndexAR-EXP-041…118
The index.
Filter by record type, lifecycle status or research field, search the full text of every abstract and finding, and sort by date, status or identifier. Two views read the same records: an editorial card view for reading, and a technical table view for scanning. Nothing is abbreviated in either.
Search abstracts and findings
Sort
Record type
Lifecycle status
Research field
Showing 14 of 14 records
No record matches
Fourteen records cannot populate every combination of eleven types, eight statuses and nine fields — most pairings are empty because the archive is small, not because the question was avoided. Clear a filter or widen the search term.
| Record | Type | Field | Status | Title, abstract and finding |
|---|
No record matches
The table is showing the same filtered selection as the card view, and that selection is currently empty. Clear a filter to widen it.
03Lifecycle VocabularyAR-VAL-020
Eight states,
and what each one costs to reach.
Status is the most misread field in any research index, because it looks like a quality score and is not one. It records what a study has survived. A Validated record with a modest finding has been through more than a Simulated record with a spectacular one, and the archive is built so that the difference is impossible to miss.
Failed records are retained deliberately
A failed study is a closed door with a reason attached, and that is worth more than an open question. Deleting it means the next researcher — including a future version of this one — pays the same cost to learn the same thing. Failure is also the only status that cannot be reached by inaction, which makes it evidence that the work was actually done.
Status never advances by decision
There is no editorial route from Simulated to Validated. The eight validation conditions below are the only mechanism, and a record that fails any one of them stops where it is. Two records in the index have been sitting at Under Review for months for exactly this reason.
Status can move backwards
If a later study contradicts an earlier one, the earlier record is demoted to Under Review or moved to Archived with a cross-reference — never quietly edited. The identifier stays, so any external reference to it keeps resolving to the same study.
04Archive CompositionAR-EXP-000·04
What the archive
is actually made of.
Two views of the same fourteen records. The first counts them by lifecycle state, including the two states currently holding nothing. The second shows how each study class is distributed across those states, which is where the archive's real imbalance becomes visible.
Counts across the eight-state lifecycle vocabulary.
Reading the census
Each bar is normalised to full width, so the bars show composition rather than count — the record count sits in the row label. Most types hold a single record and therefore appear as one solid block, which is an honest picture of an archive this size rather than a defect in the chart.
This census counts records, not evidence
Eight Validated records is not eight independent confirmations of anything. They share one engine, one set of archetype constants and one author, so their errors are correlated in ways no count can reveal. The census describes the shape of the archive; it says nothing about how much the archive knows.
05Advancement StandardAR-EXP-102
Eight conditions
between Simulated and Validated.
Specified in AR-EXP-102 and applied identically to every study class. A result that has not been through all eight is recorded as Simulated, never as Validated, however convincing it looks. Most retail conclusions fail the last condition alone.
| # | Condition | What is done | Why it is there |
|---|
How a status advances
A record enters at Proposed with an abstract and no result. It reaches In Calibration once the configuration under test is fixed and written down, which is the point after which parameters may no longer be adjusted in response to what the results look like.
Simulated means the study has produced output. This is where most work stops in the wider literature and where the archive considers it barely started. Under Review means the eight conditions are being applied and at least one has not yet been satisfied.
Conditional is the honest middle: the finding survives the conditions within a stated boundary and fails outside it. AR-EXP-071 sits here because its conclusion holds per regime and dissolves in aggregate — which is itself the finding.
Validated requires all eight conditions with no outstanding exceptions. Failed is recorded when a condition cannot be met and the study is closed. Archived is for records superseded by later work, retained with a cross-reference.
Sample adequacy, stated plainly
n* ≈ (z · σ_R / EV)²
The trade count required before an expectancy estimate has a confidence interval narrower than the estimate itself. σ_R is the dispersion of outcomes in R and EV is the mean. A configuration with 0.10 R expectancy and 1.4 R dispersion needs several hundred observations before its sign is reliable — which is why the archive states minimum counts in advance rather than reporting whatever the sample happened to be.
The same arithmetic explains why raising dispersion is expensive twice over: it widens the drawdown and it multiplies the evidence required to establish that anything was learned. Two records in this archive are held at Under Review by this condition and no other.
How many trades before you are allowed to believe it →06MethodologyAR-EXP-000·06
How a study is run
in a laboratory with no market data.
Every record in this archive answers a comparative question, never an absolute one. Not “what is the expectancy of this configuration” but “what happens to expectancy when this one thing moves and everything else is held”. The engine is deterministic, so the comparison is exact: the same inputs return the same outputs on every run, on every machine.
That constrains what can be claimed. A result of the form “partial monetisation above sixty percent reduces tail participation faster than it raises capture efficiency” is a statement about the model's structure, and it is defensible. A result of the form “sixty percent is the correct partial” would require market data and does not appear anywhere in the archive.
01 · The question is registered before the run
The abstract is written first and does not change afterwards. This is the cheapest available defence against the most common failure in retail research: running many variations and reporting the abstract that fits whichever one worked.
02 · One variable moves at a time
Sweeps hold every other control at the laboratory default. Where two variables genuinely interact — partial percentage and trail activation, concurrency and correlation — the record says so explicitly and reports the interaction as the finding rather than as a footnote.
03 · Sequence is generated from explicit seeds
No study uses an unseeded random source. Every ordering, every resampled path and every histogram in the archive can be regenerated exactly, which means a disputed result can be re-examined rather than merely disbelieved.
04 · The finding is written at the strength the evidence supports
Records that cannot support a conclusion say so in the finding field. AR-EXP-113 reports that early modelling looks encouraging and that the sample is not yet adequate for a stronger statement — which is the whole of what is known about it.
07LimitationsAR-EXP-000·07
What this archive
cannot support.
Five limits that apply to the whole index, not to individual records. Any of them is sufficient reason to treat a finding here as a hypothesis about markets rather than a fact about them.
Fourteen records is a very small archive
Whole subject areas have one record or none. Position sizing, correlation structure and portfolio construction are each represented by a single study, and one study is not a literature. Coverage will look uneven for a long time, and the index is built to make the gaps visible rather than to disguise them.
Everything is simulated, and by one engine
Every record was produced by the same deterministic model with the same archetype constants. If that model misrepresents a relationship, it misrepresents it identically across the entire archive, and no amount of internal cross-referencing would reveal it. External replication against market data does not exist in this phase.
All original, which means unreviewed
Nothing here has been through external peer review, and the validation protocol is applied by the same laboratory that runs the studies. That is a genuine weakness. The protocol is published in full so that a reader can apply it themselves, which is the best available substitute and not an equivalent one.
Validated does not mean profitable
A Validated record has survived eight conditions inside a model. It has not made money, it has not been traded, and it makes no prediction about any account. The status vocabulary describes evidential standing and nothing else — a point worth repeating because the word carries an implication the archive does not intend.
Findings do not compose
Two Validated records about different parameters cannot be added together to produce a configuration. Parameters interact, the archive has studied only a fraction of the interactions, and the practice of assembling a system from individually attractive findings is one of the more reliable ways to build something that fails in a manner no single study anticipated.
The records describe relationships.
The laboratory lets you move them.
Every finding in this index came from changing one thing and measuring what else moved. The same engine that produced them is on the parameter laboratory page with all 159 controls exposed, so a record can be re-run rather than taken on trust.
14 records · 8 lifecycle states · 1 failure retained on purpose