Research Integrity
Where errors can arise, and how we address them.
Four places a systematic process goes wrong, each drawn as the sequence it follows. Every step names the error it invites and the control that exists today.
Data and timing
A rule may only use what was knowable on the day it decided, and it is graded under an explicit execution timing and fill assumption.
Select a step to see where errors can arise and how we address them.
-
Step 01 · Information available
Where errors can arise
Missing or revised data
A vendor can publish late, fill a gap later, or revise rows a decision already used.
How we address them
Revisions recorded before use
Each daily refresh is compared with the publication it replaces, and every revised row is recorded before the data is used.
Details
A recorded revision is reviewed, not acted on automatically, so a late fill and a changed row are both visible before they can influence a decision.
-
Step 02 · Decision cutoff
Where errors can arise
Information not yet knowable
A rule can use a price or a filing that was not yet available on the day it decided.
How we address them
Dated inputs
Prices are read from a single dated publication. For research using fundamental data, filings are kept as dated, write-once snapshots, which began in October 2026.
Details
Later research sees what was known at the time of each snapshot. Fundamental history before October 2026 carries a single as-of date, a stated limitation of any research that uses it. The published backtest uses ETF price data only, so the snapshot control does not bear on it.
Where errors can arise
Units and adjustments
Research prices are adjusted for distributions and splits, so a research position is not the share count a broker holds.
How we address them
One conversion, then whole shares
Research positions are converted into broker share units once, from the adjusted and unadjusted prices of the decision date, and rounded only when orders are generated.
Details
The conversion refuses to proceed if either price is missing or invalid. Treating adjusted and unadjusted units as the same would misstate the position, so the two are never mixed after that single step.
-
Step 03 · Next-session execution
Where errors can arise
Grading at an impossible price
Grading a decision at a price it could not have traded overstates what the rule could have earned.
How we address them
Next-session execution assumption
A decision built from data through a day's close is assumed to execute at the next session's open, and paper fills are checked against that assumption each session.
Details
The published backtest uses this assumption throughout. Paper trading follows the corresponding opening-order workflow; its observed simulated fills are compared with the assumed price rather than presumed to match it, and no identical fill price or outcome is implied.
Research bias
Criteria are set before an experiment runs, failures stay on record, and the published backtest is labelled in-sample because that is what it is.
Select a step to see where errors can arise and how we address them.
-
Step 01 · Predefined criteria
Where errors can arise
Repeated experimentation
Testing many variants on the same history and keeping the best one finds noise as readily as signal.
How we address them
Criteria set in advance
Experiments state their rule and pass criterion before they run; ideas that fail are closed, not re-tuned until they pass.
Details
Closed experiments include those that fell just short of their criterion. The record of what was tried, with its outcome, is kept alongside the results that survived.
-
Step 02 · Experiment
Where errors can arise
In-sample fitting
Rules and weights fitted on a history, then shown on that same history, describe the fit rather than forecast anything.
How we address them
Fitted on the history it is shown on, and said so
The published backtest's rules were researched and its risk shares fitted on data inside the same window. It is presented as a description of the method, not as validated out-of-sample evidence, because that validation has not been done.
Details
There is no untouched holdout, and the same history has been used for several portfolio studies. Paper trading, running since June 2026, is the ongoing evaluation of the rules; no paper performance figure is published yet.
-
Step 03 · Retain or close
Where errors can arise
Selection effects
Reporting only the surviving strategies, instruments or windows hides the failures that would qualify the result.
How we address them
Failures stay in the record
Retired strategies and closed experiments remain on record, and a published result names what it leaves out.
Details
The published backtest applies today's rules across the whole history, so its composition reflects which research survived; its methodology note says so.
-
Step 04 · Recorded outcome
Where errors can arise
Mixed or unlabelled results
A figure whose basis, backtest or paper, or whose source is not stated can be misread as evidence.
How we address them
One label per figure
Every figure is labelled BACKTEST or PAPER and carries its data cutoff; the two are never combined in one number.
Details
The homepage chart is a backtest and is labelled in-sample. No paper performance figure is published; when one is, it will be labelled PAPER and shown apart from any backtest.
Execution differences
What the rules wanted, what was sent and what was filled are three different records, kept apart and reconciled every session.
Select a step to see where errors can arise and how we address them.
-
Step 01 · Target
Where errors can arise
A book that records only one thing
What the rules wanted, what was sent and what was executed can each differ, and a book that records only one cannot tell which.
How we address them
Three records, kept apart
Every session keeps the target, the order and the fill as separate records, so each difference can be located rather than inferred.
Details
The target is the position the rules asked for, in broker share units after the single conversion described under Data and timing. It is stored before any order exists.
-
Step 02 · Order
Where errors can arise
Orders that differ from targets
Whole-share rounding, a failed pre-trade check or a cancelled order mean what was sent is not exactly what was wanted.
How we address them
Checked before the open, recorded as sent
Orders pass pre-trade checks in a dry run before the open; if a check fails, the session does not trade, and what was actually submitted is recorded as its own record.
Details
Pre-trade checks cover buying power, affordability and leverage. A session whose orders were cancelled or only partly filled is reconciled from its confirmed fills and marked as such, not assumed complete.
-
Step 03 · Simulated fill
Where errors can arise
Modelled fills and costs
A backtest assumes a fill at a modelled price and cost; an actual order can fill partly, late, or at a different price.
How we address them
Modelled costs, observed fills
Research charges a fixed half spread plus the broker's commission schedule; paper fills, fees and slippage against the modelled price are recorded each session.
Details
Market impact is not modelled. Paper fills are simulated on real market data with no capital at risk, so they do not show the impact real capital would have; the comparison between modelled and observed cost is a check on the model, not evidence about real trading.
-
Step 04 · Reconciliation
Where errors can arise
Differences nobody notices
A partial fill, a cancelled order or a fee that differs from the schedule can sit in the book unexamined.
How we address them
Reconciled every session
Target, order and fill are reconciled against the broker's own records the same day, and broker fees are reconciled per order against the amounts actually charged.
Details
Reconciliation records what happened; it never prepares or submits orders. A session that fails verification is recorded as an error, not as a result, and discrepancies are investigated before any report is produced.
Traceability
Trading runs only from versioned releases on dated inputs, and every published figure ties back to the record it came from.
Select a step to see where errors can arise and how we address them.
-
Step 01 · Versioned code
Where errors can arise
Unversioned change
Code that changes between a result and its reproduction makes the result impossible to trace.
How we address them
Immutable, versioned releases
Trading runs only from immutable, versioned releases of the code, each with a recorded rollback target.
Details
Identical inputs reproduce identical decisions. This is a property of the software, not a statement about future returns.
-
Step 02 · Configuration
Where errors can arise
Silent configuration drift
A sleeve, budget share or parameter changed without a record alters decisions nobody meant to touch.
How we address them
Reviewed, versioned, dated
Configuration is kept under version control with the code; allocation changes are proposed, confirmed with a stated effective date, and released together with it.
Details
Adding or retiring a sleeve is a reviewed change to the portfolio's configuration. A failed allocation review leaves the previous allocation in force.
-
Step 03 · Dated inputs
Where errors can arise
Inputs that cannot be reproduced
A decision made on data that has since been revised or overwritten cannot be rebuilt or checked.
How we address them
Dated publications, kept
Each decision reads one dated input publication, and revisions are recorded rather than written over, so the inputs of any past decision can be read again.
Details
How revisions and snapshots are handled is described under Data and timing.
-
Step 04 · Decision
Where errors can arise
Silent regressions
A change elsewhere in the system alters a decision nobody meant to touch.
How we address them
Tests before release, observation after
Every change is reviewed and tested before release, then observed through its first paper trading session and reconciled before it is treated as settled.
Details
Replay checks rebuild recent decisions from recorded inputs and compare them with what was actually decided. Pre-release review and testing are separate from post-deployment observation; a change is settled only after both.
-
Step 05 · Report
Where errors can arise
Reports that drift from records
A published figure that no longer ties to the record it came from, or a correction that overwrites what was first published.
How we address them
Figures tied to records
Daily and monthly reconciliation documents tie every position and figure to the broker's records and the data sources; corrections are made in place with a visible note.
Details
A correction keeps the original publication date and raises a visible revision note; nothing is replaced silently.
This page describes controls as they exist today. Scope and dates are stated where a control is recent or partial. Nothing here is investment advice; Quantcore Labs does not manage client assets.