How we measure cognitive performance — and what we don't claim.
A short walk through the construct mapping, scoring, aggregation, and privacy posture behind the WelloWork platform. We publish methodology as we generate pilot data.
- 01Adaptive taskDifficulty calibrates per session
- 02Construct mappingOne primary cognitive construct
- 03Per-employee baseline normalisationCompared to your own trend
- 04Team-level aggregateThreshold-gated, weekly smoothed
What is the WelloWork methodology, in one paragraph? WelloWork measures cognitive performance via short adaptive tasks mapped to five validated constructs — working memory, processing speed, attention, problem solving, and cognitive flexibility. Per-session scores are normalised against a per-employee baseline and aggregated into team-level trends with minimum-team-size enforcement. We make no clinical or diagnostic claims about biomarker reports.
How are exercises mapped to constructs?
Each task on the platform maps to one primary construct and at most one secondary construct. The primary mapping drives scoring; the secondary mapping is captured for methodology audit but does not contribute to the headline metric. We use established paradigms — N-back, span tasks, symbol-substitution, Posner cueing, Raven-style reasoning, task-switching — adapted for short on-platform sessions.
- Sequence recallWorking Memory
- Symbol substitutionProcessing Speed
- Target detectionAttention
Each task maps to one primary construct. Secondary signals are captured but do not drive the score.
Why WelloWork scenarios — not cognitive games?
Research consistently shows people get better at cognitive games — not at the complex decisions their jobs actually require. WelloWork scenarios are grounded in behavioral field work.
Meta-analyses consistently show people improve at the game. Transfer to real workplace complexity: minimal.
Built from field observations of how people actually reason at work. Not a proxy. Not a game.
How are scores computed?
Per-session performance is normalised against the employee's own running baseline (z-scored within the last 90 days). This deliberately avoids comparing one employee against another at the individual level, since population-relative scoring is sensitive to noise that doesn't matter in a workplace context.
Comparing an employee against a peer distribution. Noisy at the individual level — and uneasy at work.
Each session is scored against the employee's own running baseline from the last 90 days.
We don't compare employees to each other. We track each person against their own established baseline.
How are aggregations done?
What do we deliberately not claim?
We do not claim transfer of cognitive training to specific business outcomes (revenue, productivity). We do not claim clinical or diagnostic value for biomarker reports. We do not claim individual employee ranking is reliable from short adaptive tasks — only that trends are. And we do not invent metrics: every claim ties back to a published construct or to a methodology note we will publish under /research.
- Cognitive trends over time
- Team-level patterns
- Construct-mapped behavioral signals
- Within-person variance
- Transfer to specific business outcomes
- Clinical or diagnostic value
- Reliable individual ranking from short tasks
- Self-invented metrics
How do we publish updates?
Methodology notes will be posted under /research/science-insight as pilot cohorts produce enough data to write something defensible. We will not publish individual customer data, and we will not publish aggregates that don't meet our minimum-team threshold.
- M1Pilot cohort data collected
- M2Methodology note drafted + reviewed
- M3Published under /research/science-insight
We publish when the data is defensible. Not before.
Privacy in the methodology
Methodology and privacy are linked. The platform's choice to normalise within an employee, not against a population, is also what makes it harder to "de-anonymise" an aggregate.
An individual baseline carries no information about peers — there is no peer distribution to back-solve against.
Small-cohort aggregates can be reverse-engineered to a single employee. We suppress them at the source.
The same methodological choices that make our measurement accurate also make de-anonymisation structurally harder.