A method that can be examined.

Start with the experimental record.

A recording is interpreted in relation to its animal, cohort, experimental phase and apparatus. Timestamp segments and spatial regions make that context explicit before measurement begins.

The reference three-chamber workflow distinguishes pretest, locomotion and stranger phases. Camera views, chamber masks and any image transformation belong to the recorded configuration; they are not universal settings for every study.

Locate anatomical landmarks.

Fylgja’s neural network pipeline obtains anatomical coordinates from video on hosted GPUs, without any labelling by your lab. For top-view mouse recordings it places 27 anatomical landmarks per animal. The exact network and configuration behind any result are read from its run record.

The pipeline runs pretrained and then adapts itself to each recording with no labels. Landmark placement is reviewed against the recording rather than read from the network’s confidence alone.

Methodological background: Ye et al. (2024), SuperAnimal pretrained pose estimation models for behavioral analysis, Nature Communications 15, 5165.

Define the behavioural measure.

Chamber occupancy

Assign tracked positions to apparatus regions, with a configured rule for transitions between chambers.

Movement

Calculate path length and velocity from a specified tracked position, accounting for sampling and coordinate scale.

Social proximity

Measure time near a specified cup or region. Proximity is an operational measure; its biological interpretation belongs to the assay.

Behavioural events

Apply the configured temporal and kinematic rules, then inspect the events against the recording.

Pixel distances, velocity thresholds and minimum event durations depend on camera resolution and analysis rate. A threshold has meaning only in conjunction with those settings.

Review, compare and retain.

Inspect difficult frames and trajectories, record any correction and compare the resulting run with the previous configuration. Quality checks, including agreement with another tracker, can identify discrepancies for review.

Export the data with the associated configuration, validation report and provenance. These records allow a reader to establish what ran, what changed and which source material produced a result.