Validation
How close are the app's numbers to the truth? To answer that, we need videos where the truth is known exactly, frame by frame. This page reports such a test for every video assay, and is explicit about what it does and does not show.
The test videos
Ten videos were generated by a program. Each shows a top-down mouse (body, head and a thin tail) moving through an arena according to a script of walking, exploring, resting, freezing and hiding. Because every frame was drawn by the script, the mouse's true position and state are known for every frame.
| Video | Assay | Length | Arena |
|---|---|---|---|
| 01 | Open field | 180 s | 40 × 40 cm square |
| 02 | Elevated plus maze | 160 s | Arms 30 × 5 cm |
| 03 | Light–dark box | 150 s | 60 × 30 cm, dark third covered |
| 04 | Place preference, 3 chambers | 150 s | 90 × 30 cm |
| 05 | Novel object recognition (test) | 180 s | 40 × 40 cm |
| 06 | Water maze, acquisition trial | 34 s | 100 cm pool |
| 07 | Water maze, probe trial | 60 s | 100 cm pool |
| 08 | Barnes maze | 30 s | 92 cm, 20 holes |
| 09 | Fear conditioning, cue test (CS at 120 and 210 s) | 270 s | 30 × 25 cm |
| 10 | Y-maze | 83 s | Arms 35 × 10 cm |
All videos are 1280 × 720 at 30 fps, H.264, the format an iPhone records.
- Ground truth. The true position is the center of the drawn mouse pixels in centimeters. The true metrics were computed by a separate Python program from these positions and the scripted states, using the definitions in this manual and the same 10 Hz time points.
- App values. The videos were processed by the app's own Swift tracking and metric code (version 1.3), with the arena calibrated exactly. Decoding mimics the iPhone: 10 Hz, frames at 192 px width, background from 21 frames.
Position accuracy
| Video | Median position error | |
|---|---|---|
| Open field | 0.17 cm | |
| Plus maze | 0.48 cm | Background repair active (5.2 % of the arena) |
| Light–dark box | 0.15 cm | Visible samples only |
| Place preference | 0.22 cm | |
| Object recognition | 0.19 cm | |
| Water maze, acquisition | 0.47 cm | |
| Water maze, probe | 0.34 cm | |
| Barnes maze | 0.47 cm | 95th percentile 2.0 cm |
| Fear conditioning | 0.26 cm | |
| Y-maze | 0.32 cm |
A mouse body is about 8–10 cm long, so the error is a few percent of a body length.
Metrics: truth vs app
Open field
| Measure | Truth | App |
|---|---|---|
| Total distance (m) | 13.62 | 13.89 |
| Center time (s) | 80.5 | 80.3 |
| Center time (%) | 44.7 | 44.6 |
| Center entries | 20 | 20 |
| Latency to center (s) | 19.3 | 19.3 |
| Immobility (s) | 20.0 | 20.0 |
| Mobile time (%) | 88.0 | 88.2 |
Elevated plus maze
| Measure | Truth | App |
|---|---|---|
| Open-arm time (s) | 27.8 | 27.8 |
| Closed-arm time (s) | 127.6 | 127.6 |
| Center time (s) | 4.6 | 4.6 |
| Open-arm time (%) | 17.9 | 17.9 |
| Open-arm entries | 2 | 2 |
| Closed-arm entries | 3 | 3 |
| Open-arm entries (%) | 40.0 | 40.0 |
| Latency to open arm (s) | 2.3 | 2.3 |
| Total distance (m) | 3.59 | 4.85 |
| Arm sequence | ON CE CW OS CE | ON CE CW OS CE |
Light–dark box
| Measure | Truth | App |
|---|---|---|
| Light-box time (s) | 85.0 | 85.0 |
| Dark-box time (s) | 65.0 | 65.0 |
| Light-box time (%) | 56.7 | 56.7 |
| Transitions | 4 | 4 |
| Latency to dark (s) | 29.4 | 29.4 |
| Visible samples (%) | 56.7 | 56.7 |
Place preference (3 chambers)
| Measure | Truth | App |
|---|---|---|
| Paired chamber time (s) | 88.4 | 88.4 |
| Unpaired chamber time (s) | 31.8 | 31.9 |
| Neutral chamber time (s) | 29.8 | 29.7 |
| Preference score (s) | 56.6 | 56.5 |
| Paired-side preference (%) | 73.5 | 73.5 |
| Paired chamber entries | 2 | 2 |
Novel object recognition
| Measure | Truth | App |
|---|---|---|
| Novel object exploration (s) | 27.8 | 27.4 |
| Familiar object exploration (s) | 11.5 | 11.3 |
| Total exploration (s) | 39.3 | 38.7 |
| Discrimination index | 0.415 | 0.416 |
| Recognition index (%) | 70.7 | 70.8 |
| Novel object bouts | 3 | 3 |
| Familiar object bouts | 2 | 2 |
Morris water maze, acquisition trial
| Measure | Truth | App |
|---|---|---|
| Path length (m) | 2.40 | 2.43 |
| Target quadrant time (%) | 25.4 | 24.9 |
| Opposite quadrant time (%) | 26.8 | 26.8 |
| Wall-zone time (%) | 42.9 | 42.9 |
| Mean distance to platform (cm) | 44.4 | 44.4 |
| Escape latency (s) | 20.5 | 20.5 |
| Swim speed (cm/s) | 11.7 | 11.9 |
| Platform crossings | 1 | 1 |
| Latency to platform zone (s) | 20.5 | 20.5 |
Morris water maze, probe trial
| Measure | Truth | App |
|---|---|---|
| Path length (m) | 10.42 | 10.60 |
| Target quadrant time (%) | 30.7 | 30.7 |
| Opposite quadrant time (%) | 21.5 | 21.5 |
| Wall-zone time (%) | 0.7 | 0.7 |
| Mean distance to platform (cm) | 33.4 | 33.5 |
| Platform crossings | 8 | 8 |
| Latency to platform zone (s) | 3.4 | 3.4 |
Barnes maze
| Measure | Truth | App |
|---|---|---|
| Primary latency (s) | 18.1 | 18.0 |
| Primary errors | 3 | 3 |
| Total errors | 3 | 3 |
| Escape latency (s) | 22.1 | 22.1 |
| Search strategy | Serial | Serial |
| Hole visits | -3 -2 -1 T | -3 -2 -1 T |
Fear conditioning, cue test
| Measure | Truth | App |
|---|---|---|
| Freezing (session) (%) | 35.2 | 35.0 |
| Baseline freezing (%) | 10.0 | 9.9 |
| CS freezing (%) | 83.3 | 83.0 |
| Post-CS / ITI freezing (%) | 36.7 | 36.3 |
| Freezing bouts | 6 | 6 |
| Mean bout duration (s) | 15.8 | 15.7 |
Y-maze
| Measure | Truth | App |
|---|---|---|
| Arm entries | 11 | 11 |
| Alternations | 6 | 6 |
| Spontaneous alternation (%) | 66.7 | 66.7 |
| Arm sequence | ABCACBABCBA | ABCACBABCBA |
| Total distance (m) | 5.81 | 5.91 |
What the numbers say
- Counts and sequences match exactly in every video. This covers center, arm and chamber entries, transitions, platform crossings, Barnes errors, hole-visit sequence and search strategy, the Y-maze arm sequence, and the number of freezing bouts.
- Times and latencies are within 0.6 s, most within 0.2 s. Percentages are within 0.5 percentage points. The discrimination index is 0.416 against 0.415.
- Freezing: CS freezing is 83.0 % against 83.3 %, and the mean bout length is within 0.1 s.
- Distance is slightly overestimated (+1 % to +2 %) in arenas that fill the frame, because residual tracking jitter adds to the true path.
- Plus maze distance is the exception: +35 %. The maze covers only about 55 % of the frame width, i.e. about 1.5 pixels per cm at analysis resolution. Small changes of the body outline while the mouse sniffs in place shift the centroid by a pixel or two, and over 160 s this adds up. Times and entries on the same video are exact. Use plus-maze distance only to compare groups filmed the same way, or frame the maze tightly.
What this test does not show
Synthetic videos are cleaner than real ones: even light, no reflections, no bedding, no shadows, a uniform floor. They verify that the tracking pipeline, the calibration geometry and every metric definition work as specified. They do not replace a check on your own setup. Before relying on a new setup, especially for freezing thresholds and the object and hole radii, score 3–5 of your own videos by hand and compare them with the app.
A second limit is in the definitions themselves: the app scores the body center. Hand scoring of object exploration or hole pokes uses the nose. The default radii (8 cm) are chosen to bridge that difference for a mouse, but the two kinds of scoring measure slightly different things.