How we know Litmus is right

The station is checked against independent evidence: a light sensor shining through the tube, manual review of the video, confounds we planted on purpose, and pairs of runs that should look identical.

7 of 7 runs in this validation are synthetic test data. Numbers from real station runs replace them as they are recorded.

7/ 7

detected color flips within 1 s of synthetic ground truth. Mean error 0.36 s, worst 0.65 s.

7/ 7

runs where the light-sensor colorimeter confirmed the camera’s flip within 1 s. Mean gap 0.29 s.

1/ 1

planted confounds surfaced as the top difference by the diff engine.

2/ 110comparisons

variables flagged as different between reference runs that followed the same protocol (10 pairs). These are false alarms.

100%

of 59 AI-written claims and report sentences cite evidence in the record.

probabilities shown anywhere. Claims use four labels: observed, possible explanation, contradicted, missing evidence.

Flip timing, run by run

Run	Detected	Truth	Error
Reference 1	110.5 s	109.94 s	+0.56 s
Reference 2	123.5 s	123.38 s	+0.12 s
Reference 3	109 s	108.72 s	+0.28 s
Reference 4	101.5 s	100.85 s	+0.65 s
Reference 5	110.5 s	110.22 s	+0.28 s
Method A	112 s	111.67 s	+0.33 s
Method B	76 s	75.71 s	+0.29 s

Reference pairs compared

Pair	Flagged as notable
Reference 1 vs Reference 2	nothing
Reference 1 vs Reference 3	nothing
Reference 1 vs Reference 4	nothing
Reference 1 vs Reference 5	nothing
Reference 2 vs Reference 3	nothing
Reference 2 vs Reference 4	temp_c
Reference 2 vs Reference 5	nothing
Reference 3 vs Reference 4	air_quality
Reference 3 vs Reference 5	nothing
Reference 4 vs Reference 5	nothing