Electromechanical latching relay pull-in voltage_Technical issue

We are using MIL standard Latching Electro-mechanical relays having dual ON/OFF coils. During screening process the relays are being subjected to pull-in voltage

measurements for both ON & OFF coils in SRC (25 degree Celsius),COLD (8 degree Celsius) & HOT (80 degree Celsius) conditions. For pull-in measurement, the coil voltage

starts at 0V and gradually increasing to 28V in continuous manner. Pull-in will occur in 6-16V range. Status monitoring (Open/Close) is done at relay poles.

The anomaly is that the pull-in voltage measured at Hot is showing values lesser than Cold and Cold value is higher than SRC. This observation is against the normal

intuition of pull-in voltage HOT>SRC>COLD.
For example, The pull-in of RL-1 ON coil shows 11.496V (ISRC), 15.042V (Cold), 17.719V (Hot). That is SRC reading is 3.5V lesser than Cold.Also the difference voltage

(3.5V) shows inconsistency under repeated tests of the same part.The curreny limit of 30V source is set at maximum coil current of relay as per data sheet.
The relay specifications are 450 Ohm coil, 15A/28V DC contact, 30V DC coil/500mA.

The deviation is observed only for 0.2% of total relays screened per batch.All others are operating well.

I hope for feedback.

Copper has a negative temperature coefficient, so when the coils are hot their resistance is slightly lower, so for a given voltage they take more current, hence a stronger magnetic field, so they are more sensitive. I suspect that this is actually a second order effect, the main being the temperature coefficient of the mechanics, how the spring reacts and the magnets them selves, or the lubricant (or lack thereof). Take some extreme examples apart and swop their bits and pieces around to isolate the culprit.
Frank

Although there may be explanations for the apparently observed temperature characteristic, my first guess would be a systematical fault in test setup.

Do you consider that the relays have a certain amount of magnetical hysteresis, in other words you need to guarantee defined initial conditions, usually perform several ON/OFF cycles before starting the measurement.