Years ago I performed silicon evaluation of a 22 bit ADC. I expected to learn, to be surprised, to be puzzled. I was.
1) your hand or face or body emits heat, and silicon junctions CLOSER to the heat source will be warmer; two nearby diodes would drift apart by 500 microvolts, and you'll experience about 60 seconds of settling time to the new offset voltage; given 0.1 meter of copper has 114 seconds of thermal time constant, we can expect heat flows to be a constant problem; I'd designed those 2 diodes onto the Eval PCB, to examine the heating by my face; one diode partially shaded the other diode, to ensure a heat flux difference.
Why are heat flows a problem? The movement of 1 watt thru a square of copper foil, from edge to edge, will produce 70 degree Centigrade temperature gradient. Yet the joining of dis-similar metals produces 5 to 40 microVolts per degree Centigrade, and PCBs have lots of such metallic transitions. The thermal mismatch of differential paths (Vin+, Vin-) becomes your challenge.
2) dielectric absorption of capacitors showed up; input filtering using RC lowpass, to explore the ADC's noise floor, showed 2 or 3 minutes of settling; when shorted briefly then opened up, nearly a millivolt of stored charge would slowly appear
3) the resistance of 1 ounce/foot^2 copper foil is 0.000500 ohms per square, for any size square; 1milliAmp thru a square will generate 500 NanoVolts of error; plan on using Finite_element modeling to design your PCBS at the 32 bit level. [edit the NanoVolts was firstly microVolts]
4) 1 amp of 60Hz pure sinusoid (no spikes) at 1 meter from 10cm by 1cm loop, will induce this voltage onto your PCB
Vinduce = 2e-7 * Area/Distance * dI/dT
Vinduce = 2e-7 *10cm*1cm/1meter * 377
Vinduce = 2e-7 * 1e-3 * 377
Vinduce = 1e-10 * 754 = 75 nanoVolts
Why? because thin copper foil will not shield against 60Hertz magnetic fields. At 60,000 Hertz, just barely. At 60,000,000 Hertz, quite well. But not at 60Hz.
5) those "quiet" digital interface pins, with either a 1 or a 0 level, are still buzzing with 200 or 500 milliVoltsPP of MCU rail noise; how close can you let a digital interface trace get to the 32-bit signals, given the MCU trash has pseudo-random (program dependent) patterns, and cannot be trusted to "average out" ?
6) some useful values for switched-cap noise
10picoFarad ................ 20 microVolts RMS
1000 picoFarad ............ 2 microVolts RMS
100,000 picoFarad ........ 200 nanoVolts RMS
10,000,000 picoFarad ..... 20 nanoVolts RMS
1Billion picoFarad ............. 2 nanoVolts RMS
using the formula: VnoiseRMS = sqrt( K*T/C)
What is use of this table? to achieve 2 nanoVolt noise levels, the equivalent energy of charging 1Billion picoFarad (0.001 Farad) must be provided from the signal source or from buffers or from amplifiers.