The current consensus model for the circadian clock in mammals is based on a transcription-translation feedback loop. In this model, CRY and PER proteins repress their own transcription by suppressing the transactivator function of the CLOCK:BMAL1 heterodimer directly (physical model) and by facilitating post-translational modifications (chemical model). Most of the data for this model come from genetic and cell biological experiments. Here, we have purified all of the core clock proteins and performed in vitro and in vivo biochemical experiments to test the physical model. We find that CLOCK:BMAL1 binds to an E-box sequence in DNA and that CRY binds stably to the CLOCK:BMAL1:E-box ternary complex independently of PER. Both CRY and PER bind to CLOCK and BMAL1 off DNA but, in contrast to CRY, PER does not bind to the CLOCK:BMAL1:E-box complex. Unexpectedly, PER actually interferes with the binding of CRY to the CLOCK:BMAL1:E-box ternary complex. CRY likely destabilizes the CLOCK:BMAL1 heterodimer on DNA by a post-translational mechanism after binding to the complex. These findings support some aspects of the canonical model, but also suggest that some key features of the model need to be revised.