Codec direction (non-normative)

constrained recursive partitions rather than maximal combinatorial partition grammars;

explicit frame dependency graphs;

approximately 2 s encoder-default independent epochs, to be measured against 1/4/8 s;

conventional block motion plus an experimental coarse motion lattice with local corrections;

non-overlapping integer transforms first; lapped transforms are not the default because their neighbor/partition coupling has historically made intra coding and variable block structure harder;

scalar inverse quantization with most perceptual intelligence pushed into the encoder;

independently restartable, interleaved rANS entropy states with table selection/deltas rather than per-symbol probability mutation;

palette, transform-skip and intra block-copy as first-class screen-content candidates;

patch dictionary only if it materially wins over simpler screen tools;

exactly one normative reconstruction-filter stage if experiments show sufficient gain.

one common transform/energy/residual path for general audio;

transient-dependent time resolution;

per-band stereo coupling;

optional LPC/pitch whitening, retained only if speech gains do not compromise general audio simplicity/performance;

rANS-family entropy coding if the video experiments validate its decoder tradeoffs.