Hypothesis

Accumulating channel statistics during first N prefill tokens enables CAT-style alignment correction without offline calibration.

Background

CAT diagonal alignment requires per-channel variance stats. Instead of offline calibration (needs separate pass), accumulate stats during first 64 prefill tokens and apply correction to remaining tokens. Plays well with graph-side WHT architecture since stats-gathering can be inserted as a graph node.

What to test

• Accumulate channel mean/variance during first 64 prefill tokens
• Apply CAT diagonal correction to tokens 65+
• Compare quality vs offline calibration vs no calibration
• Measure prefill overhead from stats accumulation
• Test sensitivity to calibration_tokens count (32, 64, 128)

Expected outcome

Similar quality to offline CAT calibration with zero extra passes. Small prefill overhead.

Priority

Medium — depends on CAT diagonal results.

Source

AutoRepl: TODO-012 (buun, fork_dc582a)