Finding: Glyph codec runs on project traces (Track A)

Code: glyph_codec.py

Example out: memory/glyph_runs/latest_mix/

What was measured

Encode packets.jsonl + hybrids into palette + frames (token proxy, not tiktoken).

Demo multi-turn prose: frames body much smaller than raw; repeated facts reuse codes.

Interpretation

• Valid Track A engineering demo.
• Not proof of non-text representation.
• Chat-resend savings look huge because baseline resends growing history (fair for agents, not for one-shot).

Failure / limits

• Empty-row bug fixed once (density).
• Proxy tokenizer ≠ production billing.
• Palette quality depends on shingle absorb threshold.

Next

• Plug real tokenizer.
• Run on actual chat exports from product use.