Storage cost and read cost differ
A library can hold one hundred thousand tokens and cost almost nothing per session when retrieved on demand. Two thousand tokens in core may be reread every time and compound into a larger bill. Disk size misses the recurring cost that matters.
Weight therefore separates total vault size from the boot-loaded subset. Per-file token estimates and load order reveal where repetition compounds.
Read the boot chain from code
If instructions load an identity document, a core glob, and selected indexes, that executable contract is the measurement source. A prose list or hard-coded UI array becomes false as soon as the loader changes.
Dynamic branches matter. Mark files as unconditional, conditional, or on-demand so a reader distinguishes recurring cost from context used only for a task.
A token estimate is a receipt, not a price tag
Tokenizers and models count somewhat differently. Displaying the number as an exact currency amount creates false precision. The purpose is relative weight and change detection, not reconstructing an invoice.
Record tokenizer and measurement time. Keep currency as a separate projection so a model change does not destroy historical token comparisons.
Delta is more actionable than total
A total of fourteen thousand is less useful in review than learning that one change added eighteen hundred boot tokens. Semantic commits can carry token delta so the cost of memory editing remains visible.
A decrease is not automatically good. Minus four thousand may be duplicate removal or sanctuary loss. The number calls for review; reason and diff complete judgment.