Two measurements, not a before and after

One figure counts the instruction files you imported. The other counts the generated material a host loads on every request. They are drawn at the same scale so their sizes are comparable, and they are never subtracted — they measure different populations of text, and both are ratios applied to characters rather than counts returned by a model. Each one prints what it assumed and what it left out.

Read these two figures separately

These two figures measure different things and are not a before-and-after of the same quantity. The first is what your pasted sources cost as they stand. The second is what the always-active generated material costs; on-demand packs are excluded and listed. Both are ratios applied to text, not counts returned by a model, and neither is a statement about spend.

Imported sources, as they stand

The instruction files you pasted, counted after redaction.

0estimated tokens

0 characters · claude-sonnet-5 · no text counted, so no ratio was applied

What this number assumes

Priced with the claude-sonnet-5 tokenizer profile at ~0 characters per token.
Ratio source: derived.
A ratio applied to the text, not a count returned by the model.
Counted on the redacted copy of your text — after import that is the only copy that exists.
No source has been imported, so this figure is zero rather than small.

Left out of this figure

Nothing. Every file named below the count is in it.

Counted

No file was counted, so this figure is zero rather than small.

Generated, always-active

The generated files a host loads on every request.

2,317estimated tokens

5,148 characters · claude-sonnet-5 · derived ratio at ~2.22 chars/token

What this number assumes

Priced with the claude-sonnet-5 tokenizer profile at ~2.2218385843763486 characters per token.
Ratio source: derived.
A ratio applied to the text, not a count returned by the model.
Excludes 10 on-demand item(s), listed separately.
Counts the 1 artifact(s) a host loads on every request, not the whole bundle.
Includes each file's provenance header and review notice — a host reads those too.
This rider is not approved, so the artifacts were rendered only to be measured, stamped with a fixed placeholder date. An approved export differs by the length of that timestamp and nothing else.

Left out of this figure

ai/AI-RIDER.md — the canonical document, for people; no host loads it on its own
ai/tech-policy.yaml — structured record, read on demand rather than loaded every request
ai/brand-policy.yaml — structured record, read on demand rather than loaded every request
ai/AGENTS.md — the same policy for a different host — a machine loads one of the four
ai/CURSOR-RULES.md — the same policy for a different host
ai/COPILOT-INSTRUCTIONS.md — the same policy for a different host
ai/api-context.md — a prompt fragment, used when a task touches one of the tools it describes
ai/context-manifest.yaml — describes the other files; nothing loads it as instructions
ai/golden-tasks.yaml — expectations on record, not instructions
ai/DECISIONS.md — a ledger for people; read when a decision is questioned

Counted

ai/CLAUDE.md

Both figures, on one axis

Two separate measurements drawn at the same scale. A shared axis makes their sizes comparable; it does not make one the result of the other.