What this branch measured, and what it changed¶

Every number here was produced on real Claude Code traffic through a local context-guru proxy, counted with internal/tokens (o200k) and priced at the tier the request actually billed. The harness is bench/ab.sh over apply/sweep_capture_test.go; the corpus is four captures of real interactive sessions (short 5 requests, long 35, mixed 21, cold 9 with real 430 s idle gaps, 3 of them verified cache_read == 0).

The measurement rules that bind every claim below¶

1. Cold measurements need n>=3 and a range. Three identical runs of the same config on the cold arm gave 5.83% / 4.30% / 1.75% removed and net +$0.0097 / +$0.0026 / -$0.0086. The sign flips. A single cold number is noise.
2. The value tier is a 12.5x lever. The 14,407-request production snapshot is 90.54% cache_read (this branch's own four captures are 86.86%), so a removed token is worth 0.209x fresh. At the $2.00/MTok this gateway billed the bench corpus that is $0.412/MTok. Two measurement tests hardcoded $0.6265, which is the same multiplier against Anthropic's $3.00 list rate and therefore ~1.26x high for this corpus; corrected. Component ranking is unaffected, absolute dollars in those tests were.
3. removed > attempted is cache invalidation, not saving. Never counted.
4. The baseline is zero. The shipped default removes 0 tokens on short, long and mixed; only 7-27% of a request lies outside the cached prefix.
5. A harness bug invalidated our history. Coldness is computed only on the Tracker path, so CG_SWEEP_IDLE advanced the clock and changed nothing. Every earlier "cold-cache sweep" figure in this repo was measured warm. Fixed in d1dbd4e.

What improved, measured¶

The win is one fix, and it came from deleting our own work rather than adding any. A keep-list built from any 4+-letter word was harvesting English prose (against, including, large) into the KEEP set and then failing its own recall check on it: 13 of 40 calls refused a good compaction for "dropping" an English word. Reverting only that predicate returns the corpus to 17/40 with 14 acceptance-check rejections, so it is +9 of the +12 accepted.

An earlier draft of this document credited the cold-sweep turnaround to no longer forcing context: full. Independent re-measurement refutes that: on the fixed code full is also net-positive (+$0.0119, accepting 3/3), so the keep-list fix alone carries -$0.039 to about +$0.012 and dropping the force adds roughly +$0.005. The two changes were conflated.

New components, all shipped off and in no preset¶

What we proved is NOT there¶

The structural finding¶

Every large token class is unreachable by a proxy, and each for the same underlying reason.

• A Read goes stale only once a later Edit lands — and by that turn it is already inside the cached prefix. Re-anchoring costs $0.471 to buy $0.0027/turn: break-even 174 turns on a 122-turn session.
• Every Read in real traffic is the newest Read of its path, which is exactly the one the agent is working from. On long's largest turn the guard declines 9 of 9.
• Tool definitions are 19.8% of an interactive request and 59.2% on SWE-bench, and no component sees the tools array at all.

The moment information becomes safe to remove is the moment removing it costs a re-anchor.

This is also why rtk's largest lever does not transfer: it substitutes the command and shrinks output before it enters the transcript, so "is this the newest read" never arises. Same technique, opposite economics, purely because of where it sits.

And it is why toolfilter is the one item measuring in dollars: a session's first request is a cold start anyway, so filtering deterministically from turn 1 costs nothing.

Metric names that assert causes they do not measure¶

Four refuted premises above trace to this, so it is recorded as a rule:

A column computed downstream of an outcome cannot validate a predictor of that outcome, and a name containing a cause is a branch label, not a measured mechanism.

• below_trigger means "no component mutated anything"; Trigger is not consulted. Renamed.
• cache_miss_reason='ttl_expiry' is only reachable when cache_read == 0, so validating a cold predicate against it yields a guaranteed fake 0% false-positive rate. Measured against the real predicate instead: 4.38% false-cold at gap>=360 s, a floor not an estimate.
• cache_saved_usd is the provider's saving ($3,339.18) one column from ours ($7.07). A test now fails if any UI line renders it without naming the provider.
• frozen_tokens=0 coexists with a warm cache on 3,749 of 5,302 rows.
• saved_gross overstates saved_unique by up to 65.9x.
• uncompressed_reason='' means "we DID compact" — 64.9% of spend. An != '' filter inverts it.
• A missing component means not deployed, not "ran and did nothing".
• The FilterMiss ledger is not persisted at all, so it cannot prioritise anything.

Bugs found that matter more than the savings¶

• The derivation floor (>=90% of an extraction must match the input in order) costs 0 acceptances and 0.2 ms, and immediately caught a fixture accepting the Starlark program's own source as the "reduction" and splicing it into the transcript in place of the tool output.
• Both reformatters silently deleted everything after the first JSON document.
• collapse had no TailOnly, no freeze, no reapplyFrozen — it re-derived the whole transcript every turn. Cost so far: zero, because it has never been deployed.
• A one-token dedup change that looked safe was refused: ContentKey collapses every whitespace run, so a file read before and after a re-indent would be labelled identical to the model.
• The extraction contract banned constructs that work (%-formatting, dict.setdefault, sorted(key=), lambdas, dict comprehensions) and failed to ban the two that do not (type annotations, set comprehensions).

Honest position¶

Deployment net over the measured 3-day window was -$10.14: $7.07 gross saved against $17.22 of our own LLM spend. The extract fixes reduce the cost side and raise acceptance, but everything new in this branch ships off, so the deployed position does not move until something is enabled. The two candidates with a positive measured net are the cold sweep (+$0.020, sign-stable) and toolfilter ($0.1186 on one real session).

Production input is 90.54% cache_read and output is 8.086% of the bill, so addressable spend is 91.914% (this branch's four captures: 86.86% and 88.77%) — but a removed token is worth 0.209x fresh. Low-single-digit percentages of the addressable bill are what the transformation components can reach on this traffic. The order-of-magnitude item is removing declarations that are never used, because it is the only work that changes the prefix before it is ever cached.