Cache frontier weekly review - 2026-06-29

This is the first dated review entry for the cache frontier operating plan. It records what the current tree proves today, without promoting thin evidence into a bigger claim. The matching machine row is in review-ledger.jsonl under schema fak-cache-frontier-review/1.

Evidence run

Commands run from the repo root on 2026-06-29:

go run ./cmd/fak nightrun score --json
go run ./cmd/fak cachevalue report --since 2026-06-22 --json
go run ./cmd/fak cachevalue review --since 2026-06-22 --date 2026-06-29 --source-markdown reviews/2026-06-29.md --json

The cache-savings Track 2 ledger was absent, so the provider-dollar track is empty.

1. What did we use ourselves this week?

Track 1 has real WITNESSED kernel reuse, but the corpus is still thin and not yet the full dev-loop dogfood target.

Field	Current value
Total cache-value sessions	5
Multi-turn sessions	1
Single-turn sessions	4
Total turns	9
Multi-turn turns	5
Prompt tokens	457
Reused tokens	260
Gate prompt tokens	413
Gate reused tokens	260
Realized reuse ratio	0.6295399515738499
Session type mix	run: 5

Read this narrowly: fak has a measured reuse signal from real fak run sessions, not yet sustained guard / serve dogfood from the coding loop. The current weekly bucket is marked thin: true; the reuse number is visible, but it is not yet trend-strong.

2. What can a new person demo this week?

The cache-frontier demo path is now one walkthrough:

go run ./cmd/ctxdemo -bars
go run ./cmd/turntaxdemo -print
go run ./cmd/ctxplandemo -selfcheck
go run ./cmd/memqdemo
go run ./cmd/unseedemo -print
fak nightrun score --json

The walkthrough is documented in run-the-demos.md. Its job is to make the multi-agent reuse win, O(1) planned context, queryable memory algebra, and addressable KV removal visible as one product story.

3. What is the next missing witness or product surface?

Gap	Evidence	Next action
Thin Track 1 corpus	multi_turn_sessions=1, multi_turn_turns=5, weekly bucket thin=true	Accumulate more multi-turn guard / serve sessions, not only run rows.
No dev-loop dogfood rows	session_type_mix={"run":5} and no guard / serve rows	Capture multi-turn guard / serve cache-value rows from the dev loop.
No Track 2 provider-dollar row	track2_observed_usd=[], no docs/nightrun/cache-savings.jsonl	Wire the OBSERVED provider-dollar append so fak cachevalue report can show real net dollars side by side with Track 1.
Multi-agent win still not dogfooded as a recurring geometry	Current cache-value rows are run, not a recorded 50-turn x 5-agent dogfood shape	Add a recurring multi-agent dogfood run or review entry that states the geometry and why it matches or differs from the benchmark.
O(1) query not yet our own agent memory workflow	Demo path exists, but the review did not query a real fak session image	Pick the first operator-facing query path and use it on a real repo session.

Verdict

Cache-frontier visibility improved: there is now a named operating plan, a local demo walkthrough, and a first dated review entry with real cache ledger evidence.

The product goal is not done. The next review must move from thin run evidence to dev-loop dogfood: guard / serve rows, Track 2 OBSERVED-dollar rows, and an O(1) query over a real fak session.