Cache frontier weekly review - 2026-07-04
This is the second dated review entry for the
cache frontier operating plan, following
2026-06-29. The matching machine row is appended to
review-ledger.jsonl under schema
fak-cache-frontier-review/1.
It also records the reconciliation done under #2184: the DEFAULT-ENABLEMENT-NEXT-50 backlog rows 22 / 26 / 27 now carry landed-increment cross-refs, and the cache-value roll-up Track 2 “Shipped” claim is date-stamped and scoped so no cache-frontier doc claims dollar witnessing the ledger does not show.
Evidence run
Commands run from the repo root on 2026-07-04:
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 2026-07-04 \
--source-markdown reviews/2026-07-04.md \
--append-ledger docs/cache-frontier/review-ledger.jsonl --json
1. What did we use ourselves this week?
Track 1 (WITNESSED kernel reuse) grew since 2026-06-29 and now includes guard sessions,
not only run rows — so the no_guard_serve_dogfood_rows gap has closed. The corpus is
still marked thin: true, so the reuse number is visible but not yet trend-strong.
| Field | Current value |
|---|---|
| Total cache-value sessions | 11 |
| Multi-turn sessions | 6 |
| Single-turn sessions | 5 |
| Total turns | 71 |
| Multi-turn turns | 66 |
| Prompt tokens | 6051 |
| Reused tokens | 4056 |
| Gate prompt tokens | 5863 |
| Gate reused tokens | 4056 |
| Realized reuse ratio | 0.6965137614678899 |
| Session type mix | guard: 2, run: 9 |
The publishable value family is unchanged: marginal-over-tuned-warm-KV (~1.0x
single-session; the vs-naive 1/(1-reuse) re-prefill multiple is excluded per #1066).
2. Track 2 exists now — but read the dollar provenance narrowly
The 2026-06-29 review named “no Track 2 provider-dollar row” as the top missing witness,
because docs/nightrun/cache-savings.jsonl was absent. That gap is now partly closed:
the ledger exists and the review fold reports track2_observed_usd.present = true. But the
dollar columns are not what “OBSERVED net-dollar savings” is meant to certify, so record the
state precisely:
Slice of cache-savings.jsonl |
Rows | Dollar state |
|---|---|---|
Total rows (2026-07-01 → 2026-07-04) |
240 | mixed, see below |
Anthropic + fak-mechanism rows, pricing_source=default:anthropic/claude-opus-4-8 |
226 | dollar columns present, but computed at DEFAULT list prices — a list-price equivalent, not a metered provider invoice |
openai-responses (codex) rows, pricing_source=none |
7 | dollar_status=dollar_blind; every USD field 0 |
2026-07-01 rows (no pricing_source) |
7 | token-equivalent only; every USD field 0 — the original state #2184 described |
So the honest headline: the read/write token counts are OBSERVED (from the provider
response), and there are now 2 fak-authored compaction_shed rows in the ledger, but the
dollars are DEFAULT-priced projections, and non-Anthropic traffic is still
dollar_blind. A metered, net-OBSERVED-economics dollar witness is still open under
#1544 (NEXT-50 item 26). The
#2179 1h-write pricing-tier fix has
landed (CLOSED), so the projected dollars now attribute upgraded writes to the correct tier.
3. What did epic #1844 land that the backlog had not credited?
The DEFAULT-ENABLEMENT-NEXT-50 rows now carry the landed-increment cross-refs, keeping the “planning artifact” header honest:
| Item | Landed increment | Witness |
|---|---|---|
22 — preserve cache_control / stable prefix through guard/serve |
#1850 (C6, 1h TTL upgrade) + #1851 (C7, uncached-remainder trim), both CLOSED under #1844 | commits dbf5aaa4b, e717e4403 (fak gateway) |
26 — join Track 2 provider-dollar rows into fak cachevalue report |
join mechanism live; item still OPEN (#1544) because the dollars are DEFAULT-priced projections | cache-savings.jsonl + two-track fold; this review’s present:true |
| 27 — provider-cache card fields | write-premium / read-rebate / spend / net _usd fields + per-owner/per-mechanism attribution card (#1491 CLOSED); false-warm + agentic-activation fields still planned |
cache-savings.jsonl fields; #2179 pricing-tier fix |
4. What is the next missing witness or product surface?
| Gap | Evidence | Next action |
|---|---|---|
| Thin Track 1 corpus | multi_turn_sessions=6, weekly bucket thin=true |
Accumulate more multi-turn guard / serve cache-value rows. |
| Metered net-OBSERVED dollars still absent | all 226 priced rows use pricing_source=default:…; 14 rows dollar_blind / token-only |
Wire a metered provider-price source so Track 2 dollars become net OBSERVED, not projections (#1544). |
| Multi-agent win not a recurring dogfood geometry | cache-value rows are guard / run, not a recorded 50-turn × 5-agent shape |
Add a recurring multi-agent dogfood run or review entry stating the geometry. |
| O(1) query not yet our own agent-memory workflow | demo path exists; this review did not query a real fak session image | Query a real fak session through the chosen O(1) memory path. |
Verdict
VISIBLE_BUT_NOT_YET_PRODUCT_DOGFOOD. Track 2 rows now exist and the docs no longer
disagree about whether the provider-dollar ledger is present — but the dollars are
DEFAULT-priced projections, so the next review must move Track 2 from projected to
metered net-OBSERVED economics (#1544), and Track 1 from thin guard / run evidence to a
sustained dev-loop dogfood.