Cache reuse trend-gate alternatives — 2026-08-10

Verdict

INCOMPLETE. The local packet executes fak’s trailing-window cache-reuse regression fold and a raw-JSONL/no-gate baseline. Prometheus, OpenTelemetry, Prometheus rules, and Datadog retain zero measurements until real ingestion, query, storage, and alert paths process the common ledger; #6157 tracks those witnesses.

Same-workload contract

Every arm receives five chronological rows: two stable four-turn sessions at 0.80 reuse, one ignored single-turn cold row, and two recent four-turn sessions totaling 0.35 reuse. Both windows meet the eight-turn minimum; the -0.45 change exceeds the 0.05 regression tolerance. The oracle requires exactly one REGRESSED alert, with the single-turn row excluded from both windows. The rows represent 1,700 prompt tokens.

Complete monitoring runs report true/false/missed regressions, query and scrape-to-alert latency, rows/second, represented tokens, CPU/RSS/network/storage, and total cost.

Arm	Class	Local availability	Honest result
fak native trailing-window trend gate	native	yes	exact one-regression oracle
raw JSONL ledger without trend gate	tuned baseline	yes	preserves rows but emits no verdict
fak + Prometheus	first-class integration	no	real scrape/query/alert path required
fak + OpenTelemetry	first-class integration	no	real export/collector path required
Prometheus recording and alerting rules	external	no	zero measurements
Datadog change and anomaly monitor	external	no	zero measurements

Adapters and local query-shaped functions are not monitoring-system witnesses.

Local native witness

go test ./internal/cachevalueledger -bench BenchmarkFoldTrendGate -benchmem -run '^$' -count=5

Windows/amd64, AMD Ryzen 9 9950X: 723.3, 693.3, 877.4, 851.8, 909.8 ns/op. Median: 851.8 ns/five-row fold, 2,644 B/op, 8 allocs/op. This is in-process fold cost, not ledger-ingest or scrape-to-alert latency.

Reproduce

go test ./internal/cachevalueledger -run TestCompareLocalKeepsTrendAlternativesExplicit
go test ./internal/cachevalueledger -bench BenchmarkFoldTrendGate -benchmem -run '^$' -count=5
go test ./internal/nativebench