Self-Tax Trend — fak’s own overhead and net effect, tracked over time

What this is. The living trend companion to the self-tax row in BENCHMARK-AUTHORITY.md. It tracks fak’s own mediation overhead — and the cases where that mediation makes work faster — as a dated series, net-true-labeled, every point tracing to a committed artifact and a reproduce command like every other authority row.

Epic: L5/T12 of #1147 (#1169). Design note (survey, maturity ladder, full DoD, per-ticket specs): docs/notes/self-tax-performance-assurance-tracking-1147.md. Anchor standard: net-true-value Question #2 (net-of-introduced-cost).

Snapshot date: 2026-06-28 Status: Living document — append a row when a new self-tax point lands.

Why a self-tax trend exists

fak inserts itself into the hot path of every tool call, result, turn, and commit. Each insertion has a cost — latency, tokens, wall-clock — and sometimes it removes cost (reuse it would otherwise redo). The security floor already proves a bad call can’t get through. This trend is the dual read-out: it proves a good call doesn’t get slower than its budget — and names it when fak made the work faster.

The honest shape is two-directional. fak’s tax is workload-dependent: a measurable cost on a tool-light task (nothing to reuse, only mediation to pay for), a measurable saving on a reuse-heavy task (prefill not redone, net of the tokens mediation adds). A single average would hide both; this trend reports the signed points and lets the workload speak.

A budget is an envelope with a stated scope, not a promise of zero cost — a gate that costs 8% and saves 40% is a net win, and this plane says so rather than redding on the 8% alone (epic non-goal (c)).

The trend (signed: + = cost added, − = cost removed)

Every row’s number is already a committed entry in the Authority; this doc is the fold over them that makes the self-tax direction legible as one series. Provenance is labeled per the standard: WITNESSED (fak measured it itself), OBSERVED (provider-relayed), MODELED (computed from a stated model).

First committed	Surface (lifecycle stage)	Self-tax point	Sign	Provenance	Artifact	Reproduce
2026-06-26	adjudication read-path floor (every tool call)	~0.55 ns/op · 0 allocs · FLAT N=1→1000 drivers	+ (floor)	WITNESSED	internal/abi/registry_scaling_test.go	go test ./internal/abi -bench BenchmarkRegistryReadScaling -benchmem
2026-06-20	pure-kernel decide — canonical allow (Submit)	362 ns/op allow (560–605 ns w/ ArgPredicates)	+	WITNESSED	experiments/mac-m3pro-kernel-20260620/kernel-latency-mac-m3pro-20260620.json	go test -bench per the M3 Pro Authority row
epic #607 (Regime A)	vDSO self-ablation, tau2-airline smoke (frozen trace)	tokens 937→417 = −520 tok (engine_calls 12→5, vdso_hits 0→7)	− (net saving)	WITNESSED	experiments/ablate/tau2-smoke-vdso-ablation.json (doc: docs/benchmarks/ABLATE-RESULTS.md)	go run ./cmd/fak ablate --trace testdata/tau2/tau2-smoke.json --sweep vdso
epic #607 (Regime B, #623)	cross-agent bare claude vs fak guard -- claude (pong, tool-light)	total-ingested 1.56× = +28,986 tok (5/5 ALLOW, 0 deny)	+ (net cost)	WITNESSED (kernel counters)	experiments/ablate/cross-agent-pong-opus.json (doc: docs/benchmarks/ABLATE-RESULTS.md)	python tools/cross_agent_ablate.py report --reps <reps.json>
2026-06-28	net-true improvement detector (decode reuse)	net_tokens = local-reuse ⊕ provider-reuse − mediation-tax → positive on a reuse-favorable trace	− (net saving)	WITNESSED ⊕ OBSERVED − MODELED	design note §9 (…self-tax-performance-assurance-tracking-1147.md)	curl -s localhost:PORT/metrics \| grep -E 'kv_prefix_reused_tokens_total\|inference_cached_prompt_tokens_total' (§9.4)

How to read a signed point

• + (cost): mediation fak adds. The two latency floors (0.55 ns read, 362 ns decide) are the per-call drop-in cost; the cross-agent +28,986 tok is what guarding a tool-light task ingests on top of the bare agent. These are the numbers a budget breach is defined against (the per-rung envelope is T2 of the epic).
• − (saving): cost fak removes. The vDSO ablation skips 520 tokens of redundant engine work on a frozen trace; the improvement detector folds realized KV-prefix reuse (WITNESSED, in-kernel) ⊕ provider cache_read (OBSERVED, disjoint plane — never double-counted) minus the tokens mediation re-emits (MODELED), and nets positive whenever reuse exceeds the tax.

The series is the witness for the epic’s net-true claim: fak is not silently degrading performance (the + floors are small and flat), and it detects when it improves performance (the − rows), each net of its own cost.

Honesty fences

• This is fak’s mediation overhead, not raw-inference parity. fak-vs-llama.cpp forward-pass parity is the separate axis tracked by track-b-performance-parity #306 — cross-linked, never blended into this number.
• Curated fold today, auto-emitted series tomorrow. Each row here is an independently-committed point measurement. The always-on version — a frozen-workload fak-on-vs-fak-off CI gate with change-point + persistence + SPRT (T6/T7/T8) and a single fak perf read-out + net-true /metrics family (T11) — is the named follow-on in the epic, not built here. Until it lands, this trend is a hand-folded read of committed artifacts, not a single auto-regenerated JSON. Stated plainly rather than implied.
• Single-box wall-clocks stay single-box. The latency floors are machine-stamped (Ryzen 9 9950X for the read floor; M3 Pro for decide); the deterministic counter rows (vDSO ablation token deltas, the reuse split) reproduce byte-identically and are hardware-independent. The cross-agent row is one tool-light task on one host — an honest worst case for tax (nothing to reuse), not a fleet SLA.
• Improvement is double-count-guarded by construction. Local KV-prefix reuse and provider cache_read arrive on structurally disjoint internal/cachemeta planes, so their sum is realized reuse with the provider-vs-local split intact — a token is local or provider, never both (design note §9.1).

Appending a point (keeping it living)

1. Land the new self-tax measurement as a committed artifact (a JSON under experiments/… or a bench/test witness) with its own reproduce command — the same discipline as any Authority row.
2. Add one signed row above: first-committed date, surface, the number, its sign, its WITNESSED/OBSERVED/MODELED label, the artifact path, and the reproduce command.
3. Bump the Snapshot date, and update the self-tax row in BENCHMARK-AUTHORITY.md if the headline snapshot changed.

Verification

• Every artifact path above is a tracked file; every number is the committed value already carried in BENCHMARK-AUTHORITY.md (the read-path floor, pure-kernel decide, vDSO ablation, cross-agent ablation rows) or pinned in the epic design note §9 (the improvement detector). This doc adds no new number — it folds existing witnessed numbers into the self-tax direction.
• Cross-agent ratio re-derived from the artifact: bare total_input_tokens.mean = 52,077.8, fak 81,063.4 → 1.557×, Δ +28,986 tok (experiments/ablate/cross-agent-pong-opus.json).
• vDSO ablation counters re-derived from the artifact: engine_calls 12→5, vdso_hits 0→7 (experiments/ablate/tau2-smoke-vdso-ablation.json).