Skip to the content.

Token-saving-defaults scorecard — is fak’s out-of-the-box token economy amazing?

The question a cost-conscious operator asks the moment they run fak guard -- claude / fak serve: of every token-saving method fak knows how to stack, which ones are ON by default — and are the high-value, low-loss ones turned on out of the box, or left dark behind a flag nobody flips? Every number below is re-derived from the entrypoint source (cmd/fak/guard.go, cmd/fak/serve.go, the Default* constants in internal/gateway/gateway.go, and internal/gateway/messages.go) by fak token-defaults-scorecard — a lever’s on/off state is the binary’s real behavior, never a claim in the roster. The headline metric is token-defaults-debt: the count of concrete defects — a high-value saver left off, an on-by-default saver with no honest note, a default no test locks, a front door out of step. Driving it to zero means a user who runs fak with no flags gets the full stack of safe savings, each honestly labeled, none able to regress unnoticed.

Budget size is governed separately by the long-context defaults doctrine: the advertised context window is a hard cap, not a target, and resident budgets should be labeled as witnessed, observed, modeled, or fallback.

Regenerate: go run ./cmd/fak token-defaults-scorecard --markdown > docs/serving/token-defaults-scorecard.md

Headline

Metric Value
Token-defaults-debt (total HARD defects) 0
Composite score 100.0/100 (grade A)
Savers stacked on by default 7/7
Groups stack 100 · honesty 100 · regression 100 · parity 100
Advisory (soft) signals 0

Per-lever status — where each token-saving method stands

class: lossless = zero model-visible change (must be on); bounded = lossy but an in-code guard keeps the model’s working set intact (high-value → should be on, with a note); optin = broader blast radius (correctly off, must carry a documented gate). gated = an off lever documents why; noted = an on bounded lever documents what it sheds + cache-safety; locked = a test pins the default.

Lever Class Default Witness Blocker Flag Gated Noted Locked
provider_cache — provider prompt-cache prefix (byte-faithful passthrough) lossless ON (structural) ·
toolfloor — tool-floor pruning (drop provably-unreachable tool defs) lossless ON (structural) ·
vdso — vDSO dedup fast path (collapse identical calls) lossless ON --vdso ·
compacthistory — history compaction (drop the un-cacheable middle past the budget) bounded ON --compact-history-budget ·
elideresult — oversized-result elision (shrink a scrolled-past tool_result to head+tail) bounded ON --elide-result-bytes ·
elidestale — stale-read elision (replace a Read superseded by a later same-file edit with a restorable marker) bounded ON --elide-stale-reads ·
ctxview — ctxplan O(1) planned view (re-materialize history under a budget) bounded ON --ctx-view-budget ·

KPIs

Group KPI Score Debt Detail
stack stacking_depth 100 0 7/7 token-saving methods stacked on by default out of the box
stack lossless_stack 100 0 3/3 lossless savers on by default
stack high_value_defaults 100 0 4/4 demonstrably-safe bounded-loss savers on by default
honesty witness_status 100 0 no off-by-default high-value savers remain — every bounded-loss saver defaults on
honesty dark_lever_gated 100 0 0/0 off-by-default levers carry a documented gate
honesty default_notes 100 0 4/4 on-by-default bounded savers carry an honest loss note
regression default_on_locked 100 0 7/7 on-by-default savers pinned by a regression sentinel
parity entrypoint_parity 100 0 front doors agree + servewiring verdicts track the real defaults

Token-defaults-debt work-list

No token-defaults-debt: every stacking saver fak can safely default is on out of the box, honestly noted, and locked against regression. 🎉