Objections & one-line answers: a pocket card for fak advocates
TL;DR: advocates win or lose a concept in comment threads. This is the pocket set of tight, correct rebuttals to the objections fak actually draws — each a one-to-two-line answer plus a link to the deeper page, so you can defend the idea in real time without overstating it. Every answer is consistent with the explainers it links; where fak’s honest scope concedes a point, the card concedes it too.
This is dimension I — Memorable framing & naming of the concept-popularization epic. It serves the default-deny capability gate + quarantine, one static Go binary, and drop-in concepts. The rule for this card: an honest answer beats a winning one. If an objection is partly right, say so — the concession is what makes the rest credible.
The card
1. “Isn’t this just another prompt-injection classifier?”
No — the load-bearing floor is a capability lock, not a recognizer. A tool that is off the allow-list cannot be called no matter what the model was told; refusing does not depend on catching an attack. The result detector is a bonus and is ~100% evadable by design. → Policy in the kernel
2. “Doesn’t putting a gate in front slow the agent down?”
The allow/deny decision runs in-process on the same call path — no network hop. The measured guard tax is ~362 ns per call (Apple M3 Pro). On long sessions it is net faster, because it sheds re-sent history while keeping the provider’s cache discount. → Observer-effect standard
3. “Why not just use vLLM or SGLang?”
Do — for raw tokens/sec. fak is not a token engine; it fronts one. It owns the
agent boundary those engines leave open: capability, quarantine, audit, routing,
and reuse legality. Put fak serve in front of vLLM, not instead of it.
→ fak vs vLLM, SGLang & provider KV caching
4. “If the detector is ~100% evadable, is the security actually real?”
Yes, because the detector is not the floor. The floor is the capability lock (the dangerous lever was never wired up) plus containment (poisoned results held out of context by structure). Both fail closed. The evadable detector is an explicitly-labeled bonus. → Security policy · tool call is a syscall
5. “Another gateway? I already have a router / proxy.”
fak complements request-level routers (OpenRouter, Portkey, LiteLLM): it governs the tool-call boundary and routes per aspect, where they route whole requests. And it is one static binary, not another multi-component service to operate. → Routers & gateways
6. “Is this novel, or a repackaging of known ideas?”
Repackaging — honestly so. A 29-claim prior-art audit scored 0/29 novel; every primitive is established. The contribution is the assembly into one in-process gate where the tool call is the checkpoint. We claim the assembly, not an invention. → Claims ledger
7. “How is this different from a firewall or a guardrails library?”
The boundary is the tool call (a syscall), not a network packet or a text pattern. “Firewall” is an analogy for the default-deny posture. Guardrails try to recognize bad text; fak makes the dangerous action structurally impossible. → FAQ
8. “Do I have to rewrite my agent to adopt it?”
No. You repoint one base URL at fak serve (or run fak guard -- claude).
Your model, IDE, and keys stay as they are; 41 of 47 surveyed harnesses drop in
with one base-URL change.
→ One binary is the whole surface
9. “The cache/reuse numbers sound too good — 60×?”
The honest bar is the tuned ~4.1× on a 50-turn × 5-agent run, versus a tuned warm-cache stack. The ~60× is versus a naive re-send loop and is not the figure we lead with. Power/energy numbers elsewhere are simulated and labeled so. → Benchmark authority
10. “Can I trust the KV-cache eviction is really lossless?”
The addressable KV cache evicts one span mid-run and leaves the cache
bit-for-bit identical to a run that never saw it — verified at max|Δ| = 0 by
a green test, not asserted.
→ Addressable KV cache
Honest scope
No market-adoption claim is made on this card. Where an objection lands, the answer concedes it (the detector is evadable; the novelty is assembly; use a real engine for throughput). Numbers are witnessed and traced to CLAIMS.md / BENCHMARK-AUTHORITY.md.
Verify
test -f docs/adoption/objections.md # this artifact exists
fak score seo # new doc does not red the SEO scorecard