A scoring API's owner asks which anti-extraction control still binds next quarter - what do you say?

## The question being asked An owner of a paid relevance-scoring API wants to know what the anti-extraction controls are worth. Not what they are - what they will still be worth after an adversary has read them and repriced. That reframing is the whole answer, because every control on this surface is a price, and prices are repriced around. ## Rank by spreadability The single most useful sorting rule: does the control raise a cost the adversary can spread across more identities and more weeks, or one they cannot? **Per-call quotas - spreadable, therefore near zero.** A limit attached to a key is divided by funding more keys. Against a party willing to spend, it changes the identity count and nothing else. Its residual value is real but small: it prices out the unfunded and the careless, bounds accidental scraping, and produces a billing record. Its residual cost is not small - it is our own high-volume buyers who meet the ceiling first. **Output coarsening - spreadable, and it charges our customers.** Fewer bits per reply raises the replies needed for a given fidelity, which is a multiplier the adversary absorbs. The same lost precision reaches the buyer whose bid is computed from that score, and they cannot buy it back with volume. So this control is best described as a transfer: paying callers give up resolution so that an adversary's bill is multiplied by a constant. Whether that trade is worth making is an argument with two numbers in it, not a configuration change. **Distribution signals - the only informative one, and decaying.** Traffic that seeks coverage looks unlike traffic that follows demand, and that difference is a property of the query set rather than its size. It buys attribution and time rather than prevention, it exists only if identities can be joined into one actor, and it fades as soon as the adversary pays extra replies to draw product-like traffic. Worth having; worth stating with its expiry. **Identity friction - the term that does not spread, paid by the funnel.** Everything above is divided by the number of accounts. What is not divided is the cost of bringing one account into existence and the risk of accounts being linkable afterwards. That is where the adversary's unspreadable money sits. It is also where our growth sits, which makes it a product decision owned jointly - not a control a security team turns up on its own. ## What I would refuse to write down I would not sign a register entry reading *rate limited, extraction risk mitigated*. It records a divisor as a barrier. The direction of the claim matters: a quota that has never been exceeded tells us about the quota; a distribution monitor that has never alarmed tells us about our definition of normal; a coarsening change with no customer-side number tells us half of a trade. If the register needs a line, it should carry an estimate of what a usable copy costs to obtain at our published prices, beside what our controls currently cost paying customers. ## The decision I would actually put to the owner The function is sold by the call. While that is true, a sufficiently funded party can obtain a copy of it, and the levers available adjust how much they pay and how visible they are, not whether they can. So the decision is not *how do we stop this* but: - what fidelity of copy are we content to have loose, and does a copy at that fidelity actually take our revenue - or is our value in freshness, inventory access, integration and the data behind the model rather than in the function itself; - how much of our own product quality and signup conversion are we willing to spend to raise their bill; - do we want attribution when it happens, which is a design requirement on identity linkage and on retaining query shape, decided before an incident rather than during one. An accepted risk written down with its reasoning is worth more than a control whose only measurable effect is on our own customers. Where attribution after the fact matters commercially, that is a different lever with its own evidentiary questions, and it belongs in its own conversation rather than being folded into a rate-limit ticket. ## What next quarter looks like If the adversary is funded and patient, next quarter the quota is worth what it was worth this quarter - a divisor - and the distribution signal is worth somewhat less than today, because it is the one they have an incentive to buy their way around. Identity friction is the only line item whose value depends on a decision we can still make. Saying that plainly, with the customer-side cost attached to each option, is the answer the owner needs; ranking the controls without pricing what our own callers lose is the failure mode.