> ## Content Index
> Fetch the complete content index at: https://www.gauravalbal.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# What did that clean history earn?
- URL: https://www.gauravalbal.com/what-did-that-clean-history-earn/
- Published: 2026-02-06T22:11:07.000Z
- Updated: 2026-10-22T16:21:27.000Z
- Description: Clean history should make growth easier. It should not make the next $5,000 automatic.
- Author: Gaurav Albal
- Tags: Token Kiting

The awkward part of the last article comes after you make the exception.

The customer uses the first allocation. Pays again. Comes back. At some point, somebody marks the account trusted.

I would expect us to make life easier for that customer. Otherwise we have built an onboarding policy with no way out of onboarding.

Now give them six clean $100 payments and a request for $5,000 of immediately available credits. I would like to know what exactly the six payments have earned.

They are evidence about the account. They are not a reason to stop noticing the size of the next request, how quickly it can be consumed, or the payment that just failed.

A good account can genuinely need more room. An account preparing to leave us with the bill can need more room too. Past behavior matters; so does the behavior we are being asked to permit next.

Let's give the history the benefit of the doubt, then look at the larger decision.

## Six payments. What have we actually seen?

Back to Relay, the fictional inference reseller. This time the account is A29\. Nobody has helpfully told us how its story ends.

A29 has completed six $100 funding cycles. At our usual 80% compute-cost assumption, that bought $480 of compute and generated $120 of contribution. The largest simultaneous unresolved compute commitment was $80\. Now the account wants $5,000 of credits, which could authorize $4,000 of provider cost.

Those are three different descriptions of the past: $600 of sales, $120 of contribution, and an $80 peak commitment. None is a $4,000 test. The historical contribution is also not collateral. We can decide to risk some of what the relationship has earned, but that should be a budget decision rather than something the word “trusted” does for us.

For the next exercise, grant the strongest version of the history: the previous funding really was good. We do not need bad historical labels to make this problem interesting.

In practice, “nothing has gone wrong yet” may be a weaker observation. Stripe notes that recent payment cohorts can appear to have lower dispute rates because customers may take weeks to file disputes.[1](#tk03-source-1) The age of the account does not give every new payment the same history.

## A clean streak is evidence against some attacks

Suppose the adverse accounts we worry about usually cause trouble quickly. Six genuinely good payments would help distinguish A29 from that group.

Now suppose building a clean prefix is common among the accounts that later make a large adverse draw. The exact same six payments would tell us less.

That is not an exotic hypothetical in credit. The Federal Reserve describes synthetic identities building creditworthiness before purchasing high-value goods and services and disappearing.[2](#tk03-source-2) It establishes that cultivated history can be part of an attack. It does not establish how often this happens at an inference provider, or that A29 is doing it.

Decision 01 / History · authored thought experiment

### Six payments. Two explanations.

**A29 at Relay · the observed history does not change** 

$100good funding$100good funding$100good funding$100good funding$100good funding$100good funding

Historical contribution

$120

Previous peak unresolved compute

$80

New credit request

$5,000

Potential new provider cost

$4,000

All six prior payments really were good in this example. They bought $480 of compute. No future outcome for A29 is revealed.

What does this history establish about the larger request? 

It supports an increaseKeep the old limitI need to know what the history distinguishes

Change the assumed process, not the history 

Adverse behavior usually appears earlyAdverse users often build this history first

The selected process changes only the assumed frequency of this whole history among adverse next requests. These are not fitted rates.

Assumed risk before this history

—

Risk conditional on this history

—

Expected loss if $4,000 is consumed

—

What changed? Inspect the assumptions and calculation 

B is adverse funding on this specified next request under the stated service treatment. H is the **whole six-payment history**. All inputs concern that same eligible population and request. H is applied once, not as six independent tests.

Prior probability of B (%)

History frequency if B (%)

History frequency if not B (%)

P(B | H) = p × P(H | B)
           ────────────────────────────────
           p × P(H | B) + (1−p) × P(H | not B)

Equal likelihoods preserve the prior. A history more common among adverse next requests can raise risk. A history assigned zero probability by both groups leaves the calculation undefined, not zero-risk. No recommended line is calculated here.

Reset history assumptions 

Read the two-world comparison instead 

Assume a 2% adverse rate for the specified next request before using H. In both worlds, 80% of good next requests have H. Only its frequency among adverse next requests changes.

__Same six payments; different likelihood of that history under an adverse process.__
| Authored process               | H if adverse | Risk after H   | EL on $4,000 |
| ------------------------------ | ------------ | -------------- | ------------ |
| Trouble usually surfaces early | 5%           | 1/785 ≈ 0.127% | $5.10        |
| Patient adverse users build H  | 80%          | 2%             | $80          |

Clean history is informative against the first process. Against the second, it does not distinguish good and adverse next requests. These invented likelihoods are not estimates for a merchant, and A29's future is not disclosed.

The first setting makes the history substantially reassuring. The second does not. We changed an assumption about which customers produce the history, not the observations on the account.

This is worth being explicit about because a policy that graduates customers after a clean streak already contains a view about the underlying process. It does not escape that assumption by being a simple rule instead of a model.

We also do not get six independent pieces of evidence merely because there are six rows. The exercise describes the likelihood of the whole history. Someone deliberately building it is not necessarily making six unrelated decisions.

I would not throw out the history. I would stop asking it to establish more than it can.

## The larger request belongs in the risk estimate

There is a second issue even when the clean history is genuinely informative. The new decision is different.

A29 has performed well under an $80 peak commitment. It is now asking us to permit $4,000\. That could be ordinary growth. It could be the point at which an adverse strategy becomes worthwhile. The amount alone does not choose between those explanations.

What we need is an assessment of this next commitment: the requested scale, current funding, recent behavior, and what the previous treatment actually let us observe. I would not take a probability that was useful for small requests, multiply it by whatever larger line sales wants, and call the result calibrated.

The distinction from the last article matters. There, we did not observe the unrestricted outcomes of rejected requests. Here, I have given you genuinely good historical outcomes. We still have to decide how far they carry into a different treatment.

**I would want the model to score the next commitment, not congratulate the account on the last six.**

## What happens when good standing stops describing the account?

Suppose the larger request goes through. The current funding episode makes $4,000 of compute drawable. A29 has consumed $400; $3,600 remains.

Then consumption accelerates and a subsequent funding attempt fails. Neither observation makes the account a fraudster. They are reasons to revisit the decision rather than let the previous approval suppress the question.

To isolate what happens next, I am supplying an illustrative conditional adverse probability: **2% before the update, 4% after it. These are assumptions for this episode, not estimates derived from the failed payment or outputs from the previous toy.**

Give the episode a $100 expected-loss budget and assume no recovery on adverse funding. At 2%, the full $4,000 fits: $80 of expected loss. At 4%, it does not: $160.

The old six payments have not changed. The current decision has.

Decision 02 / Effective response · a separate calculation

### Does the old allowance survive the new information?

**At the moment usable information arrives** 

Compute already consumed

$400

Funded compute still drawable

$3,600

Burn while the old treatment continues

$1,200 / hour

Episode expected-loss budget

$100 initially

**Supplied conditional risk: 2% before, 4% after reassessment.** Neither is derived from the payment failure, nor from Toy 1\. The probability stays fixed during each run of this short response window.

Delay from usable information to an effective control 

Immediate

Immediate effect3 hours later

The old allowance remains usable during the delay. No new funding, recovery, or informative event is added.

Apply the updated estimateReset episode

Budget-feasible upper bound — not a recommendation or a maximum realized-loss guarantee.

1 / Irreversible cost

#### COGS committed by effective response

**—**

$400 before the signal + — during response delay.

2 / Budget already occupied

#### Expected loss already committed

**—**

Evaluated at the supplied 4%. Budget: $100.

3 / Still controllable

#### Maximum further compute within budget

**—**

This is additional to cost already committed, not a new allowance that resets the episode.

**What does the restriction cost if the funding is good?** 

Contribution from the further grant

—

Contribution already made during response delay

—

Contribution forgone versus serving the remaining demand

—

At the original decision moment, all $3,600 of remaining genuine demand would contribute $900\. Each $1 of good compute contributes $0.25\. Demand is divisible; withheld work is not recovered later in this example. Costs and contribution before the signal are excluded from this incremental comparison.

Revisit the risk assumption or budget 

Unchanged state · 2%Reassessed state · 4%

This changes an authored conditional probability, not the elapsed time or observed history. The 2% case allows full growth within the original budget. A fresh run of the same model does not replenish headroom.

Assumed current adverse probability (%)

Expected-loss budget for the whole episode ($)

Explicit zero is supported as a mathematical boundary. An empty input is missing information and cannot be used as zero. This budget is not a claim about the largest loss the business can bear.

Inspect the calculation at effective response 

x = min(remaining funded COGS, burn × delay)
U = COGS at signal + x
EL already committed = p × U
q_max = min(remaining COGS − x,
            max(0, budget / p − U))

At p=0, all remaining funded compute is feasible for any nonnegative budget. If the existing exposure already exceeds the expected-loss budget, q\_max is zero but the breach remains. A high enough assumed adverse probability can make further compute unprofitable even when a large budget permits it.

Read the response-delay example instead 

With 4% supplied risk and a $100 expected-loss budget, faster enforcement preserves more of the remaining decision. It does not undo any cost.

__All amounts are USD provider cost or expected loss as labeled.__
| Effective response | Committed COGS | Committed EL | Further ceiling             |
| ------------------ | -------------- | ------------ | --------------------------- |
| Immediate          | $400           | $16          | $2,100                      |
| 15 minutes         | $700           | $28          | $1,800                      |
| 1 hour             | $1,600         | $64          | $900                        |
| 1h45m              | $2,500         | $100         | $0                          |
| 3 hours            | $4,000         | $160         | $0; $60 already over budget |

Immediate response at 4% permits a further $2,100: $525 of contribution if genuine rather than $900, forgoing $375\. At unchanged 2% risk, all $3,600 of further compute is feasible, with $80 expected loss on the full episode. This comparison does not recommend either real-world probability or a hold schedule.

With an immediate response under the 4% assumption, $400 has already been committed. That occupies $16 of the expected-loss budget. The remaining $84 can accommodate at most another $2,100 of compute.

That is a **budget-feasible upper bound**, not my recommendation. It says nothing by itself about whether the workload is useful, whether the customer is profitable, or whether the business can tolerate the realized loss if the adverse outcome occurs.

Now delay the effective response by an hour. At the example's $1,200-per-hour burn rate, another $1,200 has gone. The cost already committed is $1,600, its expected loss is $64, and the maximum further allocation falls to $900.

At one hour and 45 minutes, the full $100 expected-loss budget is committed. At three hours, all $4,000 has been consumed and expected loss is $160\. Setting the remaining allowance to zero does not put us back within budget. It leaves us $60 over it.

The $60 is an expected-loss overrun, not the bill from an observed reversal. If the adverse funding outcome occurs, this simplified example loses the whole consumed provider cost.

Notice what the clock measures: time **after useful information is available**, before the changed control becomes effective. I have not given the faster policy an earlier warning. The contrast is how long the old decision keeps authorizing spend.

A score can update correctly while a cached exception continues doing exactly what it did before. I would want to inspect both.

## Do not turn every increase into a new onboarding process

Switch the second toy back to the unchanged 2% assumption. The full remaining allocation fits the budget. There is no credit for finding an excuse to hold it anyway.

Even in the restrictive 4% case, the good-customer cost needs to stay beside the loss calculation. At immediate response, another $2,100 of compute generates $525 of contribution if the funding is good. Serving the full remaining $3,600 generates $900\. The restriction forgoes **$375** in this example.

That might be the right trade. It is not automatically acceptable because the fraud number went down.

I have made the workload divisible and treated the excluded work as lost within this window. A customer who needs a complete batch may get no useful service from the smaller allocation. [The first article](https://www.gauravalbal.com/ai-credit-fraud-looks-suspiciously-familiar/#tk01-availability) lets that assumption defeat the policy. Repeating the word “provisional” does not repair the product.

There may be a different funding route, useful assurance, or a consciously larger risk budget that lets us serve the customer. There may not. The point of reconsidering the decision is to choose between those possibilities, not to send every growing account back to the start.

Nor does scoring more frequently manufacture information. During the toy's short response window, the assumed probability stays fixed. There are no new payments, recoveries, or useful observations. Re-running the model does not make the already-consumed dollars disappear from the budget.

## What should good standing actually change?

I would look at one favorable treatment already in use: an increased allowance, a shortened hold, an automatic release, or an exception someone approved.

What earned it? Which history, at what sizes, and with how much opportunity for adverse outcomes to appear? What can that treatment authorize now, including concurrent usage and new requests? What observations reopen it, and what actually changes when they do?

Those questions should work in both directions. A material payment failure or change in utilization should not be ignored because the customer graduated last month. A recovered payment problem should not necessarily condemn them to restrictive treatment forever.

Relevant evidence can come from the merchant or from outside it. Consortium information may describe behavior across institutions that one merchant cannot see. That is a possible source of evidence, not proof that any particular signal resolves this decision or that the toy's likelihoods describe a real portfolio.

The timing also needs an honest record. Event time, the time a signal becomes available, and the time a changed control takes effect are not interchangeable. Stripe notes that even an early fraud warning can arrive after the associated dispute because the reporting systems are separate.[3](#tk03-source-3) A backtest should not give a policy information before it could have used it.

I do not want a business where every good customer has to keep proving that they deserve to grow. I also do not want six small payments to buy a standing exemption from the next decision.

The useful question is what the history supports now: at this amount, on this funding, with this recent behavior, and before the next opportunity to act. My prior is that the old account-management machinery has quite a lot to say about that. The part I would want to test is how much of it your business needs.

How the examples work · assumptions and sources 

Relay and A29 are fictional. No employer policy, customer record, incidence estimate, or vendor performance is represented. A29's next outcome is not revealed. The historical sources establish mechanisms, not their frequency among inference users.

The history exercise grants genuinely good old funding. Its 2% prior and history likelihoods concern the same specified next request and full-service treatment. The whole prefix enters the calculation once. Its posterior does not feed the response exercise.

The response exercise supplies a conditional 2% or 4% probability for one already-funded episode and holds it fixed during the short window. There is one common adverse funding outcome with no recovery. No extra survival discount, new funding, repayment, or usable new evidence is added while the response is delayed. All API keys share the episode's budget. Zero delay means immediate effect with no additional in-flight work; a production analysis must include unavoidable in-flight cost.

Expected loss is not a realized-loss ceiling. Already-consumed cost remains in the episode's exposure constraint but is sunk when comparing further grants at response. The display distinguishes the further grant from all new spend since the signal, including the delay. Good compute contributes $0.25 per provider dollar; demand is divisible and unserved work is lost in this window. Fees, recovery, review cost, future customer value, alternative funding, and strategic response to a different policy are omitted. The largest budget-feasible grant is not a recommendation.

### Sources

1. [Stripe — Disputes analytics](https://docs.stripe.com/payments/analytics/disputes?ref=gauravalbal.com). Delayed disputes can make recent payment cohorts look artificially clean; no universal finality cutoff is inferred.
2. [Federal Reserve — Synthetic-identity fraud, July 9, 2019](https://fedpaymentsimprovement.org/news/press-releases/federal-reserve-system-white-paper-examines-the-effects-of-synthetic-identity-payments-fraud/?ref=gauravalbal.com). Cultivated creditworthiness followed by bust-out, not a token-reseller prevalence estimate.
3. [Stripe — How disputes work](https://docs.stripe.com/disputes/how-disputes-work?ref=gauravalbal.com). Early warnings and disputes are distinct; their reporting systems may be out of order.

## If this is happening in your business, I’d like to take a look.

You might be eating losses, blocking customers you want to keep, or maintaining an increasingly elaborate collection of exceptions.

**Email me a few lines about what's happening. I'll tell you where I'd start and what I'd need to see before recommending a change.**

[Email Gaurav →](mailto:hello@gauravalbal.com?subject=Token%20Kiting%20%E2%80%94%20our%20fraud%20problem)

*No customer data needed to start.*