skip to content

Transparency commitments force per-field reasons on every model decision. What do you actually negotiate?

level: principalimportance: nice to knowfreq 22%

answer

  1. the yes/no is already settled
  2. width, precision, recipient, rate, consistency
  3. who is the duty actually owed to
  4. coarse bands versus six decimal places
  5. price the residual, name the owner

basics

~20 s

Not whether to explain, but the shape of the disclosure: how many fields it names, at what numeric precision, to which recipient, at what rate, and at what price. The obligation usually runs to the decision subject, one record at a time, not to a machine integrator pulling thousands a day.

solid answer

~50 s

The yes/no is already decided, so a reviewer who argues for switching explanations off has misread the room. The negotiable dimensions are width — how many input fields the disclosure names; precision — a signed rank order versus magnitudes to six decimal places; recipient — the decision subject receiving one account of their own case, against an integrator receiving thousands through an API; rate and price — whether explained calls have their own counter and their own cost; and consistency — whether re-submitting a record returns identical numbers. The cheapest large reduction is usually recipient: the obligation is almost always owed to the affected party about their own decision, and satisfying it through a subject-facing channel rather than a bulk machine interface removes most of the attack value while fully meeting the commitment. Then price the residual honestly and say what you will not claim: 'we rate limit' is not an answer when nobody can say what one reply contains.

code

text · 10 lines
text
Integration spec (vendor deck, verbatim)
----------------------------------------
Every decision ships with its reasons.
contributions returned: all 47 input features, signed
magnitude precision: 6 decimal places
price: identical per call with or without contributions
rate limit: 10,000 calls / day / integrator
metering: one counter for all calls; explained calls not separated
recipient: any party holding an integration key
...

go deeper

for a junior

Know that returning reasons with decisions is often required, and that the form those reasons take is a choice rather than a fixed output.

for a middle

Be able to list the dials — how many fields, at what precision, to whom, at what rate — and say why each changes how much model information leaves per call.

for a senior

Show you would re-derive the abuse budget after the payload changed, meter explained calls separately, and refuse a rate-limit claim that never states what one reply contains.

for a principal

Own the whole call: satisfy the obligation through the narrowest channel that discharges it, price the residual against the value of the integration, and record who accepted it.

## The decision that is actually in front of you A vendor deck or a procurement questionnaire says: *every decision ships with its reasons.* You are the reviewer deciding what that promise is worth on the attack side before signing, or the owner deciding how to honour it in your own product. The commitment itself is not up for debate — it exists because someone is owed an account of a decision made about them, and that is a good reason. What is up for debate is its shape. A lead who answers this by proposing to withhold explanations has failed the question. So has one who waves the exposure away. The work is in the dimensions. ## The five dials **1. Width.** How many of the model's input fields does one disclosure name? All forty-seven, or the three that dominated? Every named field is a coordinate of local response handed over. Narrowing genuinely reduces per-call disclosure — but be honest about the limit: across thousands of records, even short rank lists assemble into a picture of which fields drive the decision. Width is a dial, not an off switch. **2. Precision.** A signed rank order, a coarse band ("strong, moderate, slight"), or magnitudes to six decimal places? The human being owed an account needs the first; only a machine fitting something benefits from the third. Numeric resolution is frequently the largest free reduction available, because nobody chose full precision — it was the default in whatever produced the numbers. **3. Recipient — usually the biggest lever.** Transparency duties are typically owed to the *subject* of a decision, about their *own* case, one at a time. They are not owed to a commercial integrator receiving a bulk feed. Routing explanations through a subject-facing channel, and not through the high-volume machine interface, can satisfy the commitment in full while removing most of the extraction value. If the deck's promise is written as an API feature rather than as a subject entitlement, that is the first thing to interrogate. **4. Rate and price.** Are explained calls counted on their own meter? Do they cost more than bare scores? If the answer to both is no, every rational caller takes the high-disclosure path and the operator cannot even measure how much of its traffic is high-disclosure. Pricing is one of the few levers that acts on volume rather than content. **5. Consistency.** Does re-submitting the same record return identical numbers? Stability is good for the subject and good for the copyist. This is a real tension with no clean resolution, and naming it is better than pretending it away. ## What you accept, and say out loud After the dials are set, exposure remains, and a principal's job is to price it rather than to eliminate it: - What is the business loss if a competitor holds a ranked list of what drives your referral decisions? Frequently this is larger than the loss from a partial functional copy, and it needs no modelling at all. - What is the integration worth in revenue and distribution? A control that kills the integration is not a control, it is a decision to exit a channel — which may be right, but must be argued as such. - Who signs off that the residual is acceptable, and is that recorded anywhere someone can find it next year when the payload changes again? ## The claims to refuse - **"We rate limit, so extraction is bounded."** Refuse it until someone can say what one reply contains. A cap denominated in replies means nothing until you multiply by the disclosure per reply. - **"They are post-hoc approximations, so nothing leaks."** Approximate model information is model information; the sign and ranking are the parts with value. - **"They are a contracted partner."** That determines whom you can sue, not what they receive. Assume every key holder has the disclosure. - **"Explainability is a compliance matter, not a security one."** It is the same disclosure viewed from two chairs, which is exactly why this decision has to be made once, by someone who can see both. ## The recommendation that survives review State the obligation and who it is owed to. Propose the narrowest disclosure that discharges it — subject channel, coarse magnitudes, the fields that actually drove the outcome. Meter and price the machine path separately. Write down the residual exposure and who accepted it. That is a decision a reviewer can sign, and it does not require anyone to choose between transparency and security in the abstract.

  • Would cutting the disclosure to the top three drivers solve it?
    It reduces per-call width, which is real, but it is not a fix. Across thousands of records the ranks still assemble into a picture of what drives the decision, and that aggregate is often the commercially sensitive part. Treat width and numeric precision as dials that lower the rate of disclosure, then decide whether the residual is acceptable — do not present narrowing as closure.
  • The vendor answers 'we cap integrators at 10,000 calls a day'. What is your response?
    Ask what one call returns. A cap in replies is meaningless until it is multiplied by the disclosure per reply, and forty-seven signed magnitudes at full precision make ten thousand calls a very large disclosure. Then ask whether explained calls are metered separately, because if they are not, the vendor cannot tell you how much of their traffic is on the high-disclosure path.
  • Who in the organisation should own this call?
    One owner who can see both chairs, because the compliance view and the security view are looking at the same disclosure and will otherwise optimise against each other. In practice that is whoever owns the output contract of the product. The decision needs to be recorded with its residual, so that the next change to the payload triggers a re-derivation rather than an assumption.

saying these in an interview costs you the question

  • Proposes switching explanations off to close the exposure
  • Treats explainability as compliance only, never a disclosure decision
  • Accepts a call cap without asking what one reply contains
  • Presents narrowing to top-three drivers as a complete fix
  • Never asks who the transparency duty is actually owed to
  • Trusts a contracted partner not to retain what they receive

context

resolving links…