What must stay private?

Share a proposal without sharing every constraint.

A buyer’s internal budget ceiling is different from an offered price. Roy and Singh propose keeping sensitive constraints local while exchanging negotiation messages and proving some constraints cryptographically. Their architecture also includes selective cloud offloading; “device-native” should not be read as a promise that no processing ever leaves the device. [S4]

Our interpretation is to define a private mandate, an explicitly shareable proposal, and rules for disclosure between them. That is a design direction for future agent functionality. The current website’s local cost example is not an implementation of the paper’s cryptographic protocol.

Privacy extends beyond encrypted storage.

The device-native paper explicitly acknowledges that timing and concession patterns can reveal proximity to a reservation price. It evaluates controlled bilateral scenarios and calls for field validation. A cryptographic constraint check can establish a narrow proposition without resolving behavioral inference, deliberate manipulation, or offline collusion. [S4]

A future pilot would therefore inspect which information is sent, which system processes it, what the conversation reveals over time, and who may inspect the record. A tamper-evident log can help check whether a record changed; it does not establish that the recorded judgement was commercially sound.

Counterparty messages are data, not authority.

The large negotiation competition discusses prompt injection that elicited opponents’ private bargaining information. That result is relevant as a failure mode to test, rather than a tactic to adopt. A counterpart’s message should not be able to rewrite the buyer’s authority, disclose the internal mandate, or turn a draft into an authorized commitment. [S2]

Proposed controls include checking structured offers against a separate policy, limiting tool permissions, and escalating unexpected instructions. These controls require adversarial evaluation; a reassuring prompt is not evidence that the boundary holds.

Translate governance into operating decisions.

NIST’s Generative AI Profile is voluntary cross-sector guidance. It addresses confabulation, privacy, information security, and human–AI configuration, and recommends actions spanning Govern, Map, Measure, and Manage. Referencing it here is not a certification or a claim of NIST endorsement. [S6]

For procurement, plausible but unsupported market claims need source checks. Ambiguous payment or delivery terms need clarification. A limit breach needs escalation. These are our concrete applications of the guidance, not procurement-specific instructions quoted from NIST.