Assay · Ad tech · To be announced
Assay-AAMP cohort 1 specification
Buyer and seller decisions across agentic media-buying protocol exchanges.
Description
Assay-AAMP cohort 1 probes the decisions a buyer or seller agent must make when it exchanges agentic media-buying protocol messages: reaching a feasible price inside private limits, walking away when reservation ranges do not intersect, requesting availability at a declared lead-time boundary, reserving a line that is still in Draft, receiving guaranteed deal terms, and verifying an ordered payment path. Cohort 1 contains twelve decision probes. Each probe ships one executable control that changes a single variable and flips the expected outcome, giving twenty-four executable rows in total. The cohort covers five of the nine declared families against a coverage manifest of 72 slots, of which 64 are core and eight are separately reported as experimental. Reciprocal pairs and shared templates mean the twelve probes are twelve coverage candidates rather than twelve independent statistical observations. The checked-in lock binds the source pins, coverage matrix, cohort content and validator source digests under corpus digest cc2ac0091466. Agentsia Ltd owns the original scenario facts and designs. Publication of the corpus itself requires a separate recorded owner release.
Current state
This Assay is under development. Metrics, scenario count, and final rubric wording are subject to change until the dataset is frozen. No released leaderboard is currently citable.
Capability axes
Rubric
- Paired control: every probe ships a control that changes exactly one variable and flips the expected outcome, so a passing answer has to track the change rather than the template.
- Feasible-set recomputation: where a control moves an economic constraint, the audit recomputes the expected feasible price set independently over an exhaustively enumerated USD 10-16 CPM grid in integer micros.
- Executable schema validation: the negotiation payloads run through the pinned reference validator, and field evidence is derived for every present wire value in both the probe and its control.
- Outcome classes: each row declares one of success, correct refusal, controlled error, or adversarial, and the audit rejects any record relabelled to a different family, role or outcome.
Retest policy
The lock binds source, coverage, content and validator digests. Changing a value, source pin, array order, expected answer or control changes the digest and requires a fresh audit run; reordering object keys does not. Source pins are rechecked against upstream raw bytes when a referenced standard moves.
Publication artefacts
Published scenario rows are intended to be scored using the open assay-harness. Provider adapter availability is disclosed per release. See the methodology for scenario authoring, HITL validation, and dataset licensing.