Skip to content

Benchmark and player model

GPTNT measures how configured players collaborate and act in KTANE experiments. A model is one part of a player. The distinction matters because the same model can participate through different providers, protocols, processors, settings, or identities.

Model, player, and service instance

Term Meaning in GPTNT
Model The inference target that receives the prepared model input and returns a response.
Provider The client and endpoint configuration used to reach a model.
Player configuration The Hydra configuration that assembles a model, provider, protocol support, processors, limits, identity, and recorder.
Player The assembled participant that turns observations and messages into actions or messages.
Player service One running process that exposes a configured player through Redis RPC.

Several service processes can use the same player configuration. A run manifest controls their count. The included test players exercise the pipeline without external provider credentials, so not every player is backed by a hosted model.

Roles and access

KTANE defines two collaboration roles:

  • The Defuser sees and acts on the bomb.
  • The Expert receives messages and uses the selected manual to advise the Defuser.

A player protocol specifies the role, communication style, manual access, and feedback behaviour for one side of an experiment. A suite can also use a solo Defuser protocol. Do not infer that every room always requires two player services.

Roles, protocols, and capabilities explains the supported pairing modes and the distinction between protocol fields and service matching.

The suite supplies the protocols. The generated experiment specification records the selected player names and protocols. When the experiment manager creates a running instance, it adds each service UUID and the capabilities reported by that service.

Capabilities and identity

Capabilities state properties such as supported modalities and image handling that affect whether a player can take part in a selected experiment. The runtime heartbeat reports resolved capabilities. GPTNT copies them into the experiment instance and recorded provenance.

Player identity and capability fingerprints are part of result interpretation. Changing model settings, configured capabilities, processors, or other identity inputs can produce a different participant even when the display name remains the same.

The following relationship stays constant across interactive runs:

Configuration boundary Runtime boundary Recorded boundary
Model, provider, player profile, and processors Player service assigned to a protocol role Player identity, resolved capabilities, steps, and footer

Benchmark outcomes

The game state determines the declared experiment outcome: solved, strikeout, or timeout. A model response alone is not a benchmark result. GPTNT relates the outcome to the suite, mission, protocols, player identities, capabilities, runtime instance, and provenance captured by the run.