Skip to content
JANUSOBSERVATORY

Research appendix · bounded multi-agent deliberation

How does a bounded agent council compare with Janus endpoints?

This compares a bounded policy-council transcript with withheld Janus endpoints. It is not a physical, social, or scientific simulation, and it is not a forecast.

Deliberation only—not simulation or forecast

The roles exchange arguments in a bounded council. They do not execute a causal world model, evolve physical state, or reproduce civilization dynamics. Each job receives only the source-backed Table 5 morphology context; Table 9 population, energy, and growth endpoints stay hidden until a deterministic comparison step. Any resemblance is a model-generated comparison result—not model validation, a probability, or a Janus fact.

Bounded council

MODEL ROLE

Governance Steward

Propose institutions and constraints while identifying whose agency and freedoms are affected.

MODEL ROLE

Resource Steward

Track material and energy constraints and reject unsupported claims of unlimited resources.

MODEL ROLE

Ecology Steward

Protect biosphere integrity and surface tradeoffs that other council members omit.

MODEL ROLE

Technology Steward

Explore technical options without treating technology as inevitable or uniformly beneficial.

MODEL ROLE

Population Steward

Examine demographic consequences and state uncertainty instead of inventing precise quantities.

MODEL ROLE

Observer Steward

Track potential technosignatures and never equate non-detection with absence of technology.

Ten seeded job specifications

S1seed 4101
S2seed 4102
S3seed 4103
S4seed 4104
S5seed 4105
S6seed 4106
S7seed 4107
S8seed 4108
S9seed 4109
S10seed 4110

Each job is capped at 12 council steps. The seed fixes local ordering, but a remote language model may still produce different text. Raw transcripts, model identity, source allowlist, protocol version, seed, prompt hash, and transcript hash are written to ignored local artifacts for audit.

Comparison contract

  • Population, annual energy, and growth state remain separate dimensions.
  • No composite similarity score and no scenario ranking are produced.
  • Canonical values remain reported/transcribed; agent values remain model_generated.
  • Arithmetic deltas are derived and never promoted into canonical scenario data.
  • The core Story, Observatory, and Atlas never load this optional runtime.
Return to Research Companion →