Skip to content

Test harness and multi-agent architecture

Three agents participate in every bootstrap test run:

AgentRoleModel
BootstrapThe thing under test — runs the bootstrap skillVaries (test matrix dimension)
Domain ownerSimulates a human answering ontology questionsFixed (Haiku — cheap, credible)
JudgeReads repo state and scores against rubricFixed (Opus — stable scorer)

The domain owner is intentionally constrained: seed concept hints are anchors, not definitions.

Scenario schema

Scenarios drive test runs:

id: <kebab-case>
domain: <one-line domain description>
central_question: <the question this repo exists to investigate>
seed_concepts:
  - name: <string>
    hint: <one-sentence anchor>
good_bootstrap_looks_like: <1–2 sentences describing a successful result>

Snapshot path

tests/results/<id>/<model>/<YYYY-MM-DD>/
  structural.json
  quality.json
  snapshot.json