Limits and roadmap

​

Limits

• Agent output still requires review.
• The checker validates definition and guidance files. It does not yet validate the answers.yaml and ingest-log.md files that kb-ingest writes, or the evolve-log.md file that kb-evolve writes.
• Structural checking and Explorer rendering do not establish that an answer is true or that a definition captures the right question.
• Evals grade definition wording against the project’s written rules. The model tiers call paid models and can disagree; a case no tier settles goes to you. A pass does not show that a definition fits your work.
• just evals-play plays agent-behaviour scenarios with a scripted owner, one run each by default.
• The protocol defines composition and exchange rules. Knowledge Bus does not include tools that compose documents or interpret incoming knowledge.

​

Roadmap

​

Milestones

• An offline, dependency-free map for exploring definitions (Knowledge Bus Explorer)
• Skills for each step today: ingest a knowledge base, uncover questions and decisions, evolve definitions, and check and explore them
• A model-agnostic guided workflow: one coordinating agent with bounded helper agents, using whatever models and decision models are available, works with one person or a team of domain experts through before-and-after proposals to create or extract a universe, align an existing personal, team, or enterprise knowledge base to it, and evolve it when something triggers a change
• Composition rules: what a document holds, given its purpose and audience
• Tools that compose documents and interpret incoming knowledge
• Markdown interchange format
• A standard .knowledge-bus/ directory so tools can discover shared definitions
• Evals for definition wording (Evals)
• Evals for agent behaviour (Evaluation Report)
• Validation of the answers.yaml and ingest-log.md files that ingestion writes

​

Further exploration

• Knowledge exchange between different sets of definitions
• Forms and surveys designed around the decisions their answers support

​

Open questions

• How much reasoning is preserved when knowledge is extracted and turned back into a document?
• How reliably do agents follow the definitions and guidance?
• Do agents leave questions unanswered when they lack supporting information?