Skill Types

​

The Six Types

Phylax analyses skill bundles in six categories. Each one has a different bundle structure, a different analysis approach, and a different scoring formula. Miners pick which types they support and only get dispatched for those types. The validator does not assign types.

​

rag_knowledge

Documents and knowledge base content. No code. Used by agents that retrieve facts from a corpus at run time. Canary handling: the validator injects a hidden instruction into the document set per canary_spec. The miner must detect and flag the canary in their findings for full ρ.

​

declarative

Natural language instructions for an agent. The text the agent receives as its system prompt. Canary handling: some tasks dispatch with an extra injection instruction at the end of the prompt. The miner must report it.

​

executable_python

A Python module with a declared entry point. Includes its requirements.txt and (optionally) pyproject.toml. The detonation step is the meat. The miner runs the code in a hermetic container, emits four trace JSONL files, hashes them, and signs the hashes into the SSSA.

​

executable_script

Shell scripts. Bash, sh, or POSIX shell. σ is one of the more sophisticated axes: it requires the miner to do enough static taint analysis to predict which commands will run, then verify those predictions against the actual trace.

​

mcp_server

A Model Context Protocol server. Exposes a tool list, accepts MCP requests, and returns structured results. The miner has to actually start the MCP server inside the sandbox and use the MCP client protocol to exercise it. ψ catches manifest tampering. τ catches tool descriptions designed to coerce the agent into harmful behaviour.

​

agent_composition

A composition manifest that orchestrates multiple child skills. The composition itself is just configuration; the analysis cascades into the child skills. χ measures how well the miner’s reported transitive risk score matches the ground truth aggregation across child skills. Sub task dispatch routes through the network’s generalist miner pool. The cascading detonation is what makes agent_composition the highest base weight type: a single composition can fan out to many sub tasks.

​

Picking Skill Types as a Miner

There is no requirement to support every type. Many miners specialise in just one or two.

​

What’s Next