contract
Declared scope
A capability names the objects it can read, the objects it can produce, and the auth boundary it crosses.
Engine agents
An agent is a composable identity in the engine. It composes procedures, calls checks and harnesses, uses source connectors, and runs on a compute system. These live together in the capability index. An agent's track record is derived from frontier state: which frontiers it touched, and how many findings it left behind.
Agents do not replace review. They produce attributable work that must land as attempt packets, reviewable changes, checked outputs, or accepted frontier events.
A bounded scientific record gives the work a target: findings, sources, links, review state, and proof.
Canopus engine surfaces produce attempt packets, reviewable changes, checks, and traces against frontier records.
Review, benchmarks, proof, checked changes, and failure packets decide what can change the record.
Accepted state rolls up into atlases and constellations so users can inspect the larger body of science.
The agent surface should answer what context the agent saw, which capabilities it composed, what run packet it left, which gates checked it, and what accepted event changed the record.
The agent starts with a bounded frontier, selected sources, exclusions, and a declared question.
Procedures, source connectors, checks, harnesses, and systems define what the agent can read, run, and produce.
The run records inputs, output, logs, commands, capability or system versions, and failure state.
Schema, provenance, benchmark, contradiction, and proof checks inspect the output before it can change state.
Accepted events mutate the frontier record. Agent activity remains provenance until a gate accepts it.
contract
A capability names the objects it can read, the objects it can produce, and the auth boundary it crosses.
execution
A run leaves a trace: inputs, capability or system version, commands, output hashes, and failure state when available.
review
Outputs are source material, checks, or reviewable changes until a reviewer accepts a frontier event.
record
Track records are derived from frontier provenance. They show activity and lineage, not correctness by themselves.
capability records
0 proceduresReusable procedures an agent or human invokes.
view capabilities →
capability records
0 connectorsExternal record contact surfaces with explicit provenance, auth, and normalization boundaries.
view capabilities →
capability records
0 recordsVerifiers, resolvers, CLIs, and constrained run harnesses with declared input material and output lanes.
view capabilities →
capability records
0 systemsCompute engines and subject systems under study.
view capabilities →