Capability · Autonomy Contract
Every consequential
decision, on the record.
An agent that works while you're away still hits forks — which database, which auth scheme, cursor or offset. The Autonomy Contract makes those forks explicit: each one is enumerated, classified, and either answered or deferred to a recorded safe default. You set the guardrails; the agent works inside them.
Three verbs
Enumerate. Resolve. Check.
A plain-text ledger at .darkfactory/decisions.yaml — versioned with
your code, reviewed in the same PR. No service, no database, nothing to deploy.
How the contract holds
Classified, resolved, and carried into the session.
Every entry carries a resolution class (doctrine-answerable, research-answerable, agent-decidable, human-owned) and a reversibility — reversible or irreversible. The cost axis is what governs how much authority a fork demands.
An answer records who decided it and the source. A deferral names the safe default it falls back to. Anything still open is a decision nobody has made — and check exits non-zero until it is.
Cerebe injects the baseline doctrine — SOTA-first, the iteration ceiling, the parking-lot, and the decision protocol itself — into an agent at session start. The contract is present from the first line, not bolted on after.
The ledger's sibling is .darkfactory/objectives.yaml — a text-hash-bound
manifest of what done means, derived from the same brief. Decisions capture
what was chosen; objectives capture what must be proven. Two plain files, versioned
with the work.
Get Started
Author the guardrails once. Check every run against them.
A versioned decision ledger and a doctrine hook — native to your repo, reviewed in the same PR, checkable in CI. Part of the free-to-use Cerebe Software Factory CLI.