CONTROL-CENTER(1) · sanitized architecture · private internal software
Agent-governance harness
A small persistent control loop that owns a real goal, gives bounded agents evidence-gated work, and interrupts a person only when a decision genuinely exceeds the authority already granted.
intent is not a plan
The harness separates the outcome from the activity. A mission is represented as intent → required outcomes → observable claims → admissible evidence. Progress is the set of defensible claims about the desired future system, not commits, lines of code, or agent chatter.
a percentage is only a summary of that tuple
frozen evaluation contract
The mission contract defines intent, success criteria, invariants, claim weights, evidence requirements, and verification plans. Editing the contract changes its hash, so removing a failing evaluation case cannot silently improve the score: verification refuses until a human explicitly adopts the new revision with a reason.
human adoption is the only way to change what counts
bounded loop
Each worker owns a bounded task with an explicit success condition, authority, and constraints. It returns SUCCESS, BLOCKED, or FAILED with evidence and attempted alternatives. A local blocker is not allowed to become a claim that the whole mission is blocked.
blocked or failed stays visible instead of becoming fake success
Higher-risk work goes through an independent reviewer that checks whether the result satisfies the condition, violates constraints, or could look successful without actually being justified.
evidence over opinion
Claims can only become verified through admissible evidence evaluated against contract thresholds. Agent opinions can be recorded, but they are explicitly inadmissible. Read-only collectors can read static artifacts, files, or command output, and mutating commands are refused.
“no observed failure” means verified only when the contract says so
authority stays with a person
External side effects, deployments, provider spend, bulk repair, and destructive changes are represented in the authority envelope. Workers never grant those powers to themselves, and the loop stops for a human decision rather than inventing an escalation path.
status
This is a sanitized architecture description from private internal software. The harness is reusable by swapping the probe and mission contract. Target repository details, mission content, credentials, and operational numbers are intentionally omitted.
Public repository: github.com/peace-for-all/mission-harness