Force-Field is the agentic harness for safer AI from Spin State Labs. FORCE governs what the model says: no flattery, premises challenged, sources cited, reasoning shown, confidence tagged. FIELD governs what an agent may do: a per-agent manifest plus a local Enforcement Gate with a kill switch, protected paths, tool-call budgets and a hash-chained ledger.
A limit that matters is enforced by software at the point of action, not by asking a model to behave.
Consequential actions wait for approval from a second person, with a record of who approved what.
Work is routed by policy, sensitivity, cost and the task at hand, so a model can be changed without rebuilding.
Prompts, policies, knowledge, traces and evaluation sets belong to the organisation and can be exported.
Each step records what was asked, what was retrieved, which model answered and which checks ran.
Answers cite their sources, and an organisation's facts are validated by a person before agents use them.
Documentation is on its way. This site will hold it as it is written.