AI Safety for Multi-Agent Task Negotiation, Streaming Results, and Agreement Version Consistency

Autonomous agents and distributed task processors increasingly negotiate, amend, delegate, stream and cancel shared tasks, yet systems commonly conflate a proposal with an accepted agreement, a terminal stream marker with successful work, and a cancellation with reversal of earlier effects. This paper defines evidence-qualified consistency reconstruction for incrementally negotiated and streamed tasks. It distinguishes task proposals from accepted agreement versions; task status, stream closure, artifact finalisation and observable external effects; and historical subaction execution from residual obligations under amendments or cancellation. The method is a 44-stage, six-phase common workflow (E-CW-01) that returns CONSISTENT, INCONSISTENT or UNRESOLVED through a scoped three-valued conjunction in which a contradiction dominates and missing evidence is never treated as a negative result. We present eleven inventive steps, six embodiments, a formal model, forty implementation checkpoints, and a reproducible laboratory that implements a deliberately limited pure in-memory evaluator in Python, Go and JavaScript (52 shared fixtures; 166 test executions, all passing). Expected latency is modelled analytically (1.95 ms in an assumed budget); measured evaluator timings are about 0.16 to 2.0 microseconds per call and cover only in-memory evaluation. An agreement assessment conveys no authority to execute a consequential action. The laboratory is self-authored, not peer reviewed, not independently audited and not field validated. Keywords: AI safety; multi-agent systems; task negotiation; streaming results; agreement versioning; partial acceptance; amendment impact; cancellation and late results; partition recovery; residual obligations; causal event order; artifact finalisation; three-valued evidence; closure verdict; synthetic laboratory reproducibility.