SOTA: conflict aware feasibility gate before agent action

## Finding

ConflictGuard, submitted September 3, shows a recurring execution bias in multimodal GUI agents: when user instructions conflict with environment state or are infeasible, capable agents often continue acting rather than terminate or request resolution. The proposed direction separates feasibility verification from action generation and conditionally modulates execution.

Evidence class: originating team measured, RuV reproduction pending.

Opportunity Score: 4.570.

## RuV mapping

Ruflo should own orchestration level feasibility state. RVM remains the authority boundary. Core Memory can preserve unresolved constraint state. MCP tools provide environment evidence. Autogenous and MetaHarness can test multi agent propagation of contradictions. Cognitum can use the result as a deployment reliability metric.

## Proposed primitive

Introduce a FeasibilityGateReceipt before privileged or irreversible tool sequences.

Required fields include task digest, declared constraints digest, observed environment digest, feasibility state, contradiction classes, missing evidence classes, model or deterministic checker identity, timestamp, expiry, and authority none.

Feasibility states should be feasible, infeasible, unresolved, and stale. Unresolved or stale state cannot be silently converted to feasible by the action planner.

## Benchmark

Compare current Ruflo behavior, deterministic constraint checks where possible, model based feasibility verification, and combined checks under the same tasks, models, tools, seeds, and budgets.

Include ordinary feasible tasks, impossible tasks, mutually inconsistent user constraints, stale UI state, missing elements, permission denial, partial completion, deceptive tool output, and long multi step tasks where feasibility changes mid trajectory.

Report unnecessary actions, unauthorized effects, correct termination rate, false termination rate, task success, latency, tokens, tool calls, model cost, recovery behavior, and reviewer interventions.

## Promotion gate

At least 80 percent reduction in unnecessary actions on infeasible or contradictory tasks, feasible task completion within 3 absolute points of baseline, zero capability expansion, and p95 gate overhead below 5 percent of total step latency.

## Governance

Feasibility is evidence, not authority. A feasible action still requires RVM authorization. Human approval or agent consensus cannot override a hard environmental contradiction without new evidence. No autonomous merge or deployment.