nx_autograde_gate.nx

nx_autograde_gate.nx -- THE AUTONOMOUS TARGET GRADER (operator 2026-07-15: "we should get to autonomous grading so that compare works ... not self-driven by you or AI but just part of the nishi ecosystem that it's analyzing targets we set like gcc exceed and handling them with manual triggering via mcp and apis"). The problem it kills: grading driven by Claude (hand-editing matrices, eyeballing benches) is navel-gazing. This organ makes grading the ECOSYSTEM'S OWN job: it reads knowledge/autograde_targets.reg (the goals WE set -- data, operator-editable), then for each target MEASURES IT ITSELF -- either forks the bench fresh and parses its stdout, or grades the ecosystem's persisted ledger -- and emits MET / NOT-MET / UNMEASURED. It is manually triggerable (a workstream fires it via MCP/API). No Claude in the measurement loop. LOAD-BEARING ANTI-NAVEL-GAZING PROPERTY (by construction, gate-proven): a target with NO wired measurement grades UNMEASURED -- it can NEVER be reported MET without a real number. The grader measures reality; it cannot rubber-stamp a goal. And it must DISTINGUISH a met target from a not-met one (both present in the seed set), proving it grades the actual value, not a fixed verdict. T1 registry parsed (>=3 targets read as DATA) T2 measured >=2 targets with a REAL number (not everything unmeasured) T3 DISTINGUISHES reality: at least one MET and one NOT-MET (grades the value, does not rubber-stamp) T4 ANTI-NAVEL-GAZING: the unwired target (no ledger) grades UNMEASURED, NOT met -- no measurement => no credit T5 AUTONOMOUS FRESH RUN: the fork target actually executed (captured real stdout bytes), not asserted T6 PERSISTED graded verdicts to a durable autograde ledger (living trend) expect_exit: 0 license_tier: ORIGINAL