nx_schemadrift.nx

nx_schemadrift.nx -- THE CONTROL FOR THE SAFETY-FLAG DERIVER (2026-08-25). Re-derives every auto-derived row in knowledge/tool_schemas.conf from the organ's TRANSITIVE IMPORT CLOSURE and partitions the disagreement BY DIRECTION, because the two directions are not the same defect: over-warning costs information, under-warning costs safety. WHY IT EXISTS. tool_schemas.conf says in its own header that it is there to give "an honest safety profile so an agent knows BEFORE CALLING whether a tool mutates state or reaches outside". Its auto-derived rows are stamped "safety flags STATIC-DERIVED from the source's own syscalls" -- and that stamp is exactly true and exactly the problem: the source's OWN syscalls are not the tool's reach. An organ that writes through an imported helper reaches the write syscall in the LIB, so a single-file scan publishes readOnly=1 for a writer. A DERIVER MUST MEASURE ITS FALSE-POSITIVE RATE AGAINST CONTROLS BEFORE ITS OUTPUT IS TRUSTED, EXACTLY AS A DETECTOR MUST -- and this is that control, run over the FULL population and never a sample. THE PARTITION SUMS. rows = agree + permissive + conservative + mixed + unproven, printed, so a leak is visible rather than absorbed. UNPROVEN is its own bucket and never falls into AGREE: an organ whose source cannot be resolved or whose marker table cannot be loaded produced NO EVIDENCE, and scoring that as agreement is how a census flatters itself. exec reach is reported as a SEPARATE ADVISORY AXIS, not folded into the quad: neither shipping deriver claims it, and a new bucket that overlaps an existing partition breaks the reconciliation silently. nx_schemadrift [audit|worklist|row <name>] license_tier: ORIGINAL No hw writes (Rule 26). expect_exit: 0