Self-Describing Rewrite Intelligence

Self-Describing Rewrite Intelligence (SRIM) is a formal model for continual inference–learning systems whose internal state is a finite typed (hyper)graph that contains its own rewrite-rule representations as subgraphs (“rules-as-data”). The paper proposes an operational No-Meta interface: SRIM does not assume any externally correct auditor, reward signal, or environment-side evaluator. Instead, guarantees are formulated as internal self-verifiability relative to (i) a sealed, authenticated snapshot handle of the recent introspection ledger and (ii) a minimal trusted boundary (microkernel / TCB). SRIM unifies “inference” (rewriting task-relevant data) and “learning / reflective change” (rewriting embedded rule representations and self-descriptions) under a single certified rewriting semantics based on DPO (double-pushout) graph rewriting in an (M-)adhesive setting, including nested application conditions. To make classical confluence/termination results operationally meaningful for self-modifying systems, SRIM introduces a seal–freeze–commit discipline: each macro-event is processed by sealing a snapshot, freezing the inference-active rule region, executing a kernel-managed finite burst where schedulers only propose steps, and committing a burst-end record only after kernel-verified idle on an explicitly capped active workspace (or returning explicit failure statuses such as TIMEOUT / STUCK / FAIL_CLOSED). A central contribution is committed-outcome determinism: if (a) a per-step termination gate enforces strict decrease of a kernel-declared rank template, and (b) kernel-verified local confluence succeeds via bounded critical-peak enumeration and replay checking, then any two admissible schedulers that both reach an OK (idle-certified) commit produce identical committed interface records. SRIM also formalizes introspective legibility by attaching to each commit a rule-set digest and a kernel manifest (digest + readable body) that specifies the exact checker-visible attributes, snapshot query interface, caps/budgets, canonical readout procedure, and verification policies—supporting auditability from within the system’s own sealed interface under explicit resource bounds.