What the model requested. What actually ran. Synthetic British female narration: Emma (Kokoro). 0–9.4s | A typo becomes a real failure The model asks for a command with a typo. OpenCode runs it, and the validator exits with an error. That failure becomes a saved learning. 9.4–20.1s | A correction becomes an explicit rule We check the correction and compile it with Terraphim. The exported knowledge graph uses synonyms syntax. Saving a suggestion alone does not switch rewriting on. 20.1–30.9s | The hook rewrites. The tool executes. In a fresh session, the model requests the typo again. Terraphim returns the correction. The adapter applies it, and OpenCode executes the corrected command. 30.9–41.7s | Which rule changed this command? The execution record keeps the original request, the final command, and rule one. A hash identifies the exact compiled graph used for this execution. 41.7–53.4s | The result belongs to this execution The tool exits successfully. Its result and saved file carry the same execution identifier. The receipt also records the file hash, so changed bytes are detected. 53.4–65.6s | One execution. Evidence you can inspect. Regression tests pass. A separate verifier checks the real host events, rule snapshot, result and file together. Download the example to repeat the run and inspect the evidence.