15 — Risk Register and Failure Modes
Risk 1 — Reverting to verification/ranking
Symptom: final output is top worlds or claim labels but no synthesized SNO.
Mitigation: every run must emit SNO lineage and synthesis status.
Risk 2 — LLM judgments
Symptom: an LLM judge decides which narrative is true.
Mitigation: LLMs can propose or render; proof gates and calibrated models decide promotion categories.
Risk 3 — Predicate overfit
Symptom: latent predicates explain training contradictions but fail held-out examples.
Mitigation: MDL penalty, held-out synthetic contexts, grounding gates, false predicate rate.
Risk 4 — Access-state misuse
Symptom: missing records are treated as evidence.
Mitigation: access-state critic; separate absence-of-evidence from evidence-of-absence.
Risk 5 — Round-trip drift
Symptom: synthesized text re-grounds into a different logic state.
Mitigation: orthesis loop; $\chi_{LL}$ threshold.
Risk 6 — Topology theater
Symptom: topology terms appear but metrics are not used in decisions.
Mitigation: make beta-1, holonomy residual, and topology diagnostics part of acceptance criteria or remove them.
Risk 7 — Grounding destroys synthesis
Symptom: system becomes conservative fact checking and never creates new SNOs.
Mitigation: preserve Synthesizer and predicate invention; classify hypotheses separately instead of blocking all novelty.
Risk 8 — Synthesis hides contradiction
Symptom: fluent narrative erases unresolved conflict.
Mitigation: residual contradiction section required in audit report.
Risk 9 — Prior-art soup
Symptom: doc reads like a collection of known systems.
Mitigation: keep the SNO synthesis flow visible and evaluate module interactions/ablations.
Risk 10 — Dataset leakage
Symptom: labels or synthetic generator parameters appear in runtime.
Mitigation: oracle-boundary audit, split hashes, prompt scans.
Risk 11 — Citation hallucination
Symptom: evidence IDs do not resolve.
Mitigation: reject invalid inputs; citation validity required check.
Risk 12 — Calibration laundering
Symptom: likely claims are written as strict claims.
Mitigation: output category enforcement and confidence language table.