Beyond Code: Temporal Supersession Generalizes to World Knowledge
Oracle experiments on synthetic world-knowledge facts hit 1.000 accuracy / 0.000 stale. The mechanism is perfect; remaining error is extraction. Here is what Paper 5 proves and what remains.
The breadth thesis
Papers 1–4 are largely code-native. Paper 5 asks whether deterministic supersession is a general memory principle. We built a marker-free synthetic world-knowledge generator (CEO, population, price-style value flips on fictional entities to avoid real-world temporal ambiguity).
With LLM extraction (code-specialized 3B/7B models on prose), temporal_v6 reached 0.85 accuracy / 0.075 stale vs naive RAG 0.60/0.225 and advanced RAG 0.45/0.550. Better — but not zero residual.
Oracle isolation: residual is extraction, not mechanism
The decisive experiment builds the ledger from gold triples (no LLM extraction). Result: temporal_v6 1.000 accuracy / 0.000 stale. The entire residual vanishes when keys are correct. Conclusion: the supersession mechanism generalizes perfectly to world knowledge; remaining work is a domain-appropriate NL extractor and real temporal NLP benchmarks.
What comes next
Paper 5 is green on mechanism, gated on (1) a general instruction-model extractor for prose facts and (2) real suites (TEMPO, ECT-QA) using bi-temporal as_of(T) reads already built in the ledger. Entity alias tables matter when “USA” and “United States” co-refer.
For product: the same spine that keeps your codebase current can keep organizational and domain facts current — once extraction matches the domain.