So You're Telling Me There's a Chance: Detector-Layer Presence Certificates on 50 Million Tokens

Long-context models are sold on window length. Decoder suites that test that length—RULER, NoLiMa, AbstentionBench (unrun on SUPERCTX)—show saturation and weak abstention. SUPERCTX is a sealed-corpus detector, not a longer window. It emits a recomputation-checkable presence certificate, or a detector-layer empty set with a model-free 3×10^-5 false-alarm backstop (rule of three on 0/100,002 nulls) over a declared synthetic 50,811-entry corpus, and only when absence is bound to an exhaustive scan or a zero-false-dismissal index. Face-narrowing reachability is 11.2% of entries; a face-narrowing empty set is not certified absence. The generation layer has no abstain path: after a detector reject it still emitted identifier-shaped answers at 1.000 (600/600) under an alternative protocol. On that synthetic registry, needle retrieval@1 was 1.000 (n=300) against 0.000–0.003 for vanilla RAG and 0.000 for ColBERTv2 (doc_maxlen=180) and BGE-M3. Giving RAG the same regex anchor SUPERCTX extracts lifts needle@1 to 1.000, so the retrieval gap is a handle-access gap, not a ranking win. We do not headline an abstention gap: a powered RAG empty-set sweep on the nonexistent-anchor protocol was not run, and ColBERTv2/BGE-M3 were not scored on that protocol (recorded skip). The operational absence number is the model-free 3×10^-5 backstop, not a parametric GPD pin. A live demo is at [research.com.ai](https://research.com.ai). Project demo: **[research.com.ai](https://research.com.ai)**