{"path":"research/nesy-rag-note.md","content":"# NeSy-RAG: A Convergent Arrival from the QA Side, and One Design Yield\n\n**Date**: 2026-08-20 · **Source**: Gann & Gertz (Heidelberg), *NeSy-RAG: Neuro-Symbolic RAG for Explainable Question Answering*, [arXiv:2608.06292](https://arxiv.org/abs/2608.06292), founder-relayed. **Read-depth**: abstract + method sketch + sample table (the full-text fetch returned only the abstract level); the ShARC characterization below draws on prior knowledge of that benchmark and is survey-grade.\n\n## What it is\n\nAn LLM reads retrieved text chunks and synthesizes **Prolog predicates encoding Boolean claims**, some depending on user facts; queries compose the predicates; execution is deterministic, with **traces linking each reasoning step to its originating source**; and a **symbolic knowledge-gap detector identifies missing user facts whose truth values would affect the query outcome**, auto-generating the follow-up question. On ShARC (conversational reading of regulatory text — UK-government benefit rules, where the system must answer or ask the right follow-up), it scores 61.1% against a same-model RAG baseline's 42.8%, without domain training.\n\n## Three readings for this corpus\n\n**1. A structure-pays datapoint, in exactly the domain where structure was predicted to pay.** LLM-extracts → symbolic-core-reasons → traces-attribute is the arsenal-and-harness split and the world-models thesis (*coherence must be imposed, and the question is where*) implemented in Prolog instead of a typed graph — and the 18-point gain is measured in a **rule-application domain**, the structured-environment case where [world-models-and-the-ontology-revival.md](world-models-and-the-ontology-revival.md) already warned the enterprise numbers come from. So: real evidence the imposed-structure bet wins where rules exist and compose; **no license to transfer the number to open-ended argumentation**, which is not ShARC. It also lands as another independent arrival at the reasoning-layer slot from yet another side (single-user QA over rules) — adjacent, not overlapping: no contestation, no multi-party, no strength semantics, no value layer.\n\n**2. The design yield: knowledge-gap detection is hinge-guided questioning, mechanized.** Their detector asks only the follow-up whose ANSWER would change the outcome — which is \"no gap displayed without a move\" plus outcome-sensitivity, fused into one mechanism. We hold both halves separately: the hinge computes which descendant, granted or denied, moves the root; the descent asks critical questions in fixed scheme order. **The composition is buildable and cheap: order the descent's open questions by hinge sensitivity, so the machine always offers the question that matters most first.** That is a strictly better instance of the taps-not-essays scaffold (the seven holds in [depth.md](../depth.md) § The Descent, Defined), and it reuses shipped machinery (`hinge.py`) rather than adding any. **Built same day**: `deliberus/question_ordering.py` clamps each question decisively-yes/decisively-no and recomputes the claim's badge through the shipped semantics (`badge_inputs.py`, extracted behavior-identically from the badge endpoint so no constants fork); the CQs endpoint now serves open questions in swing order with the sensitivity shown beside them. One measured humility note preserved in the tests: two intuitive monotonicity claims about QEM dilution were refuted by running them — the pinned invariant is consistency-with-the-badge, not folk saturation.\n\n**3. The nearer beneficiary is the household automation, not Deliberus.** ShARC's material IS the benefit-rule domain (marriage allowance eligibility, condition trees over user facts) — a different project in this fleet reads Swedish authority letters and reasons over exactly such rules. The pattern *encode the rule text as predicates once, detect which user facts are missing, ask only outcome-relevant follow-ups, answer with a source-linked trace* is directly applicable there, and a pointer has been filed in that project's TODO. For mannaminne the relevance is thin (it is retrieval, not rule QA); the one transferable technique is joint natural-language/code embeddings for predicate retrieval, noted without action.\n\n## Cross-references\n\n[world-models-and-the-ontology-revival.md](world-models-and-the-ontology-revival.md) (coherence-imposed; the structured-domain caveat) · [structure-versus-scale.md](structure-versus-scale.md) (the evidence ledger this feeds) · [depth.md](../depth.md) § The Descent, Defined (the scaffold the yield improves) · `deliberus/hinge.py` (the shipped machinery the yield reuses)\n"}