FBN.2026.05
ScopeCompose
Two-person research build · in progress
Python, PyTorch, DeBERTa-v3, Contriever, 2026
5
classes of conflict, told apart
4
scope relations between rule and exception
54
feasibility probes, verified by hand
300
benchmark instances in the build plan
The problem
Conflict-aware RAG resolves disagreement by choosing the most credible source. But "international transactions incur a 3% fee" and "fees are waived for premium cardholders" do not contradict each other; both are true under different scopes. Selecting a winner deletes the exception and the condition that governs it, and standard answer-correctness metrics cannot even see the loss.
The approach
A pipeline that treats conditional conflict as a first-class type. Hybrid retrieval feeds a two-stage detector; a five-class classifier, compared across three architectures, labels each disagreement; a four-way scope analysis decides whether the exception refines, sits beside, repeats or opposes the rule, with an order-invariance check. Conditional pairs are routed to composition instead of selection, producing a scoped answer that keeps every valid branch. DeBERTa-v3 cross-encoders and Contriever run locally; Groq handles the judgement calls, with a local Ollama model as the offline fallback.
The outcome
In progress, and documented as it goes: the Suppression Rate metric suite that makes the deletion measurable, a benchmark of naturally occurring rule-and-exception pairs, and a 54-case feasibility probe whose gate has closed. When the first annotation pilot failed on one axis, the failure was diagnosed, written up and fed back into the manual.
What it does
4 capabilitiesSuppression Rate
A metric suite that registers when a valid, separately scoped branch of the truth has been dropped, the failure answer-correctness cannot see.
Composition, not selection
Refinements and disjoint exceptions are composed into one scoped answer; only genuinely opposed claims fall back to choosing a source.
Order invariance
A scope decision must not change when the passages swap places. The pipeline checks.
Evidence kept apart
Naturally occurring pairs and synthetically split pairs are never blended in headline results.
The strata
How it standsRetrieval
Hybrid retrieval with Contriever embeddings
Detection
Two-stage pairwise conflict detection: DeBERTa-v3 NLI filter, LLM escalation
Classification
Five-class conflict typing, three architectures compared
Scope analysis
Refinement, disjoint, redundant or opposed, with an order-invariance check
Extraction
Contrastive scope probing behind a grounding gate
Composition
Scoped answer generation that preserves every valid branch
The judgments
The reasoning, on the recordChose
Composing a scoped answer
over selecting the most credible source
A rule and its exception are both true. Choosing between them is the error the project exists to measure.
Chose
A frozen JSON contract between the two halves
over one codebase that changes in lockstep
Two researchers build detection and composition independently, and meet only at a versioned record that neither can change alone.
Chose
Groq with an Ollama fallback on one code path
over a single model provider
The study runs at zero spend. Swapping providers is a configuration change, never a rewrite.
Next exhibit — VI
Computational Entropy Lab