SAH-SPI-DDB: An Auditable Structured Reasoning Framework for Language-Structure Tasks

Large language models exhibit three structural weaknesses on language-structure tasks: constraints are enforced through prompts or post-hoc filtering, decisions cannot be replayed, and local repairs require whole-document rewriting. We present SAH-SPI-DDB, a three layer framework in which SAH compresses a linguistic object into a finite discrete skeleton, SPI constructs structurally consistent paths between skeletons, and DDB drives diagnosis and repair with a layered weighted difference ∆, enforcing boundary checks and rollback at every step. We give an interface-level formalization, a six-method contract, propositions on termination and determinism, and a hash-chained audit trail. On a runnable prototype, a small contract instance drops from two hard plus one soft violation to zero; at 25 clauses, repair uses 13% fewer edits than first-improvement search, and random improvement fails to converge in all 20 trials. Against large models on ten instances, our method converges 10/10 versus 8/10 for a reasoning model, with mean changed clauses 11.1 versus 17.2 (paired sign test p = 0.002); at 49/75 clauses each model fails on one of three variants and rewrites 1.6–1.8× more clauses. On ten real regulations, citation repair converges 10/10 and touches only 3– 12% of articles; for tasks with enumerable constraints and a small action space, strong models match our method. We therefore limit our claim to action-composition and minimal-edit language-structure tasks, and do not claim to replace Transformers.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-20
DOI
https://doi.org/10.5281/zenodo.24395181
Primary Topic
Natural Language Processing Techniques
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

SAH-SPI-DDB: An Auditable Structured Reasoning Framework for Language-Structure Tasks

Youjie Li, Shixi Wang, Baoyuan Zhang
Zenodo (CERN European Organization for Nuclear Research)
Natural Language Processing Techniques
preprint

SAH-SPI-DDB: An Auditable Structured Reasoning Framework for Language-Structure Tasks

Youjie Li, Shixi Wang, Baoyuan Zhang
preprint en

Abstract

Large language models exhibit three structural weaknesses on language-structure tasks: constraints are enforced through prompts or post-hoc filtering, decisions cannot be replayed, and local repairs require whole-document rewriting. We present SAH-SPI-DDB, a three layer framework in which SAH compresses a linguistic object into a finite discrete skeleton, SPI constructs structurally consistent paths between skeletons, and DDB drives diagnosis and repair with a layered weighted difference ∆, enforcing boundary checks and rollback at every step. We give an interface-level formalization, a six-method contract, propositions on termination and determinism, and a hash-chained audit trail. On a runnable prototype, a small contract instance drops from two hard plus one soft violation to zero; at 25 clauses, repair uses 13% fewer edits than first-improvement search, and random improvement fails to converge in all 20 trials. Against large models on ten instances, our method converges 10/10 versus 8/10 for a reasoning model, with mean changed clauses 11.1 versus 17.2 (paired sign test p = 0.002); at 49/75 clauses each model fails on one of three variants and rewrites 1.6–1.8× more clauses. On ten real regulations, citation repair converges 10/10 and touches only 3– 12% of articles; for tasks with enumerable constraints and a small action space, strong models match our method. We therefore limit our claim to action-composition and minimal-edit language-structure tasks, and do not claim to replace Transformers.

Zenodo (CERN European Organization for Nuclear Research)
Natural Language Processing Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon