Evidence-Conformant Pre-Generation Control and Completion Decisions in EQ-λ RAG v0.1

This report evaluates an evidence-first control architecture for a locally executed retrieval-augmented generation system. Retrieval evidence intended to remain usable without a generative model is formed first; v0.1 does not assess its relevance, correctness, or usefulness to maintainers. A pre-generation controller then selects RETRIEVAL_ONLY, GENERATE, or PARTIAL from the evidence state, model state, execution budget, and remaining deadline. A post-generation evidence contract prevents an answer from being marked COMPLETE when its citation set is not a subset of the supporting-source set. On macOS, twelve fixed holdout queries were divided equally among unique_known, ambiguous, and insufficient classes. In baseline B0, generation was attempted for all twelve queries; eleven timed out and all twelve ended as PARTIAL. In controlled condition B1, four queries took RETRIEVAL_ONLY, four took GENERATE, and four became pre-generation PARTIAL. Five results ended COMPLETE and seven PARTIAL, but only one COMPLETE result came from generation. The preserved results recorded zero false-complete cases, zero hidden-label runtime reads, and 24/24 replay agreement. On Windows, the retrieval foundation comprised 1,378 maintenance-history records, six standard operating procedures, and two drawing records, for 1,386 vectors. Every self-query reached the top eight. Four non-top-one cases were explained by exact-score ties, leaving zero unexplained non-top-one cases. Rebuilt float32 vectors were not bitwise identical to stored vectors, but satisfied the predefined numerical-equivalence criteria, with a maximum difference of one ULP. Independent re-verification passed. A transfer envelope reconstructed from preserved Mac actuals matched 30/30 decisions, although twelve cases were case-ID substitutions and only eighteen were independent observations. The Windows pre-generation controller, model-residency observation, cold/warm timing, deadline contract, ten-key input, and port-8798 end-to-end path remain unimplemented or unmeasured.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-09-09
DOI
https://doi.org/10.5281/zenodo.22676354
Primary Topic
Advanced Database Systems and Queries
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
preprint

Evidence-Conformant Pre-Generation Control and Completion Decisions in EQ-λ RAG v0.1

Daisuke Tsunemori
Zenodo (CERN European Organization for Nuclear Research)
Advanced Database Systems and Queries
preprint

Evidence-Conformant Pre-Generation Control and Completion Decisions in EQ-λ RAG v0.1

Daisuke Tsunemori
preprint en

Abstract

This report evaluates an evidence-first control architecture for a locally executed retrieval-augmented generation system. Retrieval evidence intended to remain usable without a generative model is formed first; v0.1 does not assess its relevance, correctness, or usefulness to maintainers. A pre-generation controller then selects RETRIEVAL_ONLY, GENERATE, or PARTIAL from the evidence state, model state, execution budget, and remaining deadline. A post-generation evidence contract prevents an answer from being marked COMPLETE when its citation set is not a subset of the supporting-source set. On macOS, twelve fixed holdout queries were divided equally among unique_known, ambiguous, and insufficient classes. In baseline B0, generation was attempted for all twelve queries; eleven timed out and all twelve ended as PARTIAL. In controlled condition B1, four queries took RETRIEVAL_ONLY, four took GENERATE, and four became pre-generation PARTIAL. Five results ended COMPLETE and seven PARTIAL, but only one COMPLETE result came from generation. The preserved results recorded zero false-complete cases, zero hidden-label runtime reads, and 24/24 replay agreement. On Windows, the retrieval foundation comprised 1,378 maintenance-history records, six standard operating procedures, and two drawing records, for 1,386 vectors. Every self-query reached the top eight. Four non-top-one cases were explained by exact-score ties, leaving zero unexplained non-top-one cases. Rebuilt float32 vectors were not bitwise identical to stored vectors, but satisfied the predefined numerical-equivalence criteria, with a maximum difference of one ULP. Independent re-verification passed. A transfer envelope reconstructed from preserved Mac actuals matched 30/30 decisions, although twelve cases were case-ID substitutions and only eighteen were independent observations. The Windows pre-generation controller, model-residency observation, cold/warm timing, deadline contract, ten-key input, and port-8798 end-to-end path remain unimplemented or unmeasured.

Zenodo (CERN European Organization for Nuclear Research)
Advanced Database Systems and Queries
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.

Evidence-Conformant Pre-Generation Control and Completion Decisions in EQ-λ RAG v0.1 — Daisuke Tsunemori · Zenodo (CERN European Organization for Nuclear Research) (2026) | TGRS Research Map | TGRS