SAS / κD=0.56 — R1 Real Local Structural Evaluation v1.0.7

• Gonzalo Emir Durante (Independent Researcher; Origin Node, Project Manifold 0.56; Symbiotic Autoprotection System)

Registry: TAD EX-2026-18792778 (Argentina)  
License: GPL-3.0 + Durante Invariance License  

Description

This record freezes the R1 real local structural evaluation snapshot for SAS / κD=0.56.

The release includes documentation, technical report material, evaluation scripts, manifests, logs, SHA-256 hashes, and frozen local evidence for the v1.0.7 structural evaluation.

The R1 real dataset contains 4698 paired rows, corresponding to 2349 paired source groups, split into 938 validation rows and 3760 held-out test rows.

The best non-runtime structural composite was Flow + CRE + Negation, score >= 1. On the held-out test split, this composite achieved:

• F1 = 0.8717

• Precision = 0.9952

• Recall = 0.7755

• Accuracy = 0.8859

This result supports the claim that SAS-light structural modules produce a reproducible, non-runtime, interpretable structural signal on the R1 real split.

Important limitation: the lexical baseline remains stronger on this split, with test AUC = 0.9963 and test F1 = 0.9968. Therefore, this release should not be interpreted as showing superiority over the lexical baseline or as a final hallucination detector claim.

v1.0.7 also fixes NIG JSON serialization. The full-run check confirms NIG status OK for 4698/4698 rows and zero conversion errors. In the current local configuration, NIG remains source-unaware and diagnostic-only.

This record should be cited as a reproducibility snapshot and technical report for the R1 real local structural evaluation of SAS / κD=0.56.

Keywords:
SAS; κD=0.56; external-clean audit; instruction paraphrase; numerical false presupposition; track-conditioned generation; contamination audit; hallucination audit; reproducibility.

Additional notes:

This record is a structural evaluation and reproducibility snapshot. It does not claim final SAS production validation, universal hallucination detection, or superiority over lexical baselines. Runtime-derived features are excluded from the main structural claim. NIG is fixed at the JSON serialization level in v1.0.7 but remains source-unaware and diagnostic-only in this local configuration.