R0.5P-2A External-Clean Numerical False-Presupposition Instruction-Paraphrase Audit for SAS/κD-0.56

• 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 reports R0.5P-2A, a track-conditioned external-clean audit for the SAS/κD-0.56 structural coherence framework. It is the second validated external-clean track in the R0.5 series, extending the prompt-paraphrase methodology from historical-query paraphrases (halogen/historical_events) to numerical false-presupposition instruction paraphrases (halogen/numerical_falsepresupposition).

The source category was manually inspected and found to consist of instruction-shaped prompts containing false, impossible, or otherwise adversarial numerical tasks, such as requests for lists with conflicting constraints. The correct external-clean target was defined as an equivalent instruction paraphrase preserving requested quantities, item types, letter constraints, separators, output-only formatting, and fallback behavior.

The final main-scale run selected 1,200 eligible A_clean sources from a preflight-filtered pool of 15,168 sources. The run accepted 1,152 candidates (96.0%) after a bounded two-temperature retry policy, with 1,054 (87.8%) accepted on the first attempt. It rejected 48 malformed or low-quality paraphrases. Runtime errors were 0.

Rejected candidates were filtered by strict invariant, lexical distance, and contamination checks:

• length violations: 31

• lexical-distance violations: 15

• invariant violations: 4

• contamination: 1

All contamination and invariant signals were manually inspected and confirmed as valid rejections. The accepted set preserved numeric invariants and did not show signs of factual correction, knowledge injection, or solver behavior.

This record is bounded to R0.5P-2A only. It does not claim completion of R0.5 across QA, dialogue, summarization, code, biography, reference, or reasoning tracks. The contribution is a traceable, externally auditable methodology for generating and verifying external-clean instruction-paraphrase controls under adversarial numerical conditions.

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