SAS Multimetric Tribunal: Experimental Validation κR=0.15 / κD=0.56

This deposit contains the implementation and experimental validation of the SAS Multimetric Tribunal, a zero-GPU, zero-LLM structural classification system for hallucination detection in AI-generated text. The tribunal evaluates (source, response) pairs across seven detection modules using a cascading multiplicative penalty strategy. It partitions the output space into three structural zones governed by two experimentally validated thresholds: κD = 0.56 (coherence boundary) and κR = 0.15 (collapse boundary). Validation was performed on 600 pairs (300 control A→A + 300 hallucination A→B) from HaluEval and TruthfulQA benchmarks. Results:- F1: 99.16%- Precision: 100.00% (0 false positives on evaluated corpus)- Recall: 98.33%- Accuracy: 99.17%- ISI separation: 0.8323 Structural zones:- Structural Collapse (F-S): ISI < 0.15 — CRITICAL- Recoverable Rupture (B): 0.15 ≤ ISI < 0.56 — HIGH- Coherent (A): ISI ≥ 0.56 — LOW Architecture:ISI_HARD = min(lexical_baseline, source_target_guard, cre_isi)ISI_FINAL = ISI_HARD × product(penalties from fired modules) Only modules that fire (detect an anomaly) apply multiplicative penalty. Silent modules return 1.0 and do not dilute the signal. Modules:1. lexical_baseline_score — Jaccard lexical overlap (weight 0.35)2. source_target_guard — Critical-entity mutation detection (weight 0.25) — new in this work3. flow_penalty — Semantic flow rupture (weight 0.12)4. negation_penalty — Logical inversion detection (weight 0.08)5. cre_isi — Ricci semantic curvature (weight 0.12)6. arithmetic_penalty — Arithmetic error detection (weight 0.04)7. reference_penalty — Reference fabrication detection (weight 0.04) Package contents:- tribunal_multimetrico.py (core implementation)- validar_tribunal_kr.py (experimental validator)- test_distribucion_isi.py (qualitative test cases)- analisis_distribucion_isi.py (statistical analysis)- REPORTE_TRIBUNAL_MULTIMETRICO.md (full validation report)- validacion_kr_v3.json / .txt (final results, 600 pairs)- validacion_kr_v2.json / .txt (intermediate results, 200 pairs)- validacion_kr_test.json / .txt (initial test, 120 pairs)- METADATA_PACKAGE.json (structured metadata)- README_ZENODO.md (bilingual documentation) Research line:- R0 Baseline Stability Audit — DOI: 10.5281/zenodo.20647532- R0-bis Nonlinear Dependence — DOI: 10.5281/zenodo.20671824- R1 Real Local Evaluation — DOI: 10.5281/zenodo.21034155- R1-D Declarative Corpus — DOI: 10.5281/zenodo.21282332- R2.1 Code Hallucination Detection — DOI: 10.5281/zenodo.21365707- R2.1-b Multimetric Tribunal — DOI: 10.5281/zenodo.19702379 Integrity:SHA-256: 10D502E93A02C06201B6BB2C08C5C040B849D9CF0FF5544D1BF07A7272BB4E25 Links:- GitHub (active): https://github.com/Leesintheblindmonk1999/SAS- Historical archive: https://github.com/Leesintheblindmonk1999/Project_Manifold_056- Public API: https://sas-api.onrender.com- Landing page: https://leesintheblindmonk1999.github.io/sas-landing/ Citation:Durante, G. E. (2026). SAS — Multimetric Tribunal: Experimental Validation κR=0.15 / κD=0.56. Zenodo. https://doi.org/10.5281/zenodo.19702379 Keywords: hallucination detection, structural coherence, AI auditing, invariant similarity index, interpretable NLP, zero-GPU, zero-LLM Author: Gonzalo Emir Durante — Project Manifold 0.56Registry: TAD EX-2026-18792778License: Durante Invariance License v1.0Date: 2026-08-15Version: v1.0.0 Experimental preprint. This is an initial three-zone structural classification experiment using control pairs A→A and hallucination pairs A→B. Important limitations: - Classification performance is dominated by κD = 0.56.- κR = 0.15 only structures the third zone and should not be interpreted as a fully validated collapse threshold.- Zero false positives is corpus-specific and should not be generalized.- Validation with external clean negatives A→C_clean is pending.- External independent review is pending. This work documents an experimental signal, not a final validation.