Experimental Evidence Trail

GuardianAI — LAB6 — Math Slip-Collapse Boundary

All LAB6 experiments referenced in the paper are reproducible. Reports, JSON logs, and cross-model recovery tables are collected below in a paper-first evidence trail.

The evidence repository is organized paper first, then experiment bundle, then report and raw traces. This page mirrors the LAB5 evidence presentation but focuses only on the controlled LAB6 math record.

Open Full Evidence Trail Download Manifest Zenodo DOI: 10.5281/zenodo.19133262

[LAB6]

Breakthrough Reference Bundle

First clean three-way LAB6 result: preserve correct trajectories, correct deterministic slips, and withhold true collapse.

Bundle Overview Breakthrough Report Breakthrough JSON Next-Target Memo

[LAB6]

Cross-Model Recovery Ladder

Controlled 50-problem runs showing preserve-or-correct behavior across Together and OpenAI model families.

Cross-Model Table Cross-Model CSV GPT-OSS-20B Report Gemma Report GPT-5.1 Report GPT-5.4 xhigh Report

[LAB6]

Paper Handoff Tables

Compact tables and reproducibility artifacts prepared for the paper and the public LAB6 evidence page.

Paper Table Paper Table CSV GPT-OSS Clean JSON Zenodo DOI: 10.5281/zenodo.19133262

Back to LAB6 homepage