Seven research papers documenting the Augle system — the seven-agent ensemble, corpus pipeline, confidence scoring, reopen conditions, and evidence admission logic. Published on Zenodo and SSRN to establish prior art and contribute to the field of augmented deliberation. All papers are co-authored by Cory Kelly and Shubhanker Saxena.
Each paper covers a distinct architectural component of the Augle system. Together they constitute a complete specification of the deliberation engine — from agent design to corpus infrastructure to calibration methodology.
Cory Kelly, Shubhanker Saxena
Introduces the Augle seven-agent ensemble — Guardian, Topic Architect, Cartographer, Methodologist, Contrarian, Synthesizer, and Pragmatist — and describes the three-phase deliberation architecture (Exploration, Deliberation, Synthesis). Defines confidence propagation, the Methodologist ceiling constraint, Grade Challenge mechanics, and the evidence nodes registry. The foundational architecture paper.
Cory Kelly, Shubhanker Saxena
Describes the document ingestion pipeline, the Structured Verification System (SVS) for source validation, and the Guardian's integrity scoring across six modes: Academic, Legal, Clinical, Financial, Editorial, and Markets. Defines the Critical/Moderate/Low flag taxonomy, phase boundary clearance mechanics, and the evidence admission decision tree that governs what enters deliberation.
Cory Kelly, Shubhanker Saxena
Introduces Structured Verbal Sparring (SVS) — the formal protocol governing Contrarian agent behaviour. Defines the steelman-first requirement, objection strength grades (Strong, Moderate, Speculative), resolution conditions, and the unresolved objection registry. Describes how dissent is preserved verbatim in session outputs and how the Contrarian's dissent register integrates with the Synthesizer's final finding.
Cory Kelly, Shubhanker Saxena
Describes the corpus infrastructure that maps Augle deliberation sessions to prediction market ground truth outcomes. Defines corpus quality tiers (Gold, Silver, Flagged), the Brier scoring methodology, and the calibration comparison framework against market consensus. Establishes the V3 Calibrator training protocol and the corpus accumulation threshold required (~5,000 resolved sessions) before supervised calibration training begins.
Cory Kelly, Shubhanker Saxena
Describes the Real-Time Evidence Admission Protocol (REAP) — the mechanism by which new evidence submitted mid-session is evaluated, verified, and integrated without disrupting the deliberation in progress. Defines the evidence queue architecture, Guardian pre-screening for mid-session submissions, the confidence recalculation protocol, and the precedence rules governing phase boundary interactions when new evidence arrives near a phase transition.
Cory Kelly, Shubhanker Saxena
Introduces the reopen condition framework — the structured specification of triggers that cause a resolved session's finding to be subject to mandatory revision. Defines reopen condition types (market event triggers, new evidence triggers, retraction triggers), the directionality requirement (upgrade or downgrade), and the session lineage graph that links parent sessions to their reopen descendants. Distinguishes reopen conditions in Markets-mode sessions from Academia evidence-triggered conditions.
Cory Kelly, Shubhanker Saxena
Describes the follow-on session architecture — the mechanism by which knowledge gaps identified by the Cartographer, unresolved objections from the Contrarian, and triggered reopen conditions automatically generate proposed follow-on research questions. Defines the three-source generation system, priority tier assignment, lineage graph construction, and the user-facing proposal interface. Introduces the compounding research loop as a mechanism for iterative calibration improvement across session chains.
All papers are published under Creative Commons Attribution 4.0 (CC BY 4.0). You are free to share and adapt with attribution. The Zenodo repository provides persistent DOIs for each paper.
BibTeX and RIS formats available on each paper's Zenodo record page.
All seven papers have been submitted to arXiv for broader visibility and prestige within the AI research community. Submission is pending endorsement — arXiv requires endorsement for new submitters in the cs.AI and cs.MA categories. Zenodo timestamps establish prior art independently of arXiv status.
If you're a researcher working on augmented deliberation, multi-agent reasoning, calibration scoring, or AI epistemic systems — we'd like to hear from you. All papers are open access.
Get in touch →