When AI Advisors Agree: Distinguishing Forecast Risk from Forecast Reliability — Research Artifacts

Version 2 research artifacts, 14 October 2026. Associated paper: When AI Advisors Agree: Distinguishing Forecast Risk from Forecast Reliability (arXiv:2610.13431, cs.MA). This deposit provides the revised PDF; arXiv replacements appear after their processing and announcement cycle. This is a preprint, not journal acceptance or human peer certification.

The frozen outcome cutoff remains 8 October 2026. The primary cohort remains 928 completed quarterly outcomes for 232 assets across four historical iPulse AI batches. No forecasts were regenerated.

V2 distinguishes the sign of the council mean from modal framework voting: the positive directional-agreement contrast is specific to mean-sign accuracy and disappears under modal voting. It adds recomputed-rank bootstraps, sector/class dependence sensitivity, period and leave-one-batch-out comparisons, scale-aware losses, framework and mode dependence, threshold grids, confusion matrices, regression diagnostics, and cohort/anchor audits. The volatility selector advantage nearly disappears after omitting Batch 3. Four overlapping periods cannot establish future-regime performance.

The public companion contains 12,270 original generation-anchor configuration forecasts, cohort dates and membership for 1,227 matched records, frozen sector/class grouping, aggregate results, the 24-page manuscript, eleven figures, twelve tables, LaTeX source, methodology, executed-analysis inventory, MIT analysis code, and a tested market-summary reconstruction adapter.

Raw market-history rows and the full market-derived outcome panel remain private. Outcome-dependent replication requires compatible authorized market inputs, including the recorded forecast-anchor basis. Author-side adapter validation is not independent external market-data replication. Two Batch 3 commodity records have different forecast and historical outcome anchor bases; additional compatible-anchor sensitivities are reported. No probabilistic calibration, practical equivalence, causal effect, trading profitability or validated live confidence rule is claimed.

Original paper, forecasts, figures and research summaries: CC BY 4.0; code: MIT; third-party material retains its own terms. Author: Russlan Ramdowar, Future Edge Group FZE / iPulse AI; ORCID 0009-0004-9311-6957; russlan@ipulseai.com. Commercial-interest and AI-assistance disclosures are retained in the manuscript. Repository: TheFutureEdge/research-when-ai-advisors-agree. V1 remains archived; the all-versions DOI is 10.5281/zenodo.23262215.