Reliable next-session risk monitoring requires models that remain interpretable under limited and mixed-frequency data. This study evaluates a lightweight macro-financial fusion framework for forecasting the direction of the Cboe Volatility Index (VIX) and a transparent credit-risk-pressure index (CRPI) built from eight FRED series. Daily market and Treasury observations are aligned with carried-forward monthly policy, labor, and price indicators; trailing features feed persistence, logistic regression, random forest, gradient boosting, gated recurrent unit, and temporal-transformer models under chronological train-validation-test splits. The VIX task uses 258 observed 2025 rows and yields 162 training, 45 validation, and 41 test sequences after a 10-session window. The longer CRPI task yields 1,614, 502, and 375 sequences after a 21-session window. Random forest provides the strongest VIX class balance (accuracy = 0.634, balanced accuracy = 0.626, F1 = 0.571, ROC-AUC = 0.657), while gradient boosting leads the CRPI task (accuracy = 0.584, balanced accuracy = 0.576, F1 = 0.519, ROC-AUC = 0.604). The compact transformer does not outperform the tree ensembles. A deterministic evidence-grounded memo layer converts probabilities and recorded drivers into eight test-period summaries; complete driver coverage and zero detected numeric violations are consistency checks by construction, not evidence of analyst comprehension or autonomous language-model reasoning. Results support cautious use as a reproducible monitoring and audit-documentation framework. Limitations include the one-year VIX span, the rule-defined CRPI proxy, release-vintage uncertainty, and the absence of human evaluation.

Keywords: VIX, credit-risk pressure, macro-financial forecasting, temporal transformers, evidence-grounded risk communication.