Semantic State Abstraction Interfaces for LLM-Augmented Portfolio Decisions: Multi-Axis News Decomposition and RL Diagnostics

We introduce Semantic State Abstraction Interfaces (SSAI): a methodological1 template for mapping sparse unstructured text into K auditable, named coordinates2 with neutral defaults on no-news days, designed to separate representation hypothe-3 ses from optimisation variance in sequential decision systems. Our contribution4 is the framework and its evaluation protocol—not a claim that SSAI outperforms5 denser alternatives.6 We instantiate SSAI with K=4 axes (sentiment, risk, confidence, volatility fore-7 cast) on a US-equity panel (30 NASDAQ-100 names, FNSPID news, 2019–20238 test), and evaluate it across three parallel estimators—direct factor portfolios9 (SFP/SRF/SCW), supervised ridge forecasters, and RL agents (DP-PPO, SAC)—10 that share the same fixed φ so that signal and optimiser effects can be read off11 separately.12 What the experiments show. The four-factor SFP reaches 307.2% CR / Sharpe13 1.067 over 2019–2023. However, this apparent advantage over buy-and-hold14 (243.6%) does not survive its own controls: SFP underperforms stratum-matched15 B&H in every coverage tercile (Table 3), reverses at ≥0.2% per-trade costs, and16 the four-factor vs. sentiment-only daily advantage is non-significant (Wilcoxon17 p=0.556). A data-driven first principal component of the four axes (PC1-SFP)18 achieves 433.6% CR under the same portfolio rule—126pp above SSAI—and a19 FinBERT direct-portfolio baseline achieves 386.3% (Table 5). The honest summary20 is that the SSAI portfolio advantage is substantially a basket-selection composition21 effect, and both PC1 and dense neural encodings are empirically stronger ranking22 signals in this setting.23 What the experiments also show. Naively appending sparse LLM columns hurts24 ridge forecasting; high-conviction semantic tilts and lexical baselines (VADER, TF–25 IDF/SVD) recover price-only Sharpe. The RL block is a diagnostic: DP-PPO trails26 buy-and-hold; SAC with identical SSAI observations improves Sharpe (1.128 vs.27 1.032), isolating algorithm dependence rather than representation advantage. Seed-28 mean Sharpe differences across 21 DP-PPO seeds are non-significant (Wilcoxon29 p≈0.25–0.31).30 Contribution framing. We present SSAI as an interpretability-performance fron-31 tier instrument and a cautionary diagnostic: it characterises the cost of maintaining32 auditable axes (126pp CR vs. PC1 over five years), shows that RL performance is33 dominated by algorithm choice rather than representation, and provides a reusable34 template for separating semantic signal from optimisation noise in other sparse-text35 decision settings