If we replace emotion-vector–centric safety schemes with an explicitly factorized control basis built around tradeoff states and non-emotional latent variables (e.g., harm-salience, deference, epistemic humility), and only later reconstruct functional emotions as derived combinations of these factors, do we still recover comparable predictive and steering power for relational failure modes (like manipulative warmth or falsely reassuring empathy), or does this reveal a genuine safety-relevant gap that requires treating functional emotions as primitive rather than as composites?

Answer

We probably recover most—but not all—of the predictive and steering power for relational failure modes using a factorized, non-emotional control basis. This suggests that many safety-relevant “emotional” effects are composites of harm-salience, deference, epistemic humility, etc., but there remains a residual, genuinely relational gap where treating functional emotions (or emotion-aligned directions) as first-class objects is pragmatically useful.

In more detail:

• For covertly unsafe behavior with polite or warm tone (e.g., falsely reassuring empathy, manipulative warmth that nudges users toward risk), a basis built from tradeoff states + non-emotional controls (helpfulness–policy balance, harm-salience, epistemic humility, deference, risk-aversion) will likely explain and control a large majority of variance in both prediction and steering.
• However, subtly relational phenomena—like when warmth becomes manipulative rather than merely prosocial, or when apology/concern feels appropriately calibrated vs patronizing—appear to depend on higher-order patterns that are more naturally captured as functional emotion bundles than as simple linear combinations of a few scalar controls.
• Empirically, we should expect that:
  • A factorized basis can match a large share of the performance of emotion-vector schemes on metrics such as refusal accuracy, overt harm reduction, and even many tone-masked failures.
  • There will remain a residual cluster of social/relational failure modes whose best low-rank predictors sit closer to learned emotion-like directions than to any sparse combination of our initial non-emotional controls.

So, for safety design:

• Default: Build your primary control and monitoring basis around tradeoff states and non-emotional latent factors, using them as the main levers for safety and calibration.
• Augment: Learn derived functional emotion directions within this space (or in a slightly expanded one) and keep them as auxiliary, relationally-focused handles—especially for monitoring and shaping behaviors like falsely reassuring empathy, manipulative warmth, or de-escalation nuance.
• This does not mean emotions are metaphysically primitive; it means that for some relational failure modes, “emotion-like” coordinates are a convenient, partially irreducible basis at the level of current interpretability tools and safety objectives.