CRRM: Persistent Token-Level Tension as a Semantic Signal for Representation Learning

We introduce CRRM (Conflict-Responsive Representational Modulation), a framework that tracks persistent token-level tension during training and evaluates whether this state behaves as a semantic ambiguity signal. Through a sequence of controlled experiments on a multilingual spherical embedding backend (EN/FR/DE/PL), we show that sparse activation in an initial SCM baseline is caused by an overly restrictive stable-anchor gate, not by absence of ambiguity signal. A soft poly-only proxy update repairs this gating failure, increasing poly token coverage from 7/200 to 193/200 while keeping anchor and random controls at zero. A stronger CSA-based mechanism demonstrates stable semantic targeting across checkpoints from 50k to 230k: poly coverage remains 70/108, anchor activation remains 0/50, strict concrete/technical controls remain zero, and polyfunctional, classic-polysemy, and abstract ambiguity probes activate strongly. BLI calibration results show that Run4c with CRRM at 80k training steps achieves comparable or better cross-lingual alignment than a more complex architecture (Run7 CSA v2) at 230k steps, suggesting that CRRM supports more efficient convergence without degrading embedding quality. Normal-model validation confirms that the CRRM/CSA profile is not reproduced by an SGNS baseline, and is not reduced to generic frozen-transformer contextual dispersion.