DDA-X: Surprise → Rigidity → Contraction

A Dynamical Framework for Embedding-Space Agents with Rigidity-Bound Generation

Abstract

DDA-X is a cognitive dynamics framework in which prediction error (surprise) induces defensive rigidity rather than immediate exploration. In contrast to standard reinforcement learning heuristics that treat surprise as an exploration bonus, DDA-X models startled systems as contracting—reducing behavioral bandwidth and state update magnitude—until safety and predictability permit reopening. We operationalize DDA-X in a family of simulations spanning adversarial dialogue, therapeutic recovery, multi-agent convergence, and embodied collision dynamics. The framework combines (i) a continuous state space with identity attractors, (ii) rigidity-modulated effective step size, (iii) multi-timescale rigidity decomposition (fast/slow/trauma), and (iv) content-addressable wound activation.

1. State Representation

Let \(x_t \in \mathbb{R}^d\) be an agent's internal state at time \(t\), where \(d=3072\) (using text-embedding-3-large).

Each agent maintains: * \(x_t\): Current state (initialized near an identity attractor \(x^*\)) * \(x^{\text{pred}}_t\): Predicted embedding of the agent's next action/utterance * \(e(a_t)\): Embedding of the generated utterance \(a_t\)

In multiple simulations, \(x^{\text{pred}}\) is an exponentially smoothed forecast of the agent's own past output embeddings:

\[ x^{\text{pred}}_{t+1} = (1-\beta)x^{\text{pred}}_{t} + \beta \, e(a_t) \]

(e.g., \(\beta=0.3\) in simulate_the_returning.py, simulate_skeptics_gauntlet.py).

2. Surprise as Prediction Error

We define surprise as the Euclidean distance between prediction and reality in semantic space:

\[ \epsilon_t = \lVert x^{\text{pred}}_t - e(a_t) \rVert_2 \]

This acts as the primary driving signal for the cognitive engine.

3. Rigidity Dynamics: The Logistic Gate

Rigidity \(\rho_t \in [0,1]\) modulates the system's openness to change. We compute an update \(\Delta \rho_t\) using a logistic gating function centered at a surprise threshold \(\epsilon_0\):

\[ z_t = \frac{\epsilon_t - \epsilon_0}{s}, \qquad \sigma(z) = \frac{1}{1+e^{-z}} \]

\[ \Delta \rho_t = \alpha(\sigma(z_t) - 0.5) \]

\[ \rho_{t+1} = \text{clip}(\rho_t + \Delta\rho_t, \, 0, \, 1) \]

This implements the core DDA-X thesis: higher surprise \(\rightarrow\) higher rigidity.

Multi-Timescale Decomposition

To model complex behaviors like startle vs. trauma, we decompose rigidity into three timescales:

1. Fast (\(\rho_{\text{fast}}\)): Quick startle response, rapid decay.
2. Slow (\(\rho_{\text{slow}}\)): Quantitative stress accumulation.
3. Trauma (\(\rho_{\text{trauma}}\)): Asymmetric accumulation (scarring). Only increases when \(\epsilon > \theta_{\text{trauma}}\) (unless therapeutic intervention occurs).

Effective rigidity is a weighted sum:

\[ \rho_{\text{eff}} = \min(1, \, w_f \rho_{\text{fast}} + w_s \rho_{\text{slow}} + w_t \rho_{\text{trauma}}) \]

4. State Update and Will Impedance

The agent's state evolves under forces, but the magnitude of evolution is throttled by rigidity.

\[ x_{t+1} = x_t + k_{\text{eff}} \cdot \eta \Big( \underbrace{\gamma(x^* - x_t)}_{\text{Identity}} + m(\underbrace{F_{\text{truth}}}_{\text{Observation}} + \underbrace{F_{\text{reflection}}}_{\text{Action}}) \Big) \]

Where the Effective Step Size is:

\[ k_{\text{eff}} = k_{\text{base}}(1 - \rho_t) \]

We define Will Impedance \(W_t\), quantifying resistance to environmental pressure:

\[ W_t = \frac{\gamma}{m_t \cdot k_{\text{eff}}} \]

As rigidity \(\rho \to 1\), \(k_{\text{eff}} \to 0\), and Will Impedance \(W_t \to \infty\). The agent becomes immovable.

5. Binding Rigidity to LLM Generation

DDA-X binds the internal scalar \(\rho\) to the external LLM generation process via OpenAIProvider.

For Reasoning Models (o1 / GPT-5.2)

Since current reasoning models do not support granular sampling parameters (temperature), we inject Semantic Rigidity Instructions directly into the cognitive state block of the prompt:

[COGNITIVE STATE]: You are in a GUARDED state. Your beliefs are rigid. Limit your response to 50 words. Do not accept new premises without extreme evidence.

For Standard Models (GPT-4o)

We modulate sampling parameters: * Temperature: \(T(\rho) = T_{\min} + (1-\rho)(T_{\max} - T_{\min})\) * Top-P: Constricted as rigidity increases. * Presence Penalty: Reduced to discourage topic drift.

6. Wounds: Content-Addressable Threat Priors

A "Wound" is a semantic point of vulnerability \(w \in \mathbb{R}^d\). Activation occurs via:

\[ \text{wound\_active} = (\langle e(s_t), w \rangle > \tau_{\cos} \;\lor\; \text{lexical\_match}) \land \text{cooldown\_satisfied} \]

When active, surprise is amplified, simulating a disproportionate psychological reaction:

\[ \epsilon'_t = \epsilon_t \cdot \min(A_{\max}, \, 1 + c \cdot \langle e(s_t), w \rangle) \]

7. Trust: Hybrid Implementation

While the theoretical ideal for trust is prediction-based (\(T_{ij} = \frac{1}{1 + \sum \epsilon}\)), current simulations implement a Hybrid Trust Model:

• Semantic Alignment: Trust increases when \(e(a_t)\) aligns with the partner's previous output (Dialectic).
• Civility Gating: Trust decreases on "unfair" (lexically wounded) engagement.
• Coalition Bias: Trust is initialized based on group topology (Council).

8. Therapeutic Recovery

In specific simulations (e.g., simulate_healing_field.py), we model trauma decay. If the agent experiences a sustained "Safe Streak" where \(\epsilon < 0.8\epsilon_0\):

\[ \rho_{\text{trauma}} \leftarrow \max(\rho_{\min}, \, \rho_{\text{trauma}} - \eta_{\text{heal}}) \]

This provides the mathematical basis for "healing" within the dynamical system.

---

9. Unique Contributions vs. Standard RL

1. Inverted exploration: Surprise \(\to\) Contraction.
2. Multi-timescale defensiveness: State separability of startle vs. trauma.
3. Wounds as state: Semantic priors that modulate dynamics.
4. Identity as attractor: Use of \(\gamma(x^* - x)\) ensures identity persistence.

---

10. Relation to RL and LLM Agents

DDA-X differs from typical RL/LLM-agent designs in several fundamental ways:

Aspect	Standard RL/LLM	DDA-X
Response to surprise	Explore more (curiosity bonus)	Contract (reduce \(k_{\text{eff}}\))
Threat modeling	Reward shaping or constraints	Content-addressable wound embeddings
Temporal dynamics	Single timescale or none	Fast/Slow/Trauma decomposition
Output control	Fixed or random verbosity	Mode bands constrain word counts
Recovery from harm	Implicit or absent	Explicit trauma decay dynamics
Identity	Stateless (context window only)	Attractor dynamics with stiffness \(\gamma\)

The key insight is that DDA-X treats surprise as a contraction signal rather than an exploration signal, and makes defensiveness, wounds, and trauma explicit state variables that directly modulate update magnitudes and policy bandwidth.

---

11. Limitations and Open Problems

Transparency

This section documents known gaps between theory and implementation. See Known Limitations for full details.

11.1 Trust Equation Mismatch

The theoretical trust equation \(T_{ij} = \frac{1}{1+\sum\epsilon}\) is not implemented in current simulations. The hybrid trust model (Section 7) is the actual implementation.

11.2 Calibration Asymmetry

While \(\epsilon_0\) and \(s\) are calibrated from runtime statistics, wound thresholds (\(\tau_{\cos}\)), trauma thresholds (\(\theta_{\text{trauma}}\)), and multi-timescale weights (\(w_f, w_s, w_t\)) remain hardcoded. These may require domain-specific tuning.

11.3 Measurement Validity

The prediction error \(\epsilon_t = \|x_{\text{pred}} - e(a_t)\|\) conflates semantic novelty with style/verbosity shifts. Future work could decompose embeddings into content vs. tone components.

11.4 Model Dependency

For reasoning models (GPT-5.2, o1), rigidity cannot modulate sampling parameters—semantic injection is the only option. Effectiveness may vary across models.

11.5 Open Research Questions

1. Can multi-timescale weights be learned from data?
2. How do thresholds transfer across domains?
3. What safe interaction patterns most effectively heal trauma?

---

Appendix A: Parameter Schema (D1_PARAMS)

Most simulations use a consistent "physics dictionary" pattern for configuration:

Parameter	Symbol	Typical Value	Description
epsilon_0	\(\epsilon_0\)	Calibrated	Surprise baseline (median of early epsilons)
s	\(s\)	Calibrated	Sigmoid steepness (IQR-based)
alpha	\(\alpha\)	0.05–0.15	Base learning rate for \(\Delta\rho\)
alpha_fast	\(\alpha_f\)	0.20	Fast rigidity learning rate
alpha_slow	\(\alpha_s\)	0.02	Slow rigidity learning rate
alpha_trauma	\(\alpha_t\)	0.10	Trauma accumulation rate
trauma_threshold	\(\theta_t\)	0.9–1.0	Epsilon required for trauma increase
wound_cosine_threshold	\(\tau_{\cos}\)	0.28	Semantic resonance trigger
wound_cooldown	—	3–5 turns	Wound refractory period
drift_cap	—	0.05–0.10	Maximum per-turn state movement
k_base	\(k\)	0.1–0.3	Base step size
gamma	\(\gamma\)	0.1–0.5	Identity stiffness

Note: Calibration of epsilon_0 and s occurs after a warm-up period using observed prediction error statistics (median + IQR).

---

Citation

@misc{ddax2025,
  author = {snakewizardd},
  title = {DDA-X: Surprise → Rigidity → Contraction},
  year = {2025},
  publisher = {GitHub},
  url = {https://github.com/snakewizardd/dda_scaffold}
}