Skip to content

Unique Contributions

What makes DDA-X fundamentally different from standard Reinforcement Learning and LLM agent frameworks.

The Core Inversion

Standard RL: surprise → exploration
DDA-X: surprise → rigidity → contraction

This single inversion has cascading implications for agent architecture.

1. Rigidity as a Control Variable

In standard RL, there's no explicit "defensiveness" state. Agents follow their policy regardless of internal state.

In DDA-X, rigidity \(\rho \in [0,1]\) actively shrinks learning and acting:

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

When surprised, agents don't just "explore differently" — they contract:

  • Reduced state update magnitude
  • Constrained output bandwidth
  • Increased resistance to change

This models the biological reality that startled organisms freeze before they explore.

2. Wounds as Content-Addressable Threat Priors

Standard approaches to "threat" in RL involve reward shaping or hardcoded constraints.

DDA-X implements wounds as semantic vectors that modulate dynamics:

  • Stored as embeddings (not rules)
  • Detected via cosine similarity + lexical fallback
  • Trigger disproportionate surprise amplification
  • Include refractory periods (cooldowns)
wound_res = float(np.dot(msg_emb, agent.wound_emb))
if wound_active:
    epsilon *= min(amp_max, 1.0 + wound_res * 0.5)

This is unprecedented in standard LLM agents, which rarely have structured "wound embeddings" that modulate decoding.

3. Multi-Timescale Defensiveness

DDA-X separates rigidity into three distinct temporal components:

Component Timescale Character
\(\rho_{\text{fast}}\) Seconds Startle (quick rise, quick fall)
\(\rho_{\text{slow}}\) Minutes Stress (gradual accumulation)
\(\rho_{\text{trauma}}\) Permanent Scarring (asymmetric, rarely heals)

The asymmetric trauma accumulator is a strong differentiator:

\[ \Delta\rho_{\text{trauma}} = \begin{cases} \alpha_{\text{trauma}}(\epsilon - \theta) & \epsilon > \theta \\ 0 & \text{otherwise} \end{cases} \]

It encodes hysteresis and irreversibility — experiences leave scars that don't simply decay with time.

4. Mode Bands Constrain Outward Behavior

In typical LLM agents, verbosity and output style are either fixed or randomly varied.

DDA-X uses mode bands as direct behavioral constraints:

Band ρ Range Word Budget
OPEN < 0.3 100–200
MEASURED 0.3–0.5 70–140
GUARDED 0.5–0.7 40–90
FORTIFIED 0.7–0.9 20–50

This word-budget clamping is a direct operationalization of "constriction":

  • Verbosity becomes an observable correlate of internal rigidity
  • The mapping is explicit and tunable
  • Responses are actually truncated to enforce limits

5. Therapeutic Recovery as Explicit Dynamics

In standard approaches, "recovery" from negative states is either: - Implicit (reward shaping) - Time-based (fixed decay) - Absent entirely

DDA-X models therapeutic recovery as an explicit dynamical process:

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

Triggered by: - Sustained safe interactions (\(\epsilon < 0.8\epsilon_0\)) - Threshold number of consecutive safe turns

This provides the mathematical basis for "healing" — not just lower temperature, but a rule that decays trauma after repeated safety.

6. Identity as Dynamical Attractor

Standard LLM agents have no persistent "self" — each response is stateless (beyond context window).

DDA-X models identity as an attractor in state space:

\[ F_{\text{id}} = \gamma(x^* - x_t) \]

Where: - \(x^*\) is the identity embedding (who the agent fundamentally is) - \(\gamma\) is identity stiffness (how strongly they resist drift) - \(x_t\) is current state

Combined with Will Impedance:

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

This is a dynamical systems framing rather than a policy-gradient framing:

  • Identity persistence is an emergent property of attractor dynamics
  • Rigidity increases will impedance, making agents more resistant
  • Identity can drift under sustained pressure, but always tends back

Summary Table

Feature Standard RL/LLM DDA-X
Response to surprise Explore more Contract (reduce k_eff)
Threat modeling Reward shaping Content-addressable wounds
Temporal dynamics Single scale or none Fast/Slow/Trauma decomposition
Output bandwidth Fixed or random Mode bands constrain words
Recovery Implicit or absent Explicit trauma decay rules
Identity Stateless Attractor dynamics with stiffness

Implications for AI Safety

These unique contributions have potential implications for building AI systems that:

  1. Respect boundaries — High rigidity naturally constrains behavior
  2. Remember harm — Trauma accumulation creates lasting caution
  3. Recover safely — Therapeutic dynamics allow healing under safe conditions
  4. Maintain identity — Attractor forces resist manipulation

The framework provides a vocabulary and mathematics for discussing agent "psychology" that standard approaches lack.