ReCoN-Ipsundrum Agent Architecture

• ReCoN-Ipsundrum is an inspectable agent architecture that integrates recurrent sensory loops and affect-coupled control to generate measurable indicators of artificial consciousness.
• It employs a script-node state-machine with recurrent persistence over sensory salience, enabling direct mechanistic inspection and causal intervention via lesion studies.
• Behavioral assays demonstrate that coupling affect not only stabilizes exploration preferences but also induces prolonged planned caution, distinguishing it from traditional state-machine agents.

ReCoN-Ipsundrum is an inspectable agent architecture designed to provide mechanism-linked, indicator-based evidence for artificial consciousness through explicit architectural features, behavioral dissociations, and causal intervention. Rooted in the ipsundrum hypothesis proposed by Humphrey, the ReCoN-Ipsundrum framework extends a ReCoN state-machine with a recurrent persistence loop over sensory salience and optionally incorporates affect-coupled (valence/arousal) control. The agent is engineered to facilitate the triangulation of indicator-like signatures—postulated as markers of conscious-like processing—in a controlled computational setting, enabling direct architectural inspection and experimental ablation (Sanyal, 26 Feb 2026).

1. State-Machine Architecture and Recurrence

ReCoN-Ipsundrum builds upon a script-node state-machine substrate. Nodes are partitioned into:

• Script nodes (S): Hold discrete states (inactive, requested, active, confirmed, failed).
• Sensor/actuator nodes (Ns, Nm): Emit continuous activations in $[0,1]$, with activations thresholded for confirmations.

Core message types—request, confirm, wait, inhibit—mediate top-down and bottom-up activation cycles. The transition function is hierarchically propagative: script nodes issue requests to children and await confirmation, progressing upon satisfaction of subgoals. The policy interface enumerates candidate actions in each environment step, leverages an internal forward model for one-step prediction of sensory drives, and scores actions via a parameterized scoring function that can incorporate affective signals depending on the agent variant.

Crucially, the Ipsundrum extension introduces a recurrent loop over the sensory salience channel $N^s$. This loop is realized as a “thick-moment” integrator:

$$X_t = d\,X_{t-1} + (1-d)\,N^s_t;\;\; M_t = \mathrm{clip}_{[0,1]}(h\,X_t)$$

$$N^e_t = d_e\,N^e_{t-1} + (1-d_e)\,M_t;\;\; E_t = \mathrm{clip}_{[0,1]}\bigl(g_\mathrm{eff}\,M_t\bigr)$$

where $d$, $d_e$ are decay parameters, $h$ a gain, and $g_\mathrm{eff}$ the feedback gain. This recurrence creates multi-step persistence of sensory evidence. Lesioning (e.g., setting $\pi_t=0, E_{t-1}=0$ or $d=0$) isolates the mechanistic contribution of recurrence to post-stimulus persistence.

2. Affect-Coupled Control and Internal Affect Proxy

The affect variant incorporates a minimal homeostatic model:

• Body budget $N^s$0 tracks cumulative internal state via

$N^s$1

with $N^s$2 (update rate), reflecting the effect of aversive ($N^s$3) or beneficial ($N^s$4) sensory evidence.

• Interoceptive proxy $N^s$5
• Valence $N^s$6 (proximity to setpoint $N^s$7)
• Arousal $N^s$8, $N^s$9

In affect-coupled control, $$X_t = d\,X_{t-1} + (1-d)\,N^s_t;\;\; M_t = \mathrm{clip}_{[0,1]}(h\,X_t)$$0 are integrated into the policy decision function:

$$X_t = d\,X_{t-1} + (1-d)\,N^s_t;\;\; M_t = \mathrm{clip}_{[0,1]}(h\,X_t)$$1

with fixed weights for affective influence. Affect can also modulate precision $$X_t = d\,X_{t-1} + (1-d)\,N^s_t;\;\; M_t = \mathrm{clip}_{[0,1]}(h\,X_t)$$2 or feedback gain $$X_t = d\,X_{t-1} + (1-d)\,N^s_t;\;\; M_t = \mathrm{clip}_{[0,1]}(h\,X_t)$$3 in the recurrence loop. This configuration enables the dissociation of persistence (recurrence) from preference stability and affective modulation of action selection.

3. Behavioral Assays: Qualiaphilia and Exploratory Play

A core behavioral probe operationalizes “qualiaphilia” (preference for sensory experience) by presenting the agent with a familiarity-controlled choice between scenic (variable, potentially beneficial) and dull (uniform) routes. During exposure, novelty scores are accrued as cross-episode visitation counts, which then inform a curiosity bonus in the scoring function.

Metrics include:

• $$X_t = d\,X_{t-1} + (1-d)\,N^s_t;\;\; M_t = \mathrm{clip}_{[0,1]}(h\,X_t)$$4: Probability of entering scenic over dull lane, conditioned on which is more novel.
• $$X_t = d\,X_{t-1} + (1-d)\,N^s_t;\;\; M_t = \mathrm{clip}_{[0,1]}(h\,X_t)$$5: Novelty difference between lanes.

Results:

• Non-affect variants exhibit novelty-sensitivity ($$X_t = d\,X_{t-1} + (1-d)\,N^s_t;\;\; M_t = \mathrm{clip}_{[0,1]}(h\,X_t)$$6), biasing exploration toward the less familiar option.
• Affect-coupled variant stabilizes preference ($$X_t = d\,X_{t-1} + (1-d)\,N^s_t;\;\; M_t = \mathrm{clip}_{[0,1]}(h\,X_t)$$7) even when scenic is less novel ($$X_t = d\,X_{t-1} + (1-d)\,N^s_t;\;\; M_t = \mathrm{clip}_{[0,1]}(h\,X_t)$$8), dissociating from mere novelty seeking.

In reward-free exploratory play (neutral gridworld, 200 steps):

• Affect variant displays pronounced local investigation (scan events $$X_t = d\,X_{t-1} + (1-d)\,N^s_t;\;\; M_t = \mathrm{clip}_{[0,1]}(h\,X_t)$$9 vs. $$N^e_t = d_e\,N^e_{t-1} + (1-d_e)\,M_t;\;\; E_t = \mathrm{clip}_{[0,1]}\bigl(g_\mathrm{eff}\,M_t\bigr)$$0 non-affect; cycle score $$N^e_t = d_e\,N^e_{t-1} + (1-d_e)\,M_t;\;\; E_t = \mathrm{clip}_{[0,1]}\bigl(g_\mathrm{eff}\,M_t\bigr)$$1).
• Action entropy is lower than random, reflecting structured rather than undirected exploration.
• Dwell-time tail (P$$N^e_t = d_e\,N^e_{t-1} + (1-d_e)\,M_t;\;\; E_t = \mathrm{clip}_{[0,1]}\bigl(g_\mathrm{eff}\,M_t\bigr)$$2) is reduced, mirroring more distributed occupancy.

4. Indicator Assays: Pain-Tail Probe and Lesion Analysis

The pain-tail probe exposes the agent to an aversive event, after which the agent is reset to a safe state in a static environment, and planned actions are recorded for 200 steps. Indicator readouts:

• Mechanistic: Area under curve (AUC) above baseline for $$N^e_t = d_e\,N^e_{t-1} + (1-d_e)\,M_t;\;\; E_t = \mathrm{clip}_{[0,1]}\bigl(g_\mathrm{eff}\,M_t\bigr)$$3 post-stimulus.
• Behavioral: Prolonged planned caution (tail duration, defined as first window where turn rate $$N^e_t = d_e\,N^e_{t-1} + (1-d_e)\,M_t;\;\; E_t = \mathrm{clip}_{[0,1]}\bigl(g_\mathrm{eff}\,M_t\bigr)$$4).

Findings:

• Only the affect variant sustains prolonged planned caution ($$N^e_t = d_e\,N^e_{t-1} + (1-d_e)\,M_t;\;\; E_t = \mathrm{clip}_{[0,1]}\bigl(g_\mathrm{eff}\,M_t\bigr)$$5 steps vs. $$N^e_t = d_e\,N^e_{t-1} + (1-d_e)\,M_t;\;\; E_t = \mathrm{clip}_{[0,1]}\bigl(g_\mathrm{eff}\,M_t\bigr)$$6 in other variants); post-stimulus AUC is significant for Ipsundrum ($$N^e_t = d_e\,N^e_{t-1} + (1-d_e)\,M_t;\;\; E_t = \mathrm{clip}_{[0,1]}\bigl(g_\mathrm{eff}\,M_t\bigr)$$7) and Ipsundrum+affect ($$N^e_t = d_e\,N^e_{t-1} + (1-d_e)\,M_t;\;\; E_t = \mathrm{clip}_{[0,1]}\bigl(g_\mathrm{eff}\,M_t\bigr)$$8), zero for ReCoN.
• Causal lesions of feedback/integration cause significant drops in post-stimulus persistence for ipsundrum variants (AUC drop $$N^e_t = d_e\,N^e_{t-1} + (1-d_e)\,M_t;\;\; E_t = \mathrm{clip}_{[0,1]}\bigl(g_\mathrm{eff}\,M_t\bigr)$$9, $d$0 for Ipsundrum+affect), but have no effect on ReCoN.

These behavioral and mechanistic dissociations demonstrate that recurrence specifically enables post-stimulus persistence, while affective coupling is necessary for preference stability, structured scanning, and lingering caution.

5. Architectural Transparency, Mechanistic Attribution, and Synthesis

ReCoN-Ipsundrum provides a concrete substrate for machine consciousness research, emphasizing three converging lines of evidence:

• Architectural inspection: Recurrence and affect circuits are explicit and inspectable.
• Behavioral markers: Engineered signatures such as persistent caution, stabilized preference, and intensive exploratory scanning.
• Causal intervention: Lesion studies directly link mechanistic substrates to behavioral indicators.

This approach illustrates the feasibility of mapping computational mechanisms to consciousness indicators in artificial agents—drawing a distinction between engineered markers and naturalistic attributions. The findings substantiate that postulated behavioral signatures for consciousness can be induced by minimal recurrent and affective loops. A broader implication is that attribution of consciousness-like properties to agents necessitates careful consideration of underlying architectural mechanisms and rigorous causal testing across multiple behavioral paradigms (Sanyal, 26 Feb 2026).