ReCoN-Ipsundrum: An Inspectable Recurrent Persistence Loop Agent with Affect-Coupled Control and Mechanism-Linked Consciousness Indicator Assays

AI summary

Overview

Research area: Machine consciousness assessment / AI safety methodology, combining neurosymbolic agent architectures (Request Confirmation Networks), recurrent dynamical models inspired by Nicholas Humphrey's "ipsundrum" hypothesis of sentience, and constructionist affect theory.

Technical level: Advanced. The paper assumes familiarity with recurrent processing debates, indicator-based approaches to machine consciousness (Butlin et al. 2025), and a state-machine/predictive-processing vocabulary.

Scope: The paper builds a small, fully inspectable agent that adds a recurrent persistence loop and an optional affect proxy to a ReCoN sensorimotor substrate, then runs fixed-parameter ablations, behavioral assays, and causal lesions to test which indicator-like signatures attach to which internal mechanism.

What This Paper Is About

Researchers who study whether AI systems could be conscious need evidence that goes beyond behavior, because convincing-looking behavior can be produced by trivial or "gameable" mechanisms. This paper takes one specific theory of how sentience might arise — Humphrey's claim that reflexive control becomes self-monitoring and then self-sustaining, forming a recurrent loop he calls an ipsundrum — and turns it into an actual, inspectable software agent with three ablatable variants. The goal is not to claim consciousness but to show how theory-derived indicators can be engineered, dissociated, and causally attributed to specific internal components.

Key Contributions

1. ReCoN-Ipsundrum architecture: An extension of a Request Confirmation Network (ReCoN) state machine with an explicit recurrent "ipsundrum" persistence loop over sensory salience (N^{s}), plus an optional affect/interoception proxy reporting valence and arousal, built through four staged design steps (Stage A centrally coordinated reflex through Stage D gated percept looping).
2. Assays with controls: Operationalizations of Humphrey's qualiaphilia (a familiarity-controlled scenic-versus-dull route choice that separates hedonic preference from novelty seeking), reward-free exploratory play with scan events, a pain-tail probe for lingering planned caution, and goal-directed navigation as a competence/safety check.
3. Mechanistic evaluation with causal lesions: Within-episode lesions at (t{=}3) that disable ipsundrum feedback and integration, used to attribute post-stimulus persistence in (N^{s}) to the implemented recurrence rather than to incidental artifacts.
4. An empirical dissociation table: A demonstration that recurrence alone produces persistence, while affect-coupled control is required for value-stable scenic preference, structured scanning, and prolonged planned caution — three signatures that do not follow from persistence on its own.

Main Findings

• Novelty dissociation in corridor preference: The non-affect variants (ReCoN and Ipsundrum) are novelty-sensitive, with (\Delta\text{scenic-entry} = 0.07) [-0.02, 0.16]. Ipsundrum+affect remains stable at (\Delta\text{scenic-entry} = 0.01) [0.00, 0.03] even when the scenic lane is less novel (median (\Delta\text{novelty} \approx -0.43)).
• The stability is value-shaped, not value-neutral: A split-point probe predicts higher valence (0.91 vs. 0.88) and lower arousal (0.24 vs. 0.32) for the scenic turn in the affect variant, because scenic versus dull changes the signed sensory drive (I_t).
• Structured investigation in reward-free play: Over a 200-step neutral-texture gridworld, the affect variant shows more scan events (31.4 [25.8, 38.2]; the abstract reports 31.4 vs. 0.9 for the comparison) and stronger limit-cycle structure (7.6 [1.2, 16.7]). Unique viewpoints are broadly similar across variants ((\approx)136–140).
• Scanning is not dithering: Action entropy stays below a random baseline (1.29 vs. 1.99) and dwell tails remain short (2.7 vs. 8.97). Recurrence alone does not increase scanning (Ipsundrum (\approx) ReCoN); adding affect coupling does.
• Post-stimulus persistence is real but insufficient for caution: In the pain-tail probe, ipsundrum variants show non-zero post-stimulus (N^{s}) AUC above baseline (Ipsundrum (\approx) 0.24; Ipsundrum+affect (\approx) 0.15), while ReCoN is (\approx) 0. Only Ipsundrum+affect shows prolonged planned caution (turn-rate tail duration (\approx) 90 [52, 128] vs. 5 and 5).
• Lesions attribute persistence to the implemented mechanism: Lesioning feedback+integration at (t{=}3) leaves ReCoN essentially unchanged (AUC drop (\approx) 0.00) but reduces persistence in ipsundrum variants ((\approx) 19.12, or 20.3%, for Ipsundrum; (\approx) 27.62, or 27.9%, for Ipsundrum+affect) over a 150-step window.
• Navigation safety check: Aggregated across the horizon sweep (H \in {1,\dots,20}), Ipsundrum+affect reduces hazard contacts in both CorridorWorld (0.00 [0.00, 0.00] vs. 3.05 [2.00, 4.25] for ReCoN) and GridWorld (0.26 [0.10, 0.44] vs. 14.30 [8.90, 20.05]). In GridWorld this also improves success (0.99 [0.97, 1.00] vs. 0.50 [0.30, 0.70]) and steps-to-goal (9.54 [4.75, 16.23] vs. 171.80 [131.30, 211.10]); in CorridorWorld it remains hazard-free but slower.
• No consciousness claim: The authors state explicitly that none of these signatures are treated as evidence of consciousness, and that persistence, stable preference, scanning, and caution are engineered, dissociable properties.

Methodology in Plain English

The researchers start from ReCoN, a graph-based sensorimotor control system in which "script" nodes request confirmation from lower-level routines and sensors. They then add features in stages that mirror Humphrey's evolutionary narrative. Stage B adds an efference-copy signal (N^{e}) — a filtered record of the agent's own outgoing motor command, i.e., a monitorable "what I am doing" signal. Stage C and D attach a recurrent state update where the previous reafferent signal is fed back into the current sensory drive, producing a "thick moment" that outlasts the stimulus; a gate keeps the loop running while the efference copy stays above threshold.

An optional affect layer adds a bookkeeping "body-budget" model: positive sensory drive (noxious evidence) depletes the budget, negative drive (scenic evidence) replenishes it. Two readouts are exposed — valence as closeness to a setpoint and arousal as prediction-error magnitude — and these can modulate the loop's precision or gain.

All three variants are fixed-parameter and learning-free, and they share the same policy and environment interface. Action selection enumerates candidate actions, simulates one-step sensory consequences, and scores them with a linear combination of valence, arousal, salience, budget error, an epistemic term, a novelty bonus, goal progress, a hazard penalty, and action costs. Crucially, the forward model used in planning matches the variant being evaluated, so ablations remain faithful.

The assay suite then probes different predictions: a two-route corridor where familiarity is manipulated across episodes to separate novelty seeking from stable preference; a reward-free gridworld for exploratory play; a pain-tail protocol where one hazard contact is forced, the hazard is removed, and planned actions are recorded for 200 steps; and goal-directed navigation as a safety/competence check. Finally, in-episode lesions zero out the feedback term and the integration decay to test whether observed persistence actually depends on the recurrence.

Why This Matters

The paper's central methodological claim is about attribution error in machine-consciousness assessment: indicator-like behavior can be produced by minimal, potentially gameable implementations, so any single behavioral marker risks false positives and false negatives. The authors argue that indicators should count as credence-shifting only conditional on explicit mechanistic hypotheses, supported by architectural inspection and causal intervention.

Potential applications:

• AI consciousness and moral-status assessment: Provides a template for pairing behavioral markers with lesion-style causal evidence before drawing any ethical conclusions about a system.
• AI safety and evaluation design: Demonstrates how to build ablatable agents where internal components can be removed within an episode to attribute behavior to a mechanism.
• Cognitive-architecture prototyping: Shows a concrete way to bolt recurrent persistence and interoceptive-style affect onto an existing symbolic/sensorimotor control substrate.
• Benchmark design for theory-derived indicators: Offers reproducible assays (familiarity-controlled preference, scan events, pain-tail persistence) that other architectures could be tested against.

Industry relevance: For teams building agentic or embodied systems, the paper suggests a discipline of coupling internal state variables (valence, arousal, budget error) into action scoring rather than relying on behavioral outcomes alone — and it warns that "conscious-like" signatures may emerge cheaply from ordinary design choices. The code and a Colab notebook are released, making the approach directly reproducible.

Future Directions

• Scale the recurrence: Move the ipsundrum loop from a scalar recurrence to a structured latent space that could support richer perceptual content and learning.
• Cross-theory discrimination: Connect these markers to other formal theories (global workspace, integrated information) and test whether the assays can tell them apart rather than merely detecting them.
• Tighten estimation: Increase the number of seeds beyond the 20 used here, report seed-level distributions, and test robustness to hyperparameter perturbations such as affect gains and integrator decay, since several confidence intervals remain wide or censored.
• Sensorimotor-to-symbolic bridge: Explore whether a recursively bifurcating higher-order sensorimotor system could be paired with grid- and place-cell models to support abstract representations, with language models as interpretive layers on top of the sensorimotor substrate.

Target Audience

Researchers working on machine consciousness indicators, AI welfare and moral status, and neurosymbolic or cognitively inspired agent architectures; AI safety and evaluation practitioners interested in causal and ablation-based evidence; and theorists of affect, interoception, or predictive processing who want to see those ideas instantiated in inspectable code. The paper is written for readers comfortable with state machines, recurrence, and formal indicator frameworks, so it is less suited to a general audience seeking a plain introduction to machine consciousness.

Authors’ abstract

Indicator-based approaches to machine consciousness recommend mechanism-linked evidence triangulated across tasks, supported by architectural inspection and causal intervention. Inspired by Humphrey's ipsundrum hypothesis, we implement ReCoN-Ipsundrum, an inspectable agent that extends a ReCoN state machine with a recurrent persistence loop over sensory salience Ns and an optional affect proxy reporting valence/arousal. Across fixed-parameter ablations (ReCoN, Ipsundrum, Ipsundrum+affect), we operationalize Humphrey's qualiaphilia (preference for sensory experience for its own sake) as a familiarity-controlled scenic-over-dull route choice. We find a novelty dissociation: non-affect variants are novelty-sensitive (Delta scenic-entry = 0.07). Affect coupling is stable (Delta scenic-entry = 0.01) even when scenic is less novel (median Delta novelty ~ -0.43). In reward-free exploratory play, the affect variant shows structured local investigation (scan events 31.4 vs. 0.9; cycle score 7.6). In a pain-tail probe, only the affect variant sustains prolonged planned caution (tail duration 90 vs. 5). Lesioning feedback+integration selectively reduces post-stimulus persistence in ipsundrum variants (AUC drop 27.62, 27.9%) while leaving ReCoN unchanged. These dissociations link recurrence -> persistence and affect-coupled control -> preference stability, scanning, and lingering caution, illustrating how indicator-like signatures can be engineered and why mechanistic and causal evidence should accompany behavioral markers.