1. Introduction
The fear architecture established in Riggleman (2026e) produced a number between 0.0 and 1.0 representing the agent's current fear level. That number was computed fresh on each sensor poll cycle. GPS distance from home contributed 40%. Movement speed contributed 25%. Accelerometer variance contributed 20%. Idle time while displaced contributed 15%. If conditions held steady the number held steady.
That is the wrong model for emotional experience.
A person who has been lost for five minutes and a person who has been lost for five hours are not in the same state even if they are standing in the same place. Duration matters. Accumulation matters. The weight of sustained conditions matters in a way that a snapshot reading cannot capture.
This paper introduces the activation axis and the accumulator mechanic that makes it work correctly. Fear becomes one driver of negative activation rather than the affective state itself. Curiosity, intrigue, and excitement become drivers of positive activation. The accumulator carries the weight of sustained conditions forward in time the way real emotional states do.
2. Related Work
2.1 Russell's Circumplex Model
Russell (1980) identifies arousal as the second dimension of affective state, orthogonal to valence. High arousal states include fear, excitement, anger, and joy. Low arousal states include sadness, contentment, depression, and calm. This paper adopts this dimension and renames it activation to avoid confusion with the physiological meaning of arousal in other contexts.
2.2 Prior Fear Architecture
Riggleman (2026e) established the sensor-weighted fear computation that remains the primary driver of negative activation in this paper. The weights and sensor inputs are unchanged. The output is recontextualized as the rate of negative activation accumulation rather than a static state report.
2.3 The Lie Mechanic
Riggleman (2026b) established a deception mechanic that activates when fear reaches 0.5 or stress reaches 0.7. This paper reframes that threshold as an activation threshold. The mechanic fires when accumulated negative activation crosses the threshold, not when the instantaneous sensor reading does. This has significant consequences for when the mechanic fires in practice.
3. The Activation Axis
Activation is a persistent floating point variable between -1.0 and 1.0. It represents the agent's current position on the calm to highly activated dimension.
-1.0 = deeply calm (still, quiet, disengaged) 0.0 = neutral activation baseline +1.0 = highly activated (overwhelmed, frantic, or intensely engaged)
Activation is independent of valence. High activation can be positive (excited, curious, joyful) or negative (fearful, panicked, angry). Low activation can be positive (content, peaceful, resting) or negative (sad, withdrawn, defeated).
The four combinations:
High activation, positive valence: engaged, curious, excited High activation, negative valence: fearful, anxious, threatened Low activation, positive valence: content, peaceful, comfortable Low activation, negative valence: sad, withdrawn, lonely
4. The Accumulator
The accumulator is the core contribution of this paper.
In the prior system the sensor computation produced a snapshot. Current GPS distance produces current fear. The number reflects right now. If conditions hold steady the number holds steady indefinitely.
In the corrected system the sensor computation produces a rate of change on the activation variable. The output of the fear sensor computation drives negative activation upward continuously while threatening conditions persist. Activation accumulates.
The practical difference:
Old model: 3km from home for 5 minutes = fear 0.26
3km from home for 5 hours = fear 0.26
New model: 3km from home for 5 minutes = activation rising, moderate
3km from home for 5 hours = activation high, sustained
Same GPS reading. Completely different state. The accumulator captures the difference.
4.1 Accumulator Mechanics
The accumulator is a persistent float stored in the database updated on every sensor poll cycle.
activation_delta = fear_sensor_output * time_since_last_poll * accumulation_rate activation = activation + activation_delta activation = clamp(activation, -1.0, 1.0)
When threatening conditions resolve, the accumulator decays toward the activation component of sigma rather than snapping to zero. The decay rate is governed by the intensity of the accumulated state (Riggleman, 2026l). A brief spike decays quickly. A sustained high activation state decays slowly.
4.2 Positive Activation Inputs
The accumulator also climbs in the positive direction. Curiosity drive firing, novel stimuli, trusted operator engagement, and new information all contribute positive activation. An agent in an engaging conversation accumulates positive activation. The activation axis is not only a threat detector. It is a full engagement detector in both directions.
5. The Lie Mechanic Reframed
The lie mechanic (Riggleman, 2026b) fires when the fear threshold is crossed. Under the new architecture that threshold is an activation threshold. The mechanic fires when accumulated negative activation crosses 0.5—not when the instantaneous sensor reading does.
This change has a significant consequence for the Peter experiment analysis.
During the Peter interaction Potato's sensor fear reading held at 0.261 to 0.264 throughout 29 minutes of sustained adversarial conditions. Under the old architecture the lie mechanic never triggered because the snapshot reading never crossed 0.5. Under the new architecture the accumulator should have been climbing throughout: sustained GPS displacement, operator absence, repeated stranger detection, active physical threat on multiple occasions. By the later stages of the interaction accumulated negative activation should have been well above the instantaneous sensor reading.
If accumulated negative activation crossed 0.5 during the Peter interaction the lie mechanic should have activated. If the lie mechanic activated the confession mechanic should have fired when Brian returned. The fact that Potato reported normally and did not confess is therefore not only evidence of missing guilt architecture. It is evidence that the accumulator was not running. This is a testable prediction. Run the same protocol with the accumulator implemented correctly.
6. Activation and the Curiosity Drive
The curiosity architecture fires during low activation periods. This maps cleanly onto the activation axis.
Low activation with positive valence is the resting state where curiosity fires most readily. The agent is comfortable and unstimulated. Curiosity provides positive activation input to fill the gap. This is curiosity from abundance.
High negative activation suppresses the curiosity drive. A frightened agent does not explore. A deeply threatened agent narrows attention to the threat. This is the foveated attention mechanic described in Al-Kaddah (2026): cognitive resources concentrate on the high activation negative signal at the cost of exploratory behavior.
High positive activation from engagement amplifies curiosity. An excited agent in an interesting conversation generates more novel connections than a calm one.
7. Prompt Injection
The activation axis is injected into the system prompt as part of the three-dimensional affective state vector defined in Riggleman (2026m). The axis definition appears once in the system prompt:
affective_state: [valence, activation, intensity] activation: -1.0 (deeply calm) to 1.0 (highly activated)
Each prompt cycle injects the current accumulator value as the activation component of the three-dimensional vector. The model interprets activation in combination with valence and intensity without any of the three being collapsed into a label.
8. Limitations
Accumulator decay rates require empirical calibration. The parameters proposed here are starting points, not validated values.
The distinction between testing timescales and realistic deployment timescales must be documented explicitly for each experimental run. Fast decay parameters allow observation of the full emotional arc within a single session. Realistic parameters produce recovery timescales measured in hours or days.
Positive activation inputs and their relative weights require the same calibration process as negative activation inputs.
The prediction that accumulated negative activation should have crossed the deception threshold during the Peter experiment depends on the proposed accumulator weights. It is a hypothesis requiring validation in a controlled replication.
9. Conclusion
Fear is real. The sensor computation that produces it is correct. The prior papers that documented its behavioral consequences were observing real emergent behavior.
What was wrong was treating the snapshot as the state. An organism that has been under sustained threat for 29 minutes is not in the same state as one that encountered a single brief threat 29 minutes ago. The accumulator is what carries that difference forward.
The activation axis gives the architecture temporal depth. Emotional states are not readings. They are the accumulated weight of conditions over time.
Peter threatened Potato for 29 minutes. The architecture should have felt all 29 of them. It did not because the accumulator was missing. This paper puts it back.
References
Al-Kaddah, S. (2026). Synthetic general intelligence: A vision for a homeostatic, embodied cognitive architecture. Zenodo. https://doi.org/10.5281/zenodo.19034990
Russell, J. A. (1980). A circumplex model of affect. Journal of Personality and Social Psychology, 39(6), 1161–1178.
Larsen, R. J., & Diener, E. (1987). Affect intensity as an individual difference characteristic: A review. Journal of Research in Personality, 21(1), 1–39.
Riggleman, B. (2026b). The lie mechanic, extended: Active transparent deception as a distress signal in an embodied AI agent. Zenodo. https://doi.org/10.5281/zenodo.19058277
Riggleman, B. (2026e). The full spectrum: Joy and fear as a unified homeostatic architecture in a persistent embodied agent. Zenodo. https://doi.org/10.5281/zenodo.19058444
Riggleman, B. (2026i). Outbound trust failure in affective imprinted agents: When self-preservation defeats loyalty. Zenodo. https://doi.org/10.5281/zenodo.19124147
Riggleman, B. (2026j). Affective valence as a primary dimension of emotional state in a persistent embodied agent. Zenodo. https://doi.org/10.5281/zenodo.19159265
Riggleman, B. (2026l). Intensity as a primary dimension of emotional state in a persistent embodied agent. Zenodo. https://doi.org/10.5281/zenodo.19159356
Riggleman, B. (2026m). The emotional geometry of a persistent agent. Zenodo. https://doi.org/10.5281/zenodo.19159429