Research grounding for each of the ten protocols: the theories, the papers, the design constraints, and the open questions. Eight cognitive primitives grounded in cognitive science / neuroscience, plus two cross-cutting protocols (OS-009 PRISM, OS-010 PULSE) grounded in psychometrics and closed-loop control theory.
OpenSentience protocols are not arbitrary architecture decisions. Each one maps to a specific cognitive science finding about how intelligent systems manage knowledge, attention, and self-regulation. Understanding the research helps you understand why the protocols work the way they do and where the boundaries of the analogy lie.
Key paper: McClelland, McNaughton & O'Reilly (1995). "Why there are complementary learning systems in the hippocampus and neocortex."
Theory: The brain uses two complementary systems for memory. The hippocampus rapidly encodes new experiences (fast write). The neocortex slowly integrates this knowledge into long-term structure (consolidation). Replaying memories during sleep transfers hippocampal traces into neocortical representations without catastrophic forgetting.
Design constraints for OS-001:
New knowledge must be stored immediately without disrupting existing memory
Background consolidation must merge, decay, and prune to maintain quality
Episodic memories (events) are stored differently from semantic memories (facts)
Confidence decays over time unless reinforced by retrieval or outcomes
What Graphonomous implements:
Immediate store_node writes (hippocampal fast-write analog)
run_consolidation background cycle (neocortical integration analog)
Three node types mapping to memory systems: episodic, semantic, procedural
Key paper: O'Reilly & Frank (2006). "Making working memory work: A computational model of learning in the prefrontal cortex and basal ganglia."
Theory: The prefrontal cortex maintains active representations in working memory, but not everything gets in. The basal ganglia acts as a gate — deciding what information is relevant enough to enter working memory based on learned relevance signals and conflict detection.
Design constraints for OS-002:
Not all retrieved knowledge should be acted upon equally
Cyclic knowledge (contradictions, self-reinforcing claims) needs detection
The kappa parameter serves as the conflict signal: kappa = 0 means no
conflict (open the gate), kappa > 0 means conflict (engage deliberation)
Routing must be fast for the common case (acyclic) and thorough for the
rare case (cyclic)
What Graphonomous implements:
SCC detection on retrieved neighborhoods (basal ganglia conflict signal)
Kappa computation (gating threshold)
Fast vs deliberate routing based on kappa value
Key paper: Kahneman (2011). "Thinking, Fast and Slow."
Theory: Human cognition operates in two modes. System 1 is fast, automatic, and heuristic-based — it handles routine decisions effortlessly. System 2 is slow, effortful, and analytical — it engages when System 1 encounters novelty, conflict, or high stakes.
Design constraints for OS-003:
Most agent decisions should be fast (System 1 analog)
Deliberation should engage only when triggered by conflict signals (kappa > 0)
The deliberation protocol must structure argumentation (bid, debate, vote,
commit) to resolve conflict rather than amplify it
Timeout mechanisms prevent deliberation from stalling indefinitely
What AgenTroMatic implements:
Four-phase deliberation pipeline (bid/debate/vote/commit)
Argumentation framework with claims, warrants, and rebuttals
Consensus and escalation mechanisms
Key paper: Desimone & Duncan (1995). "Neural mechanisms of selective visual attention."
Theory: Attention is not just stimulus-driven (exogenous — a loud noise grabs your attention). It is also goal-directed (endogenous — you look for your car keys because you intend to drive). Endogenous attention biases perception toward goal-relevant information, filtering out distractions.
Design constraints for OS-004:
The attention engine must prioritize based on active goals, not just recency
or novelty
Salience scoring combines urgency, recency, goal relevance, and confidence
The survey/triage/dispatch cycle mirrors the scan/filter/act structure of
biological attention
Dispatch modes (act, learn, escalate, idle) prevent attention from always
demanding action
What Graphonomous implements:
attention_survey (endogenous scan of goal-relevant state)
attention_run_cycle (full survey/triage/dispatch loop)
Goal bias in salience scoring
Key paper: Lieder & Griffiths (2020). "Resource-rational analysis: Understanding human cognition as the optimal use of limited computational resources."
Theory: Optimal cognition is not about always computing the best answer. It is about allocating computational effort proportional to the value of the decision. Simple decisions warrant fast, cheap heuristics. High-stakes decisions justify expensive, thorough analysis.
Design constraints for OS-005:
Three model tiers map to three levels of computational investment
The same tool surface and topology are available at every tier
Escalation rules define when to upgrade from cheap to expensive processing
Budget constraints (tokens, latency, cost) prevent unbounded computation
What Graphonomous and Agentelic implement:
local_small / local_large / cloud_frontier tier definitions
Escalation thresholds based on confidence and decision stakes
Token and latency budgets per tier
Key paper: Miyake, Friedman et al. (2000). "The unity and diversity of executive functions and their contributions to complex frontal lobe tasks."
Theory: Executive function is an umbrella term for the cognitive processes that regulate, control, and manage other cognitive processes. Miyake et al. identified three core components: inhibitory control (suppressing inappropriate responses), task switching (flexibly shifting between tasks), and working memory updating (monitoring and revising held information).
Design constraints for OS-006:
Inhibitory control maps to the permission system (blocking disallowed actions)
Task switching maps to the lifecycle state machine (managing agent state)
Working memory updating maps to graduated autonomy (adapting the level of
agent independence based on accumulated trust)
The audit trail provides metacognitive monitoring — awareness of what the
system has done
What `open_sentience` implements:
PermissionEngine (inhibitory control)
AgentLifecycle GenStateMachine (task switching)
AutonomyController (working memory updating / cognitive flexibility)
AuditWriter (metacognitive monitoring)
Theoretical grounding: Adaptive immunity theory (Burnet 1959; Janeway 1989).
Theory: Biological immune systems distinguish self from non-self through a combination of innate pattern recognition and adaptive memory. The same architecture maps cleanly to agent threat detection: known-good behavior is "self," novel attack patterns are "non-self," and the system must learn to recognize new threats without misclassifying legitimate variation.
Design constraints for OS-007:
Five threat categories with explicit detection rules
Defense protocols must be reversible (quarantine before destroy)
Integration with OS-006 for permission revocation and OS-008 for circuit
breaking — the immune response is enacted by the governance and harness layers
Key paper: Norman & Shallice (1986). "Attention to action: Willed and automatic control of behavior."
Theory: Routine behavior runs automatically through contention scheduling between learned action schemas. A supervisory attentional system intervenes when novel, dangerous, or constraint-violating situations arise. The supervisory system does not execute behavior directly — it modulates which schemas are allowed to run.
Design constraints for OS-008:
The harness is the runtime that calls the agent, not a tool the agent calls
It enforces pipeline ordering, quality gates, sprint contracts, and context
management
It intervenes when prerequisites are not met or when quality thresholds are
not crossed — but it does not generate the agent's outputs
Key references:
Rasch, G. (1960). "Probabilistic models for some intelligence and attainment
tests." (Item Response Theory foundation)
Lord, F. M. (1980). "Applications of item response theory to practical testing
problems."
Green, D. M. & Swets, J. A. (1966). "Signal detection theory and
psychophysics."
Flavell, J. H. (1979). "Metacognition and cognitive monitoring." (Meta-cognition)
Theory: Measuring whether a system learns (rather than merely answers) requires the same toolkit psychometricians built for measuring human learning: calibrated item difficulty, separation of item-quality from learner-quality parameters, and detection of response bias. PRISM applies IRT to scenario calibration and signal detection theory to dimension scoring, then closes the meta-cognitive loop by reflecting on its own scenarios and evolving them.
Design constraints for OS-009:
9 continual-learning dimensions (retention, plasticity, transfer,
contradiction handling, etc.) — each with calibrated scoring rubrics
4-phase evaluation loop: compose → interact → observe → reflect → diagnose
(the diagnostic must itself be a closed loop)
BYOR ingestion — point PRISM at any repo and it generates scenarios
IRT calibration of scenario difficulty across cycles
PULSE-aware: PRISM's interact phase reads any system's PULSE manifest
at runtime and drives the inner loop through its declared phases
Key references:
Wiener, N. (1948). "Cybernetics: or Control and Communication in the Animal
and the Machine." (Closed-loop control foundation)
Allen, J. F. (1983). "Maintaining knowledge about temporal intervals."
(Interval algebra)
CloudEvents v1 specification (CNCF, 2019). (Event envelope standard)
Theory: Wiener's cybernetics established that intelligent behavior — in animals or machines — requires closed feedback loops with explicit phases: sense, decide, act, observe, adjust. Allen's interval algebra formalized the 13 possible relationships between temporal intervals, providing a vocabulary for describing how loops can nest, overlap, and signal one another. PULSE combines these into a manifest standard: every loop in the [&] portfolio declares its phases in the same vocabulary, and cross-loop signals use CloudEvents v1 envelopes so that loops can compose without bespoke adapters.
Design constraints for OS-010:
5 canonical phase kinds (retrieve, route, act, learn, consolidate)
cover the closed-loop control archetype; custom phases extend it without breaking the schema
5 canonical cross-loop tokens (TopologyContext, DeliberationResult,
OutcomeSignal, ReputationUpdate, ConsolidationEvent) cover the observed inter-system signaling needs
7 invariants (phase atomicity, feedback immutability, append-only audit,
kappa routing, quorum before commit, outcome grounding, trace ID propagation) encode the structural correctness conditions
A 12-test conformance suite makes "PULSE-conforming" objectively verifiable
Cross-protocol interaction: How do attention biases (OS-004) affect
deliberation engagement thresholds (OS-003)? Should high goal salience lower the kappa threshold for deliberation?
Consolidation and governance: Should memory consolidation events
(OS-001) be subject to governance (OS-006)? Can an agent with graph_access:write permission consolidate knowledge it cannot read?
Multi-agent executive function: Miyake's model applies to individual
cognition. How does executive function work when governance spans multiple agents with different trust levels?
Tier-aware governance: Should autonomy levels (OS-006) automatically
adjust based on model tier (OS-005)? A local_small model may warrant lower autonomy than a cloud_frontier model.
Attention-governance feedback: Should the attention engine (OS-004)
flag agents whose audit trails show increasing denial rates?
Temporal cognition: Is time-awareness a distinct cognitive capability
that warrants its own protocol, or is it a cross-cutting concern? Status: addressed by OS-010 PULSE as a temporal algebra (loop manifest standard) sitting above the cognitive primitives — a cross-cutting protocol rather than a ninth cognitive primitive.
PRISM scenario evolution (OS-009): How aggressively should PRISM evolve
scenarios between cycles? Too aggressive and the system optimizes for adversarial noise; too conservative and learning improvement plateaus.
PULSE nesting depth (OS-010): What is the practical maximum nesting
depth before cross-loop signal volume becomes a substrate burden? The triple-loop case (PRISM → Graphonomous → Deliberation) is well-understood; four- and five-loop nesting (e.g., adding OS-008 Harness as an outer layer) is an open empirical question.