OS-001 through OS-010: what each protocol specifies, which product implements it, and the cognitive science behind it. Eight cognitive primitives plus two cross-cutting protocols (PRISM diagnostic, PULSE temporal).
OpenSentience publishes ten numbered protocols organized in two layers:
Cognitive primitives (OS-001 — OS-008) — eight individual capabilities,
each grounded in a cognitive science thread.
Cross-cutting protocols (OS-009 PRISM, OS-010 PULSE) — two sibling
protocols that sit above the cognitive primitives and above the [&] structural composition layer. PULSE declares how every loop in the ecosystem cycles, nests, and signals; PRISM measures how well those loops actually learn over time.
The protocols are independent specifications — they can be adopted individually or composed together. A system may publish a PULSE manifest without implementing any cognitive primitive, and PRISM can benchmark any PULSE-conforming system without bespoke per-system integration.
Cognitive basis: Hippocampal consolidation (McClelland et al. 1995)
What it specifies:
Three node types: episodic (events), semantic (facts), procedural (workflows)
Typed, weighted edges between nodes
Confidence scores with evidence-calibrated assignment
Consolidation cycle: decay low-confidence nodes, prune stale entries,
merge near-duplicates, strengthen co-retrieved edges
Memory timescales: working memory (session), short-term (hours), long-term
(persistent)
Design constraint from cognitive science: New memories must be stored quickly (hippocampal fast-write) without disrupting existing knowledge (neocortical slow-consolidation). This maps to immediate ETS/SQLite writes with background consolidation.
Implemented by: Graphonomous (store_node, store_edge, retrieve_context, run_consolidation)
Cognitive basis: Working memory gating (O'Reilly & Frank 2006)
What it specifies:
The kappa parameter: a measure of cyclicity in a knowledge subgraph
Strongly connected component (SCC) detection on retrieved neighborhoods
Routing decision: kappa = 0 implies fast path (acyclic, no conflict),
kappa > 0 implies deliberate path (cycles exist, may contain contradictions)
Topological features: SCC count, max SCC size, edge density within cycles
Design constraint from cognitive science: The prefrontal cortex gates information into working memory based on relevance and conflict signals. Kappa serves as the conflict signal — cyclic knowledge regions indicate unresolved tension that requires deliberation before action.
Implemented by: Graphonomous routing layer (topology_analyze, retrieve_context topology annotations)
Cognitive basis: Dual-process theory (Kahneman 2011)
What it specifies:
Four-phase deliberation: bid, debate, vote, commit
Argumentation framework: claims, warrants, rebuttals
Consensus mechanisms: majority, supermajority, unanimity
Timeout and escalation policies
Triggered when kappa routing indicates deliberation is needed
Design constraint from cognitive science: System 1 (fast, automatic) handles routine decisions. System 2 (slow, deliberate) engages when conflict is detected. The kappa threshold is the trigger that shifts from fast to deliberate processing.
Implemented by: AgenTroMatic (deliberate tool in Graphonomous provides the single-agent deliberation path)
Cognitive basis: Endogenous attention (Desimone & Duncan 1995)
What it specifies:
Three-phase cycle: survey (scan environment), triage (rank by salience),
dispatch (act on highest-priority items)
Salience scoring: combines urgency, recency, goal relevance, and confidence
Goal bias: active goals increase salience of related knowledge
Dispatch modes: act (execute), learn (gather more context), escalate
(request human intervention), idle (nothing needed)
Design constraint from cognitive science: Biological attention is goal-directed (endogenous) not just stimulus-driven (exogenous). The attention engine prioritizes based on the agent's active goals, not just what is new or loud.
Implemented by: Graphonomous attention module (attention_survey, attention_run_cycle)
Cognitive basis: Resource rationality (Lieder & Griffiths 2020)
What it specifies:
Three tiers: local_small (8B parameter models), local_large (70B+),
cloud_frontier (largest available)
Same topology and tool surface at every tier — only depth and latency differ
Escalation rules: when local_small confidence is below threshold, escalate
to local_large; when local_large is insufficient, escalate to cloud_frontier
Budget constraints: token limits, latency targets, cost caps per tier
Design constraint from cognitive science: Optimal cognition allocates computational resources proportional to decision importance. Simple decisions use fast/cheap processing; high-stakes decisions justify expensive computation.
Implemented by: Graphonomous and Agentelic (tier selection and escalation logic)
Cognitive basis: Executive function (Miyake et al. 2000)
What it specifies:
Permission taxonomy: filesystem, network, tool_invocation, graph_access
Lifecycle state machine: installed, enabled, running, disabled, removed
Graduated autonomy: observe, advise, act
Append-only audit trail with typed events
OTP-native implementation: GenStateMachine, ETS, DynamicSupervisor
Design constraint from cognitive science: Executive function provides inhibitory control (permissions), task switching (lifecycle), and cognitive flexibility (autonomy levels). Without executive function, an agent cannot self-regulate — it either does nothing or does everything.
Implemented by: open_sentience hex package (this project)
Cognitive basis: Immune system — self/non-self discrimination
What it specifies:
Five threat categories: prompt injection, model poisoning, side-channel,
identity spoofing, resource exhaustion
Detection rules per category
Defense protocols: quarantine, circuit-break, escalate
Integration with OS-006 for permission revocation and OS-008 for circuit
breaking
Implemented by: OpenSentience security module (draft)
Cognitive basis: Supervisory attentional system (Norman & Shallice 1986)
What it specifies:
Pipeline ordering enforcement (retrieve → route → act → learn)
Quality gates between pipeline stages
Sprint contracts: bounded execution with explicit goals and success criteria
Context management: 60% threshold, compaction, Graphonomous overflow
Generator-evaluator separation for adversarial grading
Implemented by: OpenSentience harness module (draft)
Cognitive basis: Meta-cognition + psychometrics (Item Response Theory, signal detection theory)
What it specifies:
9 continual-learning dimensions (retention, plasticity, transfer,
contradiction handling, etc.)
4-phase evaluation loop: compose → interact → observe → reflect → diagnose
BYOR (Bring Your Own Repo) ingestion — point PRISM at any repo and it will
generate scenarios
IRT calibration of scenario difficulty
Leaderboards, regression detection, fix suggestions
PULSE-aware: reads any system's PULSE manifest at runtime and injects
scenarios at the declared retrieve boundary, observing outcomes via the declared learn phase
Design constraint: A diagnostic that measures learning over time must itself be a closed loop — it must reflect on its own scenarios and evolve them based on what the inner loop fails on. Hence the 4-phase evaluation structure.
Implemented by: /PRISM/ Elixir/OTP codebase, Fly.io, 6 MCP machines (compose, interact, observe, reflect, diagnose, config), prism.opensentience.org
Cognitive basis: Closed-loop control theory (Wiener cybernetics 1948) + temporal cognition (Allen interval algebra 1983)
What it specifies:
Loop manifest schema (JSON Schema): pulse-loop-manifest.v0.1.json
5 canonical phase kinds: retrieve, route, act, learn, consolidate
(+ custom phases via custom_kind)
5 canonical cross-loop tokens (CloudEvents v1 envelopes):
TopologyContext, DeliberationResult, OutcomeSignal, ReputationUpdate, ConsolidationEvent
6 cadence types: event, periodic, streaming, idle,
cross_loop_signal, manual
6 substrate slots: memory, policy, audit, auth, transport, time
7 invariants: phase_atomicity, feedback_immutability,
append_only_audit, kappa_routing, quorum_before_commit, outcome_grounding, trace_id_propagation
12-test conformance suite — a runtime is PULSE-conforming when its
manifest validates and all 12 tests pass
Design constraint: Loops must be declarable in a vocabulary that is independent of their implementation language, runtime, or cognitive architecture. The same manifest schema must work for a SQLite-backed knowledge graph, a Cloudflare Worker, and a Phoenix LiveView.
Implemented by: /PULSE/ directory — JSON Schema + 11 reference manifests covering every [&] portfolio loop, pulse.opensentience.org
| Protocol | Primary Implementation | Secondary |
|---|---|---|
| OS-001 | Graphonomous | — |
| OS-002 | Graphonomous | — |
| OS-003 | AgenTroMatic | Graphonomous (single-agent path) |
| OS-004 | Graphonomous | — |
| OS-005 | Graphonomous | Agentelic |
| OS-006 | open_sentience | — |
| OS-007 | OpenSentience security module | — |
| OS-008 | OpenSentience harness module | — |
| OS-009 | `/PRISM/` Elixir/OTP | — |
| OS-010 | `/PULSE/` manifest standard | Every portfolio product publishes a conforming manifest |
Cross-protocol feedback: How should attention (OS-004) influence
deliberation thresholds (OS-003)?
Tier-aware governance: Should autonomy levels (OS-006) vary by model
tier (OS-005)?
Consolidation governance: Should the governance shim audit memory
consolidation events (OS-001)?
Multi-agent kappa: How does topological routing (OS-002) work across
agent boundaries?
PRISM scenario evolution: How aggressively should PRISM evolve scenarios
between cycles? Too aggressive and the system optimizes for adversarial noise; too conservative and improvement plateaus.
PULSE nesting depth: What is the practical maximum nesting depth before
the cross-loop signal volume becomes a substrate burden?