OpenSentience

Open research protocols for machine cognition, structured memory, deliberation topology, and agent governance.

Published by Ampersand Box Design under the [&] Protocol ecosystem.

"Intelligence is not generation. It is structured accumulation."

 ![Spec Status]() ![Docs]() ![Stack]()

What is OpenSentience?

OpenSentience is the research arm of the [&] Protocol ecosystem. It is not a product — it is a research organization that publishes:

1. Ten numbered protocols (OS-001 through OS-010) — each defining one cognitive capability or cross-cutting concern, grounded in cognitive science

2. Three runtime artifacts — a governance shim (OS-006), a benchmark engine (OS-009 PRISM), and a loop manifest standard (OS-010 PULSE)

3. Published research — cognitive science grounding, empirical benchmarks, and open questions

Other [&] portfolio products implement the protocols. OpenSentience defines them, grounds them in theory, and provides the thin enforcement layer that ties them together at runtime.

The Ten Protocols

Eight cognitive primitives (OS-001–OS-008) plus two cross-cutting protocols (OS-009, OS-010):

The Three-Protocol Stack

┌──────────────────────────────────────────────────────────┐
│  PRISM (OS-009)  — measures loops over time   diagnostic │
├──────────────────────────────────────────────────────────┤
│  PULSE (OS-010)  — declares loops + signals    temporal  │
├──────────────────────────────────────────────────────────┤
│  OS-001 … OS-008 — cognitive primitives       capability │
├──────────────────────────────────────────────────────────┤
│  [&]             — composes capabilities      structural │
└──────────────────────────────────────────────────────────┘

• [&] composes agents (*.ampersand.json) — structural layer

• PULSE gives them a heartbeat (*.pulse.json) — temporal layer

• PRISM measures their effect — diagnostic layer

Each protocol is independent. Adoption order is typically [&] → PULSE → PRISM.

Protocol Summaries

OS-001: Continual Learning

Hippocampal consolidation-inspired memory. Self-evolving knowledge graph with four node types (episodic, semantic, procedural, temporal), typed/weighted/decaying edges, multi-timescale consolidation, and no model weight modification — all learning is graph-structural.

OS-002: Topological Routing (κ-Routing)

Routes queries through knowledge graphs using topological structure, not just embedding similarity. The κ parameter balances exploitation (κ=0, follow strong edges) vs exploration (κ→∞, random walk). Inspired by prefrontal cortex working memory gating.

OS-003: Deliberation Orchestrator

Multi-agent consensus through structured phases: bid → debate → vote → commit. Uses weighted bipolar argumentation, Raft fast-path for agreement, PBFT for Byzantine tolerance, and reputation-weighted bid credibility. Maps to Kahneman's System 2.

OS-004: Attention Engine

Three-phase proactive attention cycle — survey (scan all sources, produce salience map) → triage (rank by urgency/novelty/goals) → dispatch (route to pipelines). Runs continuously with adaptive frequency. Grounded in endogenous top-down attention.

OS-005: Model Tier Adaptation

Graduated model selection across three tiers: local_small (1B–3B), local_large (7B–14B), cloud_frontier (70B+/API). Escalates when confidence falls below threshold; de-escalates by caching successful patterns as procedures. Based on resource rationality theory.

OS-006: Agent Governance Shim

The only protocol OpenSentience implements directly. Deny-by-default permissions, five lifecycle states (installed → enabled → running → disabled → removed), three autonomy levels (observe / advise / act), append-only audit trail. ETS-backed lookups at microsecond latency. Ships as the open_sentience hex package.

OS-007: Adversarial Robustness

Defense against prompt injection, model poisoning, side-channel attacks, identity spoofing, and resource exhaustion. Specification-stage protocol.

OS-008: Agent Harness

Pipeline enforcement above governance. Two enforced pipelines (reactive: query → recall → topology → deliberate → store; sprint: planner → generator → evaluator → commit). Quality gates with adversarial grading. Sprint contracts with acceptance criteria. Context management with 60% utilization threshold. Grounded in Norman & Shallice's Supervisory Attentional System.

OS-009: PRISM

Protocol for Rating Iterative System Memory. The first self-improving continual learning benchmark. Measures 9 CL dimensions (retrieval, transfer, uncertainty, feedback integration, forgetting, consolidation, temporal reasoning, multi-agent deliberation, composition). Four-phase evaluation loop: compose → interact → observe → reflect → diagnose. Features BYOR (Bring Your Own Repo), IRT calibration, three-layer judging, closed-loop verification, leaderboards, and actionable diagnostics.

OS-010: PULSE

Protocol for Uniform Loop State Exchange. Temporal algebra declaring how loops cycle, nest, and signal. Five canonical phase kinds (retrieve, route, act, learn, consolidate). Six cadence types (event, periodic, streaming, idle, cross_loop_signal, manual). Five canonical cross-loop tokens via CloudEvents v1.0 envelopes. BYOL (Bring Your Own Loop) — any system publishing a conforming manifest is automatically PRISM-evaluable.

Architecture

Governance Shim (OS-006) — OTP Supervision Tree

OpenSentience.Application
├── OpenSentience.PermissionEngine     (GenServer + ETS — deny-by-default policy)
├── OpenSentience.AuditWriter          (GenServer — batched append-only)
├── OpenSentience.AutonomyController   (GenServer + ETS — observe/advise/act)
├── OpenSentience.AgentSupervisor      (DynamicSupervisor)
│   ├── AgentLifecycle "agent-001"     (GenStateMachine → wraps real process)
│   ├── AgentLifecycle "agent-002"
│   └── ...
├── OpenSentience.MCP.Server           (Hermes MCP — governance tools)
└── OpenSentience.Telemetry            (telemetry handler)

Triple-Loop Nesting

The reference [&] ecosystem operates with three nested closed loops:

PRISM (outer)              compose → interact → observe → reflect → diagnose
  │
  └─ Graphonomous         retrieve → route → act → learn → consolidate
       │
       └─ Deliberation    survey → triage → dispatch → act → learn

PULSE manifests encode this nesting declaratively.

Cognitive Science Grounding

Every protocol maps to published research:

Empirical Results (OS-E001)

Benchmark of Graphonomous v0.3.3 on the full [&] portfolio (18,165 files across 14 projects):

Full results: `docs/spec/OS-E001-EMPIRICAL-EVALUATION.md`

κ Proof

`kappa_proof.js` is a browser-runnable proof verifying the κ-routing theorem across 1,926,351 test cases:

• 1,052,740 directed graphs (n=2–5): κ(G) > 0 ⟺ β₁(G) > 0 ⟺ nontrivial SCC

• 873,611 finite dynamical systems (n=2–7): κ(f) > 0 ⟺ periodic orbit (period > 1)

• Zero counterexamples

Repository Structure

opensentience.org/
├── docs/
│   ├── spec/
│   │   ├── README.md                       # Master spec — all 10 protocols (source of truth)
│   │   ├── OS-008-HARNESS.md               # Agent Harness Protocol
│   │   ├── OS-009-PRISM-SPECIFICATION.md   # PRISM benchmark engine spec
│   │   ├── OS-010-PULSE-SPECIFICATION.md   # PULSE loop manifest spec
│   │   └── OS-E001-EMPIRICAL-EVALUATION.md # Empirical benchmark results
│   ├── skills/                             # Agent skill guides (8 numbered modules)
│   ├── index.md                            # ReadTheDocs homepage
│   ├── architecture.md                     # OTP supervision tree details
│   ├── ecosystem-overview.md               # 10 protocols + 12 product relationships
│   ├── positioning.md                      # Governance stack & market positioning
│   ├── competitive-landscape.md            # Comparison with CrewAI, LangGraph, etc.
│   ├── harness-engineering.md              # Pipeline enforcement deep-dive
│   ├── runtime-walkthrough.md              # Concrete agent lifecycle walkthrough
│   ├── comparison-table.md                 # Feature matrix vs alternatives
│   ├── research.md                         # Cognitive science grounding (7+ hours)
│   ├── quickstart.md                       # Elixir integration guide
│   ├── faq.md                              # Common questions
│   ├── conf.py                             # Sphinx configuration
│   └── requirements.txt                    # sphinx>=8, myst-parser>=4
├── index.html                              # Marketing landing page
├── kappa_proof.js                          # Browser-runnable κ proof (1.9M test cases)
├── CLAUDE.md                               # Agent context
├── AGENTS.md                               # Agent interface guide
├── .readthedocs.yaml                       # ReadTheDocs build config
└── old_scrap/                              # Historical v1/v2 iterations (not current)

Technology Stack

Quickstart

The governance shim wraps any OTP supervision tree:

# 1. Add dependency
defp deps do
  [{:open_sentience, "~> 0.1"}]
end

# 2. Add to your supervision tree
children = [
  {OpenSentience, policy: "priv/governance/policy.yaml"},
  # ... your existing children
]

# 3. Install an agent
OpenSentience.install_agent("my-agent",
  permissions: %{filesystem: %{read: ["priv/**"]}},
  autonomy: :observe
)

Full integration guide: `docs/quickstart.md`

Ecosystem Relationships

OpenSentience defines protocols. Portfolio products implement them:

Does not use Supabase. The governance shim uses ETS (dev/edge) or pluggable backends. PRISM uses SQLite. PULSE uses embedded SQLite + sqlite-vec.

Documentation

Documentation is built with Sphinx + MyST and configured for ReadTheDocs:

# Build locally
cd docs
pip install -r requirements.txt
sphinx-build -b html . _build/html

Key documents

• [Master Spec](docs/spec/README.md) — All 10 protocols, architecture, implementation roadmap

• [OS-008 Harness](docs/spec/OS-008-HARNESS.md) — Pipeline enforcement, quality gates, sprint contracts

• [OS-009 PRISM](docs/spec/OS-009-PRISM-SPECIFICATION.md) — 9 CL dimensions, 4-phase evaluation, BYOR, IRT

• [OS-010 PULSE](docs/spec/OS-010-PULSE-SPECIFICATION.md) — Loop manifests, 5 phase kinds, 5 tokens, BYOL

• [Empirical Evaluation](docs/spec/OS-E001-EMPIRICAL-EVALUATION.md) — Graphonomous v0.3.3 benchmark results

• [Architecture](docs/architecture.md) — OTP supervision tree, component details

• [Cognitive Science](docs/research.md) — Research grounding for all protocols

• [Competitive Landscape](docs/competitive-landscape.md) — Comparison with CrewAI, LangGraph, AutoGen, etc.

• [FAQ](docs/faq.md) — Common questions answered

• [Quickstart](docs/quickstart.md) — Elixir integration guide

Design Principles

1. Research first, code second — Every protocol is grounded in cognitive science before implementation

2. Thin shim, not thick runtime — Hex package dependency, not a daemon

3. Graduated autonomy — Three levels (observe, advise, act) build trust incrementally

4. Deny by default — No implicit permissions; every operation requires explicit policy

5. Append-only audit — Every event is logged immutably; no updates, no deletes

6. Protocol-driven — Numbered specs (OS-00X) are the source of truth; implementations reference them

7. OTP-native — GenServer, ETS, supervision trees; no external dependencies beyond BEAM

License

Apache 2.0 — Open research.

Author

Travis Burandt — Ampersand Box Design