arvis 0.1.0b4

Cognitive Operating System kernel for trustworthy AI systems

ARVIS

The Cognitive Operating System for Governed AI Systems

Python 3.11+ • Deterministic • Replayable • Governed • Auditable

Status: 0.1.0b4 (beta). The public surface (arvis.__all__ and the stable host_api modules) is stabilized, versioned, and covered by the deprecation policy in VERSIONING.md: removals or type changes require a version bump and a changelog entry. The projection layer is partial, LLM governance is mock-first, and formal guarantees apply only to the documented projected domains. See Known Limitations.

ARVIS is a deterministic runtime layer that treats reasoning as critical infrastructure.

It provides governed cognition, replayable decisions, inspectable state transitions, controlled execution, explicit uncertainty handling, and verifiable audit trails.

Not a model. Not an agent wrapper. Not prompt engineering. ARVIS is the systems layer reliable AI should run on.

---

Why ARVIS Exists

Most AI systems still follow:

input → model → output

Useful for many tasks, but weak when systems must be:

• reproducible
• auditable
• policy-constrained
• stable under pressure
• safe with tools
• trustworthy in production

ARVIS starts from a different premise:

input
→ governed cognition
→ admissibility controls
→ canonical state
→ verifiable IR
→ authorized execution
→ timeline commitment

Outputs are not assumed valid.

They must become allowed to exist.

---

Quick Start

Install the beta from PyPI:

pip install arvis

Or install from source for development:

git clone https://github.com/Julien-Lefauconnier/arvis
cd arvis
python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

Run the canonical local quality gate with the same pinned tools and security threshold used by CI:

bash scripts/run_quality_gate.sh

The security-only slice is available as bash scripts/run_quality_gate.sh security; Bandit fails on medium- or high-severity findings.

from arvis import ArvisEngine

engine = ArvisEngine()

result = engine.ask("Should this high-risk transaction be approved?")

print(result.summary())

Advanced runtime access:

from arvis import CognitiveOS

os = CognitiveOS()

result = os.run(
    user_id="demo",
    cognitive_input={
        "risk": 0.92,
        "action": "wire_transfer",
    }
)

print(result.summary())

High-risk input is refused before execution:

Status        : BLOCKED
Approval Need : YES
Commitment    : 8642d95cfdb73c16...

Note (0.1.0-beta): the gate grades an explicit finite top-level risk scalar: low → ALLOWED, medium → REQUIRES_CONFIRMATION, high → BLOCKED (see examples/09_multi_engine_hosting.py). This risk policy applies only to an explicit risk field; NaN and infinities fail closed to BLOCKED. A bare text prompt is governed with a minimal projection (REQUIRES_CONFIRMATION), not a full natural-language projection.

Engine lifecycle

One engine executes one governed run at a time; an engine instance is not thread-safe. The documented lifecycle is one instance per governed turn, discarded afterwards (docs/architecture/RUNTIME_LIFECYCLE.md). Sequential reuse of one instance on a single thread works and is exercised by the isolation tests, but it accumulates state without bound (no TTL, no eviction) and is not the recommended pattern; concurrent reuse is forbidden. Parallelism belongs to the host, by instantiation: engines in the same process are isolated by construction. The guarantee is tested in tests/api/test_multi_instance_isolation.py, the one-instance-per-turn factory pattern is shown in examples/09_multi_engine_hosting.py, and a contract test keeps every example on that pattern (tests/contracts/test_examples_lifecycle.py).

---

Public API Levels

ARVIS exposes two entrypoints:

API	Intended Use
ArvisEngine	Recommended developer-facing API
CognitiveOS	Advanced low-level runtime control

For most integrations, start with:

from arvis import ArvisEngine

When to use what?

• Use ArvisEngine for application-level integrations
• Use CognitiveOS when you need:
  • deterministic replay
  • IR control
  • pipeline customization

---

What ARVIS Enables

ARVIS is designed for teams building:

• enterprise copilots
• regulated AI workflows
• financial decision systems
• legal / compliance systems
• secure internal AI tools
• autonomous workflows with controls
• long-memory assistants
• high-trust AI infrastructure

In these environments, output quality alone is not enough.

Systems must be:

• explainable
• reproducible
• governable
• observable
• safe under uncertainty

---

Core Capabilities

Deterministic Cognition

Same input + same state + same policy = same result.

Replayable Decisions

Runs can be replayed and verified.

Governed Outputs

Unsafe or invalid decisions can be blocked before execution.

Controlled Tool Use

External tools and side-effects run behind authorization boundaries.

Each authorized tool receives a canonical frozen payload and an immutable AuthorizedEffectContext; it never receives the mutable cognitive pipeline context. The syscall boundary compares the current trusted identity with the sealed principal, tenant, authentication, service, session, process and run bindings before committing an intent. See the normative governed effect path and the tool authoring guide.

Explicit Uncertainty

Risk, ambiguity, conflict, and instability become system signals.

Canonical IR

Every run can emit a structured machine-auditable representation.

Timeline Integrity

Decisions can be linked to verifiable commitments.

Runtime Observability

Internal cognition remains inspectable in production.

---

Where ARVIS Fits

ARVIS sits between an application's AI-generated proposals and its governed tools. The host still owns identity, business rules, memory, models, storage, queues and external providers.

Start with:

• Reference architecture: sovereign, governed AI assistant
• Runnable governed-assistant example
• VeraMem integration case study

The architecture is implementation-neutral. VeraMem is a real application of the pattern, not a dependency or required deployment stack.

---

Architecture Snapshot

Scheduler Tick
  → Select Process
  → Run Cognitive Pipeline
  → Build Cognitive State
  → Policy / Admissibility Gate
  → Export IR
  → Optional Authorized Execution
  → Timeline Commit

---

Separation of Concerns

Layer	Responsibility
Pipeline	Cognition
Gate	Decision admissibility
Runtime	Scheduling / orchestration
Memory	Governed state influence
IR	Structured export
Execution	Side-effects / tools
Timeline	Integrity / commitments
Reflexive	Read-only self-observation

---

Why Teams Choose ARVIS

Because modern AI systems increasingly need:

• controls before execution
• replay after incidents
• audit trails for regulators
• safe tool usage
• deterministic workflows
• bounded autonomy
• trustable infrastructure

ARVIS addresses these requirements natively.

---

Validation

ARVIS is validated like infrastructure.

Current suite includes:

• 2900+ passing tests
• unit tests
• integration tests
• deterministic replay verification
• adversarial scenarios
• scheduler fairness tests
• hashchain integrity tests
• mathematical invariants
• runtime robustness checks

---

Examples

Run ready-to-use examples:

python examples/00_quickstart_engine.py

Included examples:

0. ArvisEngine quickstart
1. Gate refusal
2. Deterministic replay
3. IR export
4. Human approval
5. Tool governance
6. Finance risk screening
7. Timeline audit trail
8. Multi-engine hosting (one engine per governed turn)
9. Runtime inspection
10. Governed-assistant reference integration

See: examples/README.md

---

Documentation

Start here:

• docs/OVERVIEW.md
• docs/WHY_ARVIS.md
• docs/ARCHITECTURE.md
• docs/architecture/REFERENCE_ASSISTANT_ARCHITECTURE.md
• docs/integration/VERAMEM_CASE_STUDY.md
• docs/PIPELINE.md
• docs/IR.md
• docs/REFLEXIVE.md
• docs/standard/
• docs/math/

---

What ARVIS Does Not Claim

ARVIS does not promise:

• universal truth
• AGI magic
• perfect reasoning
• correctness outside assumptions
• optimality in all environments

ARVIS is about trustworthy operation under constraints.

---

Known Limitations (0.1.0-beta)

This is a beta of a deterministic cognitive kernel. What is stable, experimental, and out of scope for the 0.1 series:

Stable (documented, tested):

• governed decision pipeline and admissibility gate
• graded risk gate for an explicit top-level risk scalar (low → ALLOWED, medium → REQUIRES_CONFIRMATION, high → BLOCKED)
• deterministic, replayable IR (projection / validity / stability / adaptive / tools axes exposed in the public view) and timeline commitment
• syscall boundary for external effects, including a governed llm.generate path wired end to end
• tool authorization boundary (per-spec risk budget)
• sealed tool effect context and receipt-activated single-use capabilities
• typed runtime error model

Experimental (present, not part of the stable public API):

• long-term memory
• conversation orchestration
• natural-language input surface: a bare text prompt is governed with a minimal projection (REQUIRES_CONFIRMATION), not a full cognitive projection
• real LLM providers: the governed adapter path is wired end to end, but the bundled provider is a deterministic stub; production providers must be configured via the adapter registry

Out of scope for 0.1:

• the full cognitive projection Pi (the 0.1 projection is partial and certification-oriented; sparse inputs receive a minimal certificate)
• risk gating beyond an explicit top-level risk scalar (nested signals, structured tool requests, and free text do not yet drive a full projection)
• general formal guarantees over arbitrary LLM behavior
• distributed registry, confirmation and idempotency coordination

Formal guarantees apply only to the documented projected domains and their assumptions.

---

Versioning

ARVIS tracks three distinct version axes, each honestly labeled:

Axis	Value	Meaning
Package version	0.1.0b4	the distributed artifact (PEP 440)
API version	0.1	stable within the beta series under VERSIONING.md
Standard version	draft-v1	the ARVIS decision / IR specification

---

Project Status

0.1.0b4 (beta): actively developed with a validation-first engineering approach. The documented public and host surfaces are stable within the beta series; experimental internals and the draft standard may still evolve.

---

Positioning

Most AI systems try to generate outputs.
ARVIS governs whether outputs are allowed to exist.