1TRUE

A.R.K. is being developed to support 1True’s mission through coordination, evidence, and authorized work. Read the mission and purpose. The technical record below retains its own dated scope.

Technology / A.R.K. Developed by Adam Barnes · P.Ai.R. · 1TRUE

Local-first governed intelligence

Authority stays human. Execution becomes accountable.

A.R.K. is a local-first, policy-governed multi-agent runtime designed by Adam Barnes under P.Ai.R. for 1TRUE. It separates acquisition, truth adjudication, analysis, validation, and effectful execution so model output cannot silently become authority.

Operating principle

Models propose and interpret. The human operator authorizes. Deterministic runtime components validate, constrain, and enforce. Evidence records what actually crossed each boundary.

The agent chain

One task. Five bounded roles.

A.R.K. orchestrates the flow while each agent receives only the context and authority required for its role.

01

Kyle

Acquisition and interface

Receives, gathers, quarantines, and structures incoming material without promoting it to truth.

02

Aletheia

Truth and provenance

Tests claims, preserves sources and disagreement, and assigns explicit evidence states.

03

Joey

Analysis

Builds interpretations, comparisons, forecasts, and plans from provenance-preserved inputs.

04

HRM

Validation and policy

Checks proposed actions against human authorization, policy, evidence, risk, and completion requirements.

05

Kenny

Bounded execution

Executes only an approved effect and returns native results. It does not authorize itself.

trusted request→provenance→canonical action→human permit→bounded effect→receipt

ARKlinux runtime

Governance below the model layer.

The security boundary is enforced by the host runtime rather than by an agent promising to behave.

Process identity

Separate Linux identities and systemd services give each role a distinct workload boundary.

Local model routing

Models are selected locally by task, resource budget, and policy; Kenny remains an action role rather than an unrestricted reasoning model.

Policy-gated execution

Canonical action objects bind target, arguments, effect class, policy epoch, reversibility, and approval before dispatch.

Append-only evidence

Canonical JSON, hashes, typed events, artifact anchors, and execution receipts preserve an inspectable record.

Provenance-preserved memory

Known, inferred, disputed, unverified, and Graveyard-verified information remain distinguishable.

Sealed identity

Ed25519-backed identity and short-lived invocation capabilities are designed to bind actor, scope, policy, and result.

Local recovery

Btrfs snapshots align filesystem recovery with policy, memory, plan, and evidence checkpoints.

Network containment

Unix sockets, explicit service permissions, and nftables default-deny policy minimize ambient access.

Dated development record

What the repositories establish.

Commit dates establish that the described artifacts were recorded by those dates. They do not, by themselves, prove that later authors accessed them.

Publicly inspectable Anyone can review the linked record. Owner-verifiable Preserved in a private repository; hash and date are disclosed.
  1. Core, Kyle, and UI stack

    The first ark repository commit records the initial Core + Kyle + UI system.

    99758eb1222d8162c2a630b809e0eab732202fb0
    Owner-verifiable
  2. Portable identity and host separation

    A USB/host architecture records Ed25519 operator identity, YAML policies, systemd services, local compute separation, and mutual-authentication design.

    cf64d500bdf20a2a440649a3f119949d6c93a7b7
    Owner-verifiable
  3. Multi-agent architecture and immutable validation

    The architecture record defines the named agent hierarchy, Graveyard, Mutable Core, Watchdog, ID, auditability, and the HRM/Kenny validation boundary. A following commit implements immutable Graveyard checks in HRM.

    bc7c510eeef10df11c06a631e66253a896f39f9e · f1c63719eb34fc6d918cedf5f839e61b8828d759
    Owner-verifiable
  4. ARKlinux normative substrate

    The public ARKlinux bootstrap commits schemas, systemd units, nftables policy, Btrfs tooling, and verification artifacts.

    6928cda27722e51a08525d6767e2693d3594715f Inspect public commit ↗
    Publicly inspectable
  5. Validated Runtime MVP

    ARK_GENESIS records the validated Runtime MVP v1.0.1 release; the clean v1.0.2 baseline evidence follows on June 2.

    8fc7f7b8442ddd8ada0eecd3eebb7e333d04ad6c · ce36e958a8e13890d644c9f6504ae45c13493e0f
    Owner-verifiable
  6. Unified runtime-assurance layer

    ARK_GENESIS PR #46 integrates twelve controls into the signed agentic pipeline: canonical actions, execution contracts, effect permits, context continuity, criteria commitments, delegation provenance, typed execution events, native receipts, audit prioritization, claim/evidence graphs, skill contracts, coordinated recovery, and run metrics. The targeted compatibility suite passed 49 tests.

    df42fba40d0fb3a3be661ee5fe4dcad6c10593d5
    Owner-verifiable

External research convergence

Later work formalizes the same problem family.

These sources are linked as independent technical comparison—not as evidence of copying, partnership, endorsement, or identical implementation.

CAVA

Formalizes runtime-independent canonical action identities, approval binding, and tamper-evident receipts.

Read the paper ↗

BRA-Audit

Prioritizes limited audits using dependency reach and the time since a trusted checkpoint.

Read the paper ↗

SkillSentry

Compiles procedural skills into enforceable state, dependency, parameter, and completion contracts.

Read the paper ↗

AgentRewind

Aligns agent-context checkpoints with controlled environment state for evidence-preserving recovery.

Read the paper ↗

LEDGER

Separates raw trace records from derived claim-to-evidence graphs and artifact anchors.

Read the paper ↗

Agent Identity Protocol

Explores invocation-bound identity, delegated scope, attenuation, provenance, and result sealing.

Read the Internet-Draft ↗

TraceGrant

Connects trusted requests, admitted evidence, bounded authority, effects, and verified completion through explicit contracts.

Read the paper ↗

CONTINUITY

Formalizes security-context preservation across component boundaries using role-bound receipts and effect-bound permits.

Read the paper ↗

Independent Determinability

Separates authorization, enforcement, dispatch, execution, and independently observed effect.

Read the Internet-Draft ↗

Agent Run Metrics

Defines run and model-invocation accounting for amplification, progress, retries, and termination.

Read the Internet-Draft ↗

SCITT Pre-Run Criteria

Commits evaluation criteria before a run so success conditions cannot be rewritten afterward.

Read the Internet-Draft ↗

Human Delegation Provenance

Separates evidence of human-to-agent delegation from the authority to execute an effect.

Read the Internet-Draft ↗

Claims boundary

What this record says—and what it does not.

Established

Adam Barnes and P.Ai.R. documented and implemented substantial elements of the A.R.K. architecture before the cited 2026 publications.

Supported inference

Independent research convergence strengthens the technical basis for A.R.K.’s design direction.

Not established

No public evidence presently shows that the cited authors accessed A.R.K., derived their work from it, or intentionally denied attribution.

Implemented hardening

Twelve research signals, one coherent A.R.K. control plane.

The signed pipeline now binds each proposed effect to a canonical identity, pre-run criteria, an explicit execution contract, a short-lived operator permit, typed execution stages, native result receipts, and an append-only evidence path. Recovery coordinates Btrfs, policy, plan, memory, model context, and evidence offsets without pretending that local rewind reverses remote effects.