// engineering tools / shared boundaries
One Ecosystem. Clear Roles. エコシステム // 2026
Explore possibilities. Keep evidence. Control consequential actions. Howl brings specialized tools together around a common rule: intent is not authority. Each component owns a distinct problem so creative work, factual claims, orchestration and execution do not silently become the same thing.
Engineering orchestration
Creative development
Controlled experiments
Independent verification
Writing and provenance
The Howl CLI installs and maintains registered components. It does not orchestrate AI work.
SECTION // 01 Proposals Do Not Authorize Effects
ECO-MOD-ARCHITECTUREArchitecture / conceptual relationships. The arrows below describe responsibilities, not a claim that every pair has a native runtime adapter.
Human objective → HowlPlane
Planning, agent routing, evidence and policy evaluation.
HowlCreate ↔ HowlDream
Intentional invention and controlled speculative experiments. Candidates return for evaluation.
HowlFrame → HowlProof
What is built is challenged before it is trusted. HowlProof produces findings and evidence; it never repairs the artifact it evaluates.
HowlProof → HowlPlane → the builder
Remediation is routed by the control plane to whoever owns the artifact, and HowlProof independently re-verifies the repair rather than accepting a claim that one was made.
HowlProof → HowlChangeOps
Evidence informs promotion. The verdict is advisory: HowlProof holds no authority to release, and HowlChangeOps owns approval and rollback.
HowlFrame → HowlChangeOps
Capability-bounded runtime and deterministic policy underpin governed change execution.
HowlRelay carries work state and handoffs. HowlWriter communicates reviewed knowledge. HowlDream and HowlProof both export advisory artifacts only; neither can authorize an effect.
SECTION // 02 Core Installation & Optional Companions
ECO-MOD-COMPONENTSThe standard installer manifest includes HowlFrame, HowlChangeOps, HowlPlane and HowlWriter. Core membership describes installation scope, not production maturity. HowlCreate, HowlDream, HowlProof, HowlRelay and reference applications are separately installed companions.
Plans and routes engineering work, coordinates agents, records evidence, and applies authority gates. Start here for governed AI engineering.
AI-native language, compiler, and capability-bounded VM. It verifies program and policy constraints; it is not a general factual-truth oracle.
Collects trusted evidence, evaluates HowlFrame policy, requires applicable human approvals, and executes bounded changes.
Writing, citation, provenance, and review workflows. Model-backed roles require explicit integration; standalone deterministic tools remain useful.
Deliberate invention through reframing, creative search, mutation, challenge, synthesis, and concept lineage.
Independent red-team QA. Challenges an artifact across functional, security, reliability, AI and operator adversaries, reproduces what it finds, and issues a verdict backed by an evidence bundle. It does not repair what it judges, and holds no authority to release. Experimental, milestone one.
Controlled divergent exploration. Explores possibilities that may be incorrect, measures known failures, and checks speculative ideas before downstream evaluation.
Collects observable work state, journals, evidence and handoffs so engineering can continue asynchronously without worker surveillance.
A deterministic task board that exercises HowlFrame browser compilation, backend state transitions, and persistent storage. Reference application.
A notes application used to expose concrete HowlFrame language, runtime and native-storage gaps. Reference application.
A Discord operations adapter with deterministic HowlFrame policy and bounded actions; not a core orchestration component. Pages site unavailable at the September 2026 audit.
Two kinds of exploration
HowlCreate invents intentionally: it searches for useful concepts through deliberate creative operators and convergence.
HowlDream explores speculatively: it treats outputs as experimental material, compares conditions with a baseline, induces failures, and preserves unresolved claims after checks.
HowlDream intentionally explores possibilities that may be incorrect, measures known model failures, and verifies speculative ideas within explicit limits before downstream evaluation. A candidate can deserve investigation while remaining unverified.
[ VISIT HOWLDREAM ]SECTION // 03 Start With One Component
ECO-MOD-STARTNewcomer paths.
Run an offline experiment
Clone HowlDream, install in a virtual environment, and run its deterministic sample and Dreambench. No API credentials required.
Coordinate engineering
Read HowlPlane's workflow and authority model. Add other tools when the work needs them.
Install registered tools
Howl resolves a tested release manifest and manages installation lifecycle. Experimental companions are not silently added.
SECTION // 04 Names and Maturity Are Evidence-Based
ECO-MOD-NAMINGChangeOps is the compatibility name for HowlChangeOps, not another active component. The former AI knowledge library now lives within HowlPlane. HowlBoard, HowlNotes and HowlBot are dogfood applications, not required platform services. Experimental does not mean deprecated; repository existence alone does not mean production-ready.
HowlProof and HowlChangeOps are adjacent and easily confused. HowlProof asks whether an artifact can be demonstrated untrustworthy and produces evidence; HowlChangeOps asks whether a change is authorized to happen and can be undone. An approval over an unverified artifact is a governed mistake, and a verdict with no gate behind it is an ungoverned success. A future HowlGuard, watching whether the ecosystem's security posture has changed enough to require re-evaluation, is a documented boundary and is not built.
Audit: 11 September 2026, based on current repository documentation, code and Pages metadata. Read the integration audit · Installer architecture