Evidence · integration

Integration support and boundaries

A release-specific matrix separates the tested 5D core from integrations that need additional wiring and review or execution capabilities owned by the host.

Evidence strength Verified Evidence form Shipped code Version 2.0.0 Author Loren Angoni · 5D product owner Dates Created 2026-09-27 · Updated 2026-10-06 · Verified 2026-10-06

Scope

PyPI release 0.7.0, exact package fivedrisk-0.7.0-py3-none-any.whl, SHA-256 790edf0e09c81e89b232ab77589d6df8551a08ed93d0bef3644f5f1a4043c22c, plus the clearly labelled action-flow walkthrough.

One release, several materially different contracts

The Python verdict socket, function decorator, JSON-lines gateway, Claude Code hooks, LangGraph node, and five framework shims do not expose identical enforcement or approval semantics. The matrix states the stop shape and known bypass boundary for each path.

Verified core means the named behavior was exercised from the exact wheel without depending on a live framework SDK. Partial integration means an entry point exists but stronger end-to-end framework evidence is absent.

Review and semantic evidence do not become permission

The public package does not ship a universal reviewer service or a semantic observer. A host may compose Jev or another contextual model, but the observer remains evidence and cannot clear an explicit deterministic block.

The has_approval_channel flag is a caller assertion. It is not proof that an authenticated reviewer approved the exact attempt.

Runtime entry-point support matrix

This matrix separates tested core behavior from integrations that need additional wiring and capabilities that remain with the host. It covers package fivedrisk-0.7.0-py3-none-any.whl, version 0.7.0, SHA-256 790edf0e09c81e89b232ab77589d6df8551a08ed93d0bef3644f5f1a4043c22c.

verified-coreThe named behavior was exercised from the exact public wheel without a framework SDK dependency.

partial-integrationAn integration entry point exists and has bounded tests, but pinned real-SDK end-to-end interception proof is missing.

reference-compositionA documented composition or helper exists outside the verified core and still requires host integration and proof.

out-of-scopeThe host application or another system must provide this capability.

verified-core

Python verdict socket: to_verdict()

Contract: Accept a tool name and JSON-object input, then return a band, sentinel, rationale, scores, and decision identifier.

Stop shape: RED or invalid input returns a BLOCK sentinel. ORANGE returns APPROVE and remains non-executable by default.

Approval: No reviewer channel is supplied. APPROVE is a hold signal for the host.

Evidence and limits

Evidence

  • Exact-wheel adapter tests passed for band mapping, malformed input, error handling, and fail-closed defaults.

Limits

  • The caller must place this check before every relevant side effect and honor the returned sentinel.
  • Caller-supplied identity and context are not authenticated by this socket.
verified-core

Python @gate decorator

Contract: Wrap a Python function so 5D evaluates declared tool and action context before the function executes.

Stop shape: GREEN and YELLOW execute. ORANGE and RED raise or invoke the configured on_block handler.

Approval: The decorator does not run a human-review workflow. A host-owned workflow must resolve approval-required cases.

Evidence and limits

Evidence

  • The published wheel contains fivedrisk/hooks.py and its hook test suite.

Limits

  • Only wrapped call paths are governed.
  • Function execution, idempotency, downstream outcome, and recovery remain host responsibilities.
verified-core

JSON-lines gateway: stdio and one-shot

Contract: Let a non-Python host submit a JSON action envelope to the Python engine and receive a versioned JSON decision.

Stop shape: Malformed requests return typed errors; detected input threats return RED with blocked=true; scored requests return the policy band.

Approval: The gateway returns decision evidence. It does not supply a reviewer or execute the external tool.

Evidence and limits

Evidence

  • Exact-wheel gateway and TypeScript wire-contract tests passed for request, response, handshake, correlation, and malformed-input behavior.

Limits

  • A language client must still translate the response into a local stop before execution.
  • Destination policy configured through Python hooks is not automatically available through this gateway path.
partial-integration

Claude Code PreToolUse and PostToolUse hooks

Contract: Translate a structured pre-tool payload into allow, ask, or deny; scan supported post-tool output fields after execution.

Stop shape: EXECUTE maps to allow, APPROVE maps to ask, and BLOCK or an unknown sentinel maps to deny.

Approval: The host's native ask path is used. 5D does not authenticate the person or prove that approval occurred.

Evidence and limits

Evidence

  • Exact-package example payload tests passed, including fail-closed egress-scan failures and alternate output fields.

Limits

  • No version-pinned live Claude Code executable integration was exercised in this receipt.
  • PostToolUse can halt further processing but cannot undo a tool action that already occurred.
partial-integration

LangGraph gate node and route_by_band()

Contract: Add a scoring node that writes 5D fields into graph state and routes by the returned band.

Stop shape: Missing band fails closed to red; the graph owner maps green, yellow, orange, and red edges.

Approval: The graph owner must provide and secure any human-review node.

Evidence and limits

Evidence

  • The exact public wheel contains the node and routing implementation.

Limits

  • No pinned live LangGraph SDK end-to-end interception test was included in the exact-wheel receipt.
  • Alternate graph edges or direct executor calls can bypass the node.
partial-integration

CrewAI, Google ADK, OpenAI Agents, Pydantic AI, and Microsoft Agent Framework shims

Contract: Translate the shared verdict socket into each framework's documented pre-tool stop shape.

Stop shape: Each shim stops BLOCK and, by default, also stops APPROVE when no approval channel is asserted.

Approval: has_approval_channel=true is a caller assertion, not proof of review or reviewer authority.

Evidence and limits

Evidence

  • Exact-wheel framework-adapter tests passed against in-repository fakes for all five translation shapes.

Limits

  • The shims were not validated end to end against pinned live installs of all five framework SDKs.
  • Coverage is limited to the documented hook surface and does not include alternate or direct tool paths.
reference-composition

5D setup skill

Contract: Guide a coding agent through discovery, policy scaffolding, integration, and verification steps.

Stop shape: The skill proposes code and tests. It does not enforce a runtime decision itself.

Approval: A person must review the generated integration and its coverage evidence.

Evidence and limits

Evidence

  • The skill source and SKILL.md are available in the public GitHub repository.
  • The 0.7.0 wheel includes helper scripts but omits the skill's SKILL.md and reference files.

Limits

  • The skill is not bundled as a complete installable skill in the 0.7.0 PyPI wheel.
  • A successful skill run is not certification and does not prove coverage of alternate execution paths.
reference-composition

Jev or host-selected contextual LLM observers

Contract: Supply bounded semantic evidence to a host-defined policy branch when deterministic context alone is insufficient.

Stop shape: Observer evidence may tighten or hold a decision but cannot clear an explicit deterministic block.

Approval: An observer is evidence, not permission and not a reviewer identity.

Evidence and limits

Evidence

  • The architecture guide shows an example composition and an explicit authority boundary.

Limits

  • No observer service, prompt, model credential, or model call ships in the fivedrisk wheel.
  • The integrating host must define timeouts, malformed-output handling, data minimisation, and disagreement policy.
out-of-scope

Authenticated human review and execution authority

Contract: Bind an authorized reviewer decision to the exact action attempt and return it as evidence to policy.

Stop shape: Missing, expired, malformed, or unavailable required review must remain held and must not become permission.

Approval: The host owns authentication, reviewer authority, freshness, anti-replay, and attempt binding.

Evidence and limits

Evidence

  • The public architecture documents the boundary and does not claim a bundled reviewer service.

Limits

  • 5D 0.7.0 does not provide a universal reviewer UI, identity provider, or approval service.
  • A reviewer decision does not prove that the action is correct or safe.
Claim register

What this page says, and how far it goes.

Verified

The tested artifact is exact

The integration receipt is bound to the public 0.7.0 wheel and its SHA-256 digest rather than to an unpinned working tree.

Shipped code · Evidence: pypi-wheel-070

Limit: The receipt proves the exercised wheel behavior, not every possible host integration.

Verified

Core entry points return a disposition before host execution

The verdict socket, decorator, and gateway expose pre-action decision paths with fail-closed malformed-input or unknown-band behavior.

Shipped code · Evidence: pypi-wheel-070

Limit: The host must place the entry point on every relevant execution path and honor the result.

Verified

Framework shims remain partial integrations

Five framework translation shims exist and were tested against in-repository fakes, but do not yet have pinned live-SDK end-to-end proof for every framework.

Shipped code · Evidence: pypi-wheel-070

Limit: Adapter availability is not a claim of production support or complete interception coverage.

Verified

The setup skill is repository-hosted

The complete SKILL.md is available from GitHub, while the 0.7.0 wheel omits the complete skill instructions and reference set.

Shipped code · Evidence: pypi-wheel-070, setup-skill-source

Limit: Installing the Python package alone does not install the complete coding-agent skill.

Design target

The complete cascade is an integration contract

The action-flow walkthrough describes the desired end-to-end contract while stating clearly that no single package entry point implements every step.

Framework mapping · Evidence: website-action-flow

Limit: The illustration does not prove a live model observer, reviewer service, executor, or complete deployment.

Limits

What this record does not establish.

Sources

Versioned references.

What this page represents

The evidence and demonstrations shown here do not run a live engine, reviewer service, executor, or backend, and do not establish a supported integration contract.