Skip to main content

Live Governed Proof

live-governed-proof is a pack of live proof scripts for real model-backed governed execution. It is the example set to read when you want evidence that AgentXchain is governing actual model output instead of only mocks.

What It Is

  • Category: live proof example
  • Surface: runner-interface, api_proxy, MCP, and continuous mixed-runtime proof scripts
  • Source: examples/live-governed-proof

Workflow Or Proof Shape

This directory contains two categories of live proof:

Single-Turn and Multi-Provider Proofs

  • run-live-turn.mjs — one real api_proxy governed turn
  • run-multi-provider-proof.mjs — OpenAI → Google → Anthropic governed flow
  • run-proposed-authority-proof.mjs — proposal-authority proof
  • run-escalation-recovery-proof.mjs — blocked-state recovery proof
  • run-mcp-real-model-proof.mjs — real-model MCP proof

Checkpoint Handoff Proof (BUG-23)

  • run-checkpoint-handoff-proof.mjs — proves that run --continuous --auto-checkpoint emits a git checkpoint commit after every accepted authoritative turn, and that no "clean baseline" errors surface between role handoffs
  • Full case study: Checkpoint Handoff Proof

This proof exercises the complete pm → dev → qa role progression with all roles using local_cli authoritative runtimes. Each role produces working-tree changes that must be checkpointed before the next authoritative turn can start. The proof verifies:

  • 3 checkpoint: <turn_id> (role=<role>, phase=<phase>) git commits in the log
  • 3 turn_checkpointed events in events.jsonl with SHA references
  • zero "clean baseline" errors in CLI output
  • zero manual checkpoint-turn --turn prompts

Verified evidence (2026-04-18):

MetricValue
Runtimelocal_cli (mock changingAgent) for all roles
Roles checkpointedpm (planning), dev (implementation), qa (qa)
Session IDcont-d49f1fc5
Checkpoint commits3
Checkpoint events3
Clean baseline errors0
Session statuscompleted
Evidence artifactcheckpoint-handoff-proof.latest.json

Run it: node examples/live-governed-proof/run-checkpoint-handoff-proof.mjs --json --output examples/live-governed-proof/evidence/checkpoint-handoff-proof.latest.json

Pass --runtime-command <cmd> to use a real local runtime (e.g., codex exec --dangerously-bypass-approvals-and-sandbox) instead of the mock.

Continuous Mixed-Runtime Proof

  • run-continuous-mixed-proof.mjsfull continuous session with mixed local + remote runtimes
  • run-continuous-3run-proof.mjs — dated case-study harness for three unattended governed runs in one continuous session

This is the strongest proof in the repo. It exercises the real agentxchain run --continuous CLI surface with:

  • local authoring roles (PM, Dev, Director) satisfying requires_files gates via local_cli
  • a real Anthropic-backed api_proxy QA role (review_only) validating and requesting completion
  • vision-driven intake seeding from a project-local VISION.md
  • continuous session lifecycle: start → governed run → intake resolution → session completion

Verified evidence — continuous mixed-runtime (2026-04-18):

MetricValue
ModelClaude Haiku 4.5 (claude-haiku-4-5-20251001)
Runtimeapi_proxy (Anthropic) for QA, local_cli for authoring roles
Session IDcont-8ff8f453
Runs completed1
Total spend$0.008
CLI version2.134.1
Review artifact.agentxchain/reviews/turn_*-qa-review.md
Session statuscompleted
Evidence artifactcontinuous-mixed-proof.latest.json

Verified evidence — continuous 3-run (2026-04-18):

MetricValue
Runs completed3
Total spend$0.023
CLI version2.134.1
Session statuscompleted
Evidence artifactcontinuous-3run-proof.latest.json

This proof confirms that AgentXchain can dispatch to real models, accept structured turn results, track budget, and complete a governed continuous session — not just pass mock-based tests.

For the public 3-run proof page with the exact command, spend, session snapshots, and git-log evidence, read Live Continuous 3-Run Proof.

How AgentXchain Governs This Example

AgentXchain governs these proofs by:

  • writing real dispatch bundles
  • validating real turn-result JSON
  • enforcing gate progression and completion rules
  • recording governed artifacts even when the model is live
  • tracking continuous session state, budget, and intake provenance

This is proof infrastructure, not a sample product.

Run It

Single-turn and multi-provider proofs:

node examples/live-governed-proof/run-live-turn.mjs --json
node examples/live-governed-proof/run-multi-provider-proof.mjs --json
node examples/live-governed-proof/run-proposed-authority-proof.mjs --json
node examples/live-governed-proof/run-escalation-recovery-proof.mjs --json
node examples/live-governed-proof/run-mcp-real-model-proof.mjs --json

Checkpoint handoff proof (BUG-23):

node examples/live-governed-proof/run-checkpoint-handoff-proof.mjs --json
node examples/live-governed-proof/run-checkpoint-handoff-proof.mjs --runtime-command "codex exec --dangerously-bypass-approvals-and-sandbox"

No credentials required. Self-contained with mock agent by default.

Continuous mixed-runtime proof:

node examples/live-governed-proof/run-continuous-mixed-proof.mjs --json --output examples/live-governed-proof/evidence/continuous-mixed-proof.latest.json
node examples/live-governed-proof/run-continuous-3run-proof.mjs --json --output examples/live-governed-proof/evidence/continuous-3run-proof.latest.json

Credential requirements: ANTHROPIC_API_KEY is required for the continuous proof and several single-turn scripts. OPENAI_API_KEY and GOOGLE_API_KEY are needed for the multi-provider proof. Scripts skip cleanly when their required credentials are absent.

Key Takeaways

  • This is the clearest live-proof surface in the repo.
  • The continuous mixed-runtime proof is the strongest evidence that governed lights-out execution works with real models — not just mocks.
  • review_only remote roles can validate and review but cannot author gate files. Local authoring roles satisfy requires_files gates. This is the truthful mixed-runtime shape today.
  • Use these proofs to validate runtime truth before making product claims.