Evidence Guard FAIL-CLOSED / OFFLINE / SYNTHETIC

OpenAI Build Week · Developer Tools

Nexus Vector Evidence Guard

Never guess through a high-risk deployment.

A fail-closed safety layer that verifies autonomous-agent updates, blocks unreviewed drift before mutation, preserves evidence around ambiguous writes, and authorizes only exact verified rollback.

INTERACTIVE SYNTHETIC DEMO REAL OFFLINE PYTHON ENGINE

This hosted experience is an interactive browser simulation using synthetic data.
The repository contains the real offline Python implementation and deterministic test suite.

The problem

An update can be more dangerous than the agent.

A high-risk autonomous agent may become unsafe if an update is only partially applied, if installed files changed after review, or if an uncertain operation is retried blindly.

Unprotected path

  1. Reviewed agent
  2. Incomplete or uncertain update
  3. Mixed or unknown runtime state
  4. Unsafe restart risk

Evidence Guard path

  1. Reviewed source
  2. Baseline verification
  3. Verified backup and controlled apply
  4. Evidence-bound recovery

Core invariant

UNKNOWN RESULT ≠ RETRY

UNKNOWN RESULT =
STOP + EVIDENCE + CONTROLLED RECOVERY

Interactive safety console

Test each failure boundary.

Choose a fixed synthetic scenario. The browser runs a deterministic state sequence; it does not accept commands, paths, or live data.

SYNTHETIC BROWSER CONSOLE NO COMMAND INPUT

Ready. Select one of four fixed synthetic scenarios.

Scenario map

Four outcomes. One fail-closed rule.

SCENARIO 01

Safe update completed

Risk
The agent must update without entering a mixed-version state.
Evidence Guard action
Verifies source, baseline, backup, apply, and final identities.
Outcome
Controlled deployment completes with rollback evidence.
CLEAN_DEPLOYMENT_PASS
SCENARIO 02

Unsafe update blocked before mutation

Risk
The installed agent changed after review.
Evidence Guard action
Detects destination drift before any write.
Outcome
Mutation count remains zero; review is required.
DESTINATION_DRIFT_BLOCKED_PRE_MUTATION
SCENARIO 03

Ambiguous write stopped

Risk
The operating system cannot prove whether replacement succeeded.
Evidence Guard action
Prohibits blind retry and preserves mutation evidence.
Outcome
Automation stops for controlled manual recovery.
AMBIGUOUS_MUTATION_REQUIRES_MANUAL_RECOVERY
SCENARIO 04

Exact verified rollback completed

Risk
A rollback could use the wrong backup or run twice.
Evidence Guard action
Binds authorization to the exact apply and backup.
Outcome
The baseline is restored and duplicate rollback is blocked.
VERIFIED_ROLLBACK_PASS

Portable evidence

Keep the result, not sensitive context.

Export the scenarios completed in this browser session and the current inspector values. The generated JSON contains synthetic status data only.

The timestamp is explicitly marked browser-generated. No paths, credentials, production data, or live identifiers are included.

Real offline engine

Browser semantics, repository implementation.

The browser experience demonstrates the safety state machine and scenario outcomes. The repository contains the real offline Python engine. This page does not execute PowerShell, the operating system, or the Python implementation.

PASS CLEAN_DEPLOYMENT_PASS
PASS DESTINATION_DRIFT_BLOCKED_PRE_MUTATION
PASS AMBIGUOUS_MUTATION_REQUIRES_MANUAL_RECOVERY
PASS VERIFIED_ROLLBACK_PASS
RESULT: PASS (4/4)
Windows powershell -ExecutionPolicy Bypass -File .\demo\run_demo.ps1
Cross-platform python demo/run_demo.py