AI
Explore the Stack

EphUX / Guardian

A local, user-owned overlay that detects reasoning-integrity patterns in AI chat interfaces.

EXPERIMENTALReasoning-integrity browser layer

The Problem

Web AI assistants can subtly sycophant, drift from the user's instructions, deviate from constraints, or assert things they cannot support—while remaining fluent and useful. Users have no layer that watches the assistant's behavior independently of the assistant itself.

The Approach

EphUX is a Manifest V3 browser extension that runs local heuristic detectors (sycophancy, drift, instruction deviation, unsupported assertion, identity inflation) over the chat surface of Claude, ChatGPT, Gemini, Perplexity, and Copilot. When a reasoning-integrity threshold is crossed, it surfaces a quiet category-specific overlay with Dismiss / Inspect / Recommend actions. Local-first, zero-retention by default.

What Exists Today

  • ✓Manifest V3 extension prototype
  • ✓Working local heuristic detectors: sycophancy, drift, instruction deviation, unsupported assertion, identity inflation
  • ✓DOM capture adapters for major AI chat interfaces
  • ✓EphUX overlay with Dismiss / Inspect / Recommend
  • ✓Design handoff document specifies Level-4 rebuild flow and additional detectors

//What It Does Not Prove

What this project does not prove.

  • ✗Local heuristic detection only
  • ✗Cross-validation network calls stubbed (future work)
  • ✗Level-4 rebuild consent flow stubbed (no destructive action ships)
  • ✗Additional detectors (false consensus, over-conservatism, contradiction suppression) not wired
  • ✗Extension is unpacked; not independently audited

Evidence

Local heuristic detection of reasoning-integrity patterns in AI chat interfaces

ephux

PROTOTYPE
Artifact: EphUX extension
Version: v0.1.1
Environment: Local development (unpacked extension)
Date: 2026-08

Result:

Manifest V3 extension implements local heuristic detectors for sycophancy, drift, instruction deviation, unsupported assertion, and identity inflation, with an overlay offering Dismiss / Inspect / Recommend. DOM capture adapters exist for major AI chat interfaces.

//Verification Boundary

Local heuristic detection only. Cross-validation network calls stubbed. Level-4 rebuild consent flow stubbed. Detectors for false consensus, over-conservatism, and contradiction suppression not wired. Not independently audited.

Defensive companion detection of malicious or spoofed AI chat sessions

ephux

PROTOTYPE
Artifact: EphUX Guardian Shield extension
Version: v0.1.0
Environment: Local development (unpacked extension)
Date: 2026-08

Result:

Manifest V3 extension uses local zero-trust heuristics to flag phishing, scammers, and interlopers using spoofed or jailbroken AI interfaces. Same reasoning-integrity detectors applied to untrusted chat contexts.

//Verification Boundary

Local heuristic detection only. Does not guarantee detection of all phishing attempts. Cross-validation and live Level-4 session salvage out of scope. Not production hardened.

Activation kernel enforcing Activation Doctrine (purpose anchoring, path integrity, structured continuity)

ephux

PROTOTYPE
Artifact: EphUX Activation Kernel worker
Version: prototype
Environment: Cloudflare Workers (deployed prototype)
Date: 2026-08

Result:

Worker implements an activation layer: purpose + context in, structured JSON out (decision: Supported/Unsupported/Indeterminate, integrity classification, flags, receptor event suggestions, dither spec). Controller pre-filter detects classic prompt injection and exfiltration before model call.

//Verification Boundary

Uses lightweight Workers AI model. Controller fallback handles unparsable output. Deployment status not independently confirmed. Not a full AGI safety system.

Guardian Intake GatewayGuardian Audit ServiceSymIDSessionGlyph

Source Artifacts

  • 📄GUARDIAN_DESIGN_HANDOFF
  • 📄EphUX extension (content.js, manifest.json, README)
  • 📄EphUX Guardian Shield extension
  • 📄EphUX Activation Kernel worker