The Critic — Background Audit System

The Critic is a silent second-opinion engine. After each AI generation, it runs a background audit using a separate LLM to detect logic gaps, sycophancy, and shallow reasoning — without interrupting the workflow.

How It Works

graph LR
    G["AI Generation<br/>(Synthesis Model)"] --> C["The Critic<br/>(Background Model)"]
    C --> |"PASS"| P["✓ Silent"]
    C --> |"ISSUE"| W["Badge appears<br/>in chat"]
    C --> |"Click"| D["Expand critique<br/>+ fix suggestion"]

Every time the Brainstorm Reactor completes a generation (recipe step or direct response), the Critic runs a parallel background API call with the full output, user briefing, and last 5 messages as context. The result is a structured audit verdict.

Audit Categories

Critic Confidence Indicator

The Critic shows its own audit confidence alongside the verdict — preventing Meta-Cognitive Surrender (blindly trusting the auditor).

The confidence level is derived server-side from observable response properties (critique length, fix suggestion presence, parse status) — never from the LLM's self-assessment. This is a deliberate architectural decision: empirical research shows that model-reported confidence scores exhibit an inverse correlation (r = −0.40) with actual accuracy.

The Brain-Quietness principle applies: high-confidence audits show nothing extra. Only when the Critic's own analysis quality is questionable does a visual indicator appear.

Two Audit Modes

The Critic operates in two modes, configurable in the System Config sheet:

• System 2 (Logic Audit) — Default. Evaluates logical depth, contextual consistency, and mission alignment. Named after Kahneman's dual-process theory.
• Creative Director — Evaluates novelty, strategic boldness, and creative friction. Detects when the AI plays it safe instead of pushing boundaries.

Critic Onboarding — Taste of Critic

New users encounter the Critic through a one-time contextual invitation called Taste of Critic. The flow:

1. User completes their first recipe (Critic is disabled by default for new users).
2. The system silently runs a background audit anyway — a "probe."
3. If the probe returns a non-PASS result, a full-width invitation card appears showing the finding.
4. The card offers two options: "Always activate" or "Not now."
5. If activated, the Critic is permanently enabled for all future sessions.

This onboarding is non-intrusive — consistent with the Brain-Quietness design philosophy. No pop-ups, no alerts. The invitation only appears when the Critic has something meaningful to show.

Gate Logic

The probe fires only when all conditions are met:

• criticEnabled === false (not yet activated)
• hasSeenCriticInvitation === false (first time)
• tasteOfCriticResult === false (no prior result)
• A valid recipe has completed

If the user dismisses the invitation, hasSeenCriticInvitation is set to true and the probe never fires again.

UI Behavior

The Critic result appears as a compact badge beneath the AI response:

• Collapsed: Small The Critic label with a semantic status pill (teal/amber/rose) and optional confidence indicator.
• Expanded: Click to reveal the critique text and optional fix suggestion.
• Loading: Animated glasses icon with reading dots — no text, no distraction.

The Critic never blocks the UI. All requests are fire-and-forget fetch() calls. If the critic API is unavailable, the result is silently null — no error toast, no user impact.

Cross-Model Badge

When Cross-Model Verification (CMV) is active, the Critic header shows a provider badge:

The Critic [Audit Model ↬ Synthesis Model] ● Pass

This indicates the audit model is evaluating output from a different provider. The badge only appears when the audit provider differs from the synthesis provider — confirming cognitive diversity is active.

## Configuration

In the **System Config** sheet (gear icon → Orchestrator section):

- **Background Model** — The default Critic model when CMV is disabled. Typically a fast, lightweight reasoning model.
- **Cross-Model Verification Toggle** — When enabled, overrides the background model with a verified cross-provider judge (see [CMV Architecture](/docs/engineering/ai-and-models/cross-model-verification)).
- **Critic Mode** — Switch between "System 2" (logic) and "Creative Director" (novelty) audit modes.

---

## Distinction: The Critic vs. De-Risking Critic Audit (EVPI)

The platform offers two complementary verification tools: **The Critic** (under general Orchestrator settings) and the **De-Risking Critic Audit (EVPI)** (under Settings → *Experimentell*):

| Tool | **The Critic** | **De-Risking Critic Audit (EVPI)** |
|---|---|---|
| **Execution** | Asynchronous **2nd Model Call** (Secondary Background Model) | In-Stream **Primary Model Suffix** |
| **Placement** | Bottom audit badge / card below response | Markdown section (`### 🔬 De-Risking & EVPI Audit`) at end of text |
| **Primary Goal** | Detect logic gaps, sycophancy, or hallucinations via a 2nd opinion | Identify the 1 empirical question (EVPI) that reduces risk to <5% |
| **Best For** | High-stakes logic & contract verification (4-eye principle) | Operational project execution & risk reduction steps |