Media Generator Dashboard (Demo)

Total Assets

in pipeline

Generating

ChatGPT-Web sim

Consensus Pass

vision bridge ≤85px

HALT / Review

drift >85px

Placed

ready to deliver

What it does

Spec-driven asset factory. Each asset moves through a 4-stage pipeline: GPT-4o renders the scene art, Gemini re-extracts geometric anchor coordinates, and a consensus engine reconciles both models. Drift beyond 85px halts the run — fail-closed.

Inputs

Asset spec JSON (canvas size, layout zones, scene brief, forbidden elements). Style palette tokens. Gold reference images for visual style matching. Operator queue of assets to process.

Outputs

Final composite PNG at 1024×1536 with deterministic callout geometry. asset.runtime.json with verified anchor coordinates. Gate reports for every stage. Structured failure-learning summaries when halted.

Stage 1

Generate

ChatGPT-Web renders blueprint scene art via Playwright automation (production; simulated here)

›

Stage 2

Primary Vision

GPT-4o extracts anchor centroids from the rendered image

›

Stage 3

Secondary Vision

Gemini API independently locates the same semantic targets

›

Stage 4

Consensus

Euclidean drift check. ≤85px PASS, >85px HALT

Asset Pipeline Board — drag or click “Advance” to move cards

⚖ Vision Consensus Engine — 85px Fail-Closed Rule

Select an asset in the Verify stage and run consensus to simulate GPT-4o vs Gemini geometric agreement. Drift > 85px triggers HALT. Consensus logic is a faithful deterministic port of scripts/validate-consensus.mjs. Production pipeline uses ChatGPT-Web via Playwright automation; all model calls here are simulated.

ChatGPT / GPT-4o

waiting...

↔

Gemini Vision

waiting...

⇨

Drift Analysis

Run a check to see results

◯

Idle — no consensus run yet

Select an asset and click Run Consensus Check

The 85px Fail-Closed Threshold — Why it matters

85px Euclidean distance (√(Δx² + Δy²)) between GPT-4o centroid and Gemini centroid for the same anchor target. ≤85px PASS — Models agree. Primary (GPT-4o) math is written to asset.runtime.json. Asset advances to Placed. >85px HALT — Models are guessing. The pipeline exits code 1. No runtime file is written. Asset enters Review for manual repair.

85px AGREE EDGE HALT

The threshold is intentionally tight. At >85px a model is placing its anchor in the wrong anatomical region of a 1024×1536 canvas. Using disagreed math would silently produce misaligned callout arrows in the final poster.

Demo · All asset IDs, specs, and coordinates are fictional sample data. No client assets, no real images, no secrets. Pipeline logic mirrors scripts/validate-consensus.mjs from the real factory.