V Guard

Inspiration

Image generation makes it easy to create more files, but new is not the same as better. A polished live image can be replaced by a generic layout render simply because the replacement exists. I built V Guard to ask a more disciplined question: does this candidate fit the brand, placement, audience, and truth of the product better than the current image?

What it does

V Guard presents three public-safe comparisons:

1. Protect a strong existing image from a weaker layout draft.
2. Promote purpose-built campaign art while preserving a real screenshot as product proof.
3. Reject abstract decoration when a placement requires evidence of a working product.

For each case, V Guard returns a policy-owned verdict, five visual-quality scores, concrete reasons, mistakes avoided, a next direction, and a public-safe receipt containing the analysis source and both asset hashes. Accessibility is presented as a design heuristic, not a guarantee.

How it works

The browser sends only an allowlisted sample ID to a vinext server route. The route resolves two local public samples and their fixed FVE policy verdict. When OpenAI API capacity is available, the configured GPT-5.6 analysis lane contributes structured scores and explanations. It cannot override local policy.

Cloudflare D1 is used when available for cached reviews and best-effort daily usage accounting. If the model request cannot run, V Guard returns a deterministic, sample-specific policy review and labels that source honestly.

How I used Codex and GPT-5.6

Codex was my implementation partner throughout Build Week. It helped isolate a safe public slice from my private TheFAYTH Visual Empire system, build and harden the interface and API, separate policy verdicts from model explanations, create and inspect the V Guard identity, repair quality-gate tests, prepare the Sites deployment, generate the real-screen demo video, and preserve public-safe receipts.

GPT-5.6 is the optional structured analysis layer. Human-authored FVE policy remains responsible for the verdict. This separation lets model vision add detail without silently taking authority over the design decision.

Challenges

The central challenge was preserving human and product judgment while still using model vision effectively. A second challenge was making fallback behavior reliable without presenting it as live AI. Every result identifies its actual source.

Accomplishments

• A working three-case visual review flow with fixed public samples.
• Policy verdicts that model output cannot reverse.
• Structured scoring across brand fit, visual quality, specificity, accessibility heuristic, and trust.
• Public-safe receipts with hashes and source disclosure.
• Deterministic fallback behavior when live analysis is unavailable.
• A narrow public/private boundary that demonstrates the product without exposing the private FVE engine.

What I learned

Visual quality control is not a single aesthetic score. It depends on placement purpose, baseline quality, specificity, accessibility, trust, provenance, and whether the image is honest evidence. Model vision becomes more useful when it operates inside an explicit, auditable decision policy.

Privacy and safety

The public demo accepts no uploads, arbitrary URLs, free-form prompts, credentials, or publishing instructions. It contains no WordPress or COAI write endpoint. Private FVE code, customer data, target registries, rollback internals, and production credentials remain outside the contest repository.

Try V Guard

• Live demo
• Public contest repository
• Two-minute demo video

Built by Faith Cheltenham. Owned by XXYYZZ Society LLC.

Built With

• cloudflare-d1
• cloudflare-workers
• codex
• drizzle-orm
• gpt-5.6
• next.js
• openai-api
• react
• typescript