GOFRE Horror Forge

A provenance-first AI horror studio that generates, verifies, stores, and assembles cinematic media with Genblaze and Backblaze B2.

Comment

GOFRE Horror Forge

GOFRE Horror Forge is a production-oriented AI media studio for turning a creative brief into a verified horror production. It connects creative planning, real media generation, durable storage, provenance, human review, and cinematic assembly instead of leaving creators with disconnected prompts and expiring output URLs.

The problem

Independent filmmakers, agencies, and small creative teams can generate images, audio, and video with AI, but the production trail fragments quickly. Prompts disappear, outputs expire, versions become ambiguous, storage is scattered, and teams cannot reliably prove which provider, model, parameters, or source material produced the final asset.

GOFRE Horror Forge solves that operational gap.

The production chain

  1. A creator enters a horror-film brief, tone, rating, duration, and composition strategy.
  2. NERHIA Brain structures the creative intent into an executable production plan.
  3. GOFRE sends an HMAC-signed request to an isolated Genblaze worker.
  4. Genblaze orchestrates the provider call, records provider/model data, and creates canonical provenance manifests.
  5. Generated media, storyboards, manifests, hashes, and production evidence are stored in a private Backblaze B2 bucket.
  6. OpenMontage imports the generated hero asset through a short-lived B2 capability, verifies SHA-256 integrity, strips the temporary URL, and initializes its cinematic pipeline.
  7. GOFRE presents the result, provenance, production state, review boundary, and delivery path.
NERHIA Brain
→ Genblaze
→ OpenAI Images
→ Backblaze B2
→ SHA-256 verification
→ OpenMontage cinematic pipeline
→ GOFRE review and delivery

Real-world utility

GOFRE Horror Forge is designed for creators and teams that need more than a single impressive generation. It provides:

  • repeatable generative-media workflows;
  • durable storage instead of expiring provider links;
  • workspace-scoped production boundaries;
  • provider and model traceability;
  • verifiable SHA-256 integrity;
  • canonical provenance manifests;
  • cinematic handoff and human approval checkpoints;
  • a path from experimental generation to a deliverable media production.

The first creative mission is a short horror film targeted for Halloween, October 31, 2026. The hackathon build demonstrates the complete golden path with a hero frame and verified production packet.

How Backblaze B2 is used

Backblaze B2 is the durable media vault, not a decorative upload destination.

The Genblaze worker stores:

  • structured storyboards;
  • generated hero media;
  • canonical Genblaze provenance manifests;
  • content hashes and production evidence;
  • workspace and project-scoped object paths.

Buckets remain private. The browser receives only a short-lived signed preview URL. OpenMontage consumes that capability once, validates the trusted Backblaze host, bounds the download size, checks the declared SHA-256, writes the media atomically into its production workspace, and removes the temporary capability before persisting any JSON record.

This makes B2 the durable bridge between generation, verification, review, and cinematic assembly.

How Genblaze is used

GOFRE uses Genblaze as the generative-media orchestration and provenance layer.

The isolated Python worker uses Genblaze primitives to:

  • create a production run;
  • call the configured image provider;
  • record provider, model, prompt, and parameters;
  • attach generated assets;
  • write canonical manifests;
  • verify the canonical manifest hash;
  • send assets to the Backblaze B2 storage sink;
  • return a bounded provenance envelope to GOFRE.

GOFRE does not recreate the SDK in application code. It communicates with the worker through an explicit HMAC-signed server-to-server boundary.

Production readiness

The tournament stack includes:

  • separate Cloud Run services for GOFRE, Genblaze, and OpenMontage;
  • Secret Manager references for OpenAI, B2, and HMAC secrets;
  • fail-closed gates requiring real media, B2 persistence, and verified OpenMontage import;
  • private object storage and short-lived previews;
  • timeout, redirect, content-type, maximum-size, trusted-host, and hash checks;
  • persistent OpenMontage project storage;
  • workspace authorization and role checks;
  • idempotent media import;
  • responsive UI with an explicit layer hierarchy;
  • automated Node, Python, Docker, Firebase route, and security regression tests;
  • a manual, confirmation-gated tournament deployment workflow;
  • smoke checks that validate infrastructure without triggering a paid generation.

Security design

  • Browser clients never receive B2 or provider credentials.
  • B2 application keys are limited to the tournament bucket.
  • Worker endpoints require HMAC signatures.
  • Genblaze and OpenMontage remain disabled until explicitly configured.
  • Tournament mode rejects storyboard-only or memory-only fallback results.
  • Signed B2 preview URLs are never written to provenance, handoff, checkpoint, or asset seed files.
  • OpenMontage remains a separately deployed AGPL-3.0 service with its public source URL exposed.

What was built during the hackathon

GOFRE existed before the submission period as a broader AI content-operations platform. During this hackathon it was substantially extended with:

  • a dedicated Horror Forge product experience;
  • a Genblaze production and provenance worker;
  • real OpenAI image generation through Genblaze;
  • a Backblaze B2 durable storage sink;
  • canonical storyboard and media manifests;
  • short-lived B2 preview delivery;
  • verified B2-to-OpenMontage media ingestion;
  • OpenMontage cinematic production handoff;
  • persistent production storage;
  • fail-closed tournament controls;
  • responsive layer and overlap corrections;
  • deployment automation, runbooks, and regression tests.

Challenges

The difficult problem was not generating a horror image. It was building a trustworthy production boundary between a multi-tenant Node.js platform, a Python generative-media SDK, an external model provider, private object storage, and an agentic cinematic pipeline without exposing credentials or losing provenance.

A second challenge was request duration. Firebase Hosting has a short dynamic-request ceiling, while real media generation can take longer. The deployed architecture keeps Firebase for the public interface and fast API calls, while the production request uses a validated direct Cloud Run origin published through a non-secret runtime configuration.

What is next

  • execute the live tournament deployment with bucket-scoped secrets;
  • generate and record the first production evidence set;
  • complete OpenMontage edit, compose, review, and publish checkpoints;
  • render the first MP4;
  • expand from the hero production into the full Halloween short film;
  • package the same governed media engine for agencies, branded content, education, and entertainment teams.

Built With

Share this project:

Updates