Most index projects die in the business case, not the build. Someone estimates a year of migration, a rack of new storage, and a re-training program, and the whole thing gets deferred to a budget cycle that never comes. The four week pilot exists to skip that argument. Instead of a proposal about what an index might do, you get a working one on your own archival data, and a findings report in your own numbers, before anyone signs anything larger.
The rules are deliberately small. A media index pilot runs on archival data you choose, on a single production or facility, single-tenant in Fortify's cloud. Nothing touches your live pipelines. At wrap, the pilot data is deleted and the deletion is verified, so the only thing you keep is the graph you built and the report that came out of it. Four weeks, four moves.
The pilot starts with something you already have, not something you have to make. That is usually a ShotGrid export, and sometimes a shooting script. Either one becomes the seed of a context brain.
If it is a ShotGrid export, Fabric joins it into one governed knowledge graph. Fabric natively reads ShotGrid, FileMaker and Airtable, with 300+ more sources available through the connector catalog, and maps what it finds onto the MovieLabs OMC model: 340 fields across 14 entity types. The result is not a spreadsheet with search on top. It is a graph that knows a shot from a scene from a rights holder, enforces permissions inside the query, and flags a current version against a superseded one. Embeddings are computed inside the deployment environment, so the retrieval layer does not make external AI calls to understand your material.
If the seed is a script instead, Codex parses the shooting script into a 12-category breakdown, 9 of those categories auto-tagged, with cast rollups. There is no LLM in the parse, so it is deterministic and diffable: run it on two revisions and you get an exact diff of what changed. Either way, week one ends with a brain that has structure, not just a pile of files.
With a brain in place, week two points it at real storage. Gateway indexes the archival data you nominated and gives you one search across it, returning files with their checksums and metadata. Gateway is read-only by design, so an index pass and every query after it can never change your data. It resolves permissions per question and writes an audit row on each one, with RBAC, MFA and SSO on the front door.
Finding a file is only useful if you can judge it, so Lens rides along. Lens previews about 140 formats in the browser with no restore and no pre-processing: camera RAW, EXR, DPX, broadcast video, audio, multi-page documents, legacy 3D as turntables. Video previews are roughly 50x lighter in our tests, so a reviewer scrubs a proxy instead of pulling a master. By the end of week two a producer can search the pilot archive and open what they find, which is the first moment the index stops being a promise and starts being a tool.
Reading is safe. The harder question is what happens when new material arrives, so week three runs a controlled ingest. Nexus takes a test set into a sandbox destination you name, and the ingest is context-aware: it lands against the graph from week one, not into an anonymous bucket. Nexus does atomic frame-sequence assembly, so you never end up with half a shot, and it runs checksum verification on intake so what landed is provably what left. Write paths in Nexus are explicit and opt-in, which is exactly why a test ingest can be scoped to a single sandbox and nothing else.
Note: The two sides stay cleanly separated. Gateway reads and reports; Nexus writes. The week two index never moves a byte, and the week three ingest only writes to the sandbox you chose. That separation is not a pilot convenience, it is how the platform is built.
By the final week you have a graph, a searchable index, and a proven ingest path. Week four opens all of it to agents. Fabric exposes 23 entity MCP tools, and MCP, the open agent-connector standard, is how an agent platform like Claude connects to your graph without bespoke integration work. The agent does not get a raw database. It gets the same governed graph a person queries, with permissions and version time-awareness enforced inside the query, so an agent asking what exists for a scene is bound by the same rules a producer is.
This is the part that usually reads as science fiction in a slide and turns concrete in a pilot. You watch an agent answer a real question about your real material, correctly, because the context brain underneath it is real too.
A ShotGrid export or a script becomes a governed graph in Fabric, or a diffable 12-category breakdown in Codex. Structure first.
Gateway indexes your chosen archive read-only and returns files with checksums; Lens previews about 140 formats in the browser.
Nexus runs a context-aware test ingest into a sandbox you name, with atomic assembly and checksum verification on intake.
Fabric's 23 entity MCP tools open the graph to agents over MCP, bound by the same permissions a person queries under.
The point of a four week pilot is not the demo, it is the paperwork. You walk out with the knowledge graph you built and a findings report expressed in your own numbers: what was found, what previewed cleanly, how the test ingest verified, and what an index of your estate is worth against your storage rates. That report is what a real budget conversation is built on, because it is measured on your material rather than a reference dataset.
What you do not keep is the pilot's copy of your data. The pilot runs single-tenant in Fortify's cloud on archival data you choose, and at wrap that data is deleted and the deletion is verified. If the pilot graduates to a full deployment, it runs in your own AWS account or on-prem, and in production deployments the content stays in your environment. The whole engineering posture is SOC 2 Type II aligned and designed for MPA content-security best practices, which is the standard the pilot is run to as well.
Four weeks is short on purpose. It is long enough to build something real and too short to turn into a project that outlives its own justification. If you want to see the index on your own estate rather than read about someone else's, that is the entire deliverable. For adjacent reading, teams weighing this usually look at tape is a feature, not a graveyard and give your DAM a brain next.