Forge ships as one versioned system, identical for every customer. Onboarding fits it to your estate: we map your schemas, set your confidence gates, and label a ground-truth set with your experts. Fixed scope, fixed price, fixed end date. After that you are running software, not a project.
You buy a statement of work with a fixed fee and a date on it. There is no open-ended build phase, because there is nothing customer-specific left to build — everything Forge does for you is configuration.
| Week | What happens | Who from your side | Ends with |
|---|---|---|---|
| 01 Estate & schema | Your document families are inventoried against the pack schema. Fields your systems of record need but the pack does not carry are added as schema entries. Connectors are pointed at your sources and targets. | A data owner and a platform engineer, part-time | schema mapping signed |
| 02 Gates & ground truth | Your risk function sets the per-field confidence threshold. Your experts label a ground-truth set from your own estate — the set that will score every release from here on. Steward routing is wired to your queues. | A steward or SME, and whoever owns the risk policy | thresholds set, ground truth labelled |
| 03 Pilot gate & go-live | Forge runs the estate and is scored against the set your experts labelled. If it clears the gate criteria in the run-book, it goes to production and the tuned weights are handed over. If it does not, you stop here. | Whoever signs off on a wrong value reaching a client | scored, live, weights handed over |
It would be easier to leave this vague. Most software companies do, because naming the humans in delivery invites the question of whether you are buying software or people. The honest answer is that three weeks of expert work sits between your estate and a working deployment, and pretending otherwise would be the first thing you discovered was untrue.
The engineer who fits your estate writes the runtime. Not a delivery organisation, not a partner bench, not a solutions team measured on billable hours. Whatever they learn in your estate lands in the product because they are the person who ships the product.
Anything built during a fitting must become part of the runtime for every customer, or part of a versioned pack, or it does not get built. There is no third option and no customer branch. When a request genuinely cannot be generalised, we decline the work rather than fork the codebase.
The metric on this function is the share of what was built during fittings that shipped to everyone within a quarter — not utilisation, not hours. You can check it yourself: every promotion appears on the public changelog with the release it landed in.
Four things distinguish them, and all four are checkable. The fee is fixed and the end date is in the statement of work. The engineer belongs to engineering and their output goes to the main repository. There is no per-customer branch to maintain, which you can verify against the changelog. And the work is transferable — certified implementation partners can run a fitting instead of us, which is not something a consultancy would ever design for. A delivery model that a vendor is willing to hand to somebody else is a product.
Every customer runs the same Forge runtime at the same version. What differs between two deployments is a set of files you could read.
The master fields your systems of record expect, and how each document family maps onto them. A versioned file in the pack, extended for you at Fitting.
Per field, set by your risk function. A number, not a code path. Change it on a Tuesday without a release.
Which exception goes to which queue, which cross-field checks must hold, which entitlements apply before a record is served.
Which sources, which masters, which targets, under which credentials. From the released connector library.
The ground truth your experts labelled. It scores every release we ship you, and it is yours to take.
If you need something Forge does not do, it goes into the runtime for everyone, or into a pack — or it does not get built. Where a request genuinely cannot be generalised, we decline the deal rather than fork the codebase.
Read the public changelog. It is very hard to maintain a per-customer branch and a public changelog at the same time. Every version we have shipped is on that page with a date, and the version you are running is printed in your control plane.
Deployment location is a modifier on the licence, not a rebuild. The image, the version and the release notes are the same on gothink Cloud as they are inside a bank's VPC; what changes is the runbook and what the connectors are permitted to reach.
| Profile | Runs on | Same image? | Typical reason |
|---|---|---|---|
| gothink Cloud | An isolated tenant in the region you choose | Yes | Fastest path to production; no infrastructure work on your side |
| Customer VPC | Your AWS, Azure, GCP or Kubernetes, your IAM and keys | Yes | Policy or regulator requires the documents never leave your boundary |
| Air-gapped / sovereign | On-prem or a national cloud, with an offline update channel | Yes | No outbound network path is permitted at all |
There is no pod to leave, because there was never a pod. What recurs after go-live is software and an SLA.
Quarterly feature releases, security patches for the current and previous major, every one of them on the public changelog.
When ACORD revises a form or a custodian changes a statement layout, the pack ships the change as a new version. You do not re-fit.
Standard is business hours through the portal. Enterprise adds priority response, 24×5 and a named engineer. Both are documents you can read before you sign.
The ground truth your experts labelled at Fitting runs against each new version before you take it, so an upgrade cannot quietly cost you accuracy.
Accuracy, straight-through rate, drift and lineage completeness, formatted for model-risk and audit functions — and if the number has fallen below the floor set at Fitting, gothink retrains at gothink’s cost. The guarantee, in full.
The weights, records and lineage are yours permanently. The system keeps running on the schemas and the accuracy it had the day you stopped. Why people renew anyway.
Before any of the above, score Forge on twenty-five of your own documents against a ground-truth set you label yourself.
Your documents, your labels, your number. No contract and no call.
Fixed scope, fixed price, fixed end date, and an exit at the pilot gate.
The tuned weights, the records and the lineage are handed over at go-live.