Forge reads your document estate into validated records with a page reference behind every field — trained on your own documents, deployed where your regulator requires, and the weights stay yours.
Score it on your own documents before you sign anything.
Five layers between a document nobody can query and a record your systems can trust. One owner, all five.
Swipe to follow the pipeline →
A forwarded phone photo of a signed amendment is the normal case, not the edge case. The ingest layer takes the estate exactly as it arrives — native PDFs, scans, spreadsheets, email threads, handwriting — and gives every page a stable identity it keeps for the rest of the pipeline.
OCR + layoutHandwritingEmail & attachmentsAny resolutionNo document is triaged out because a human would find it inconvenient.
Real estates are full of split scans, duplicate copies and amendments that silently supersede the terms above them. This layer stitches the pages back together, resolves which version governs, and reconciles every amendment against the original before anything is read for meaning.
Deskew & denoisePage stitchingVersion resolutionDe-duplicationExtraction never runs against a superseded version.
The classify layer names the family, resolves the parties against your existing masters, segments the clauses, and maps each term onto your canonical schema — yours, not a generic one, because the schema is what everything downstream is actually promised.
Document familyParty resolutionClause segmentationCanonical schemaEvery value lands in a named field, or it lands nowhere.
Every field carries a confidence score and a page reference. Anything under the threshold you set stops and waits for a named steward rather than flowing through quietly — and that correction is captured as a labelled example from your own documents.
Confidence gatesSteward routingField-level lineageAudit traceA confident error costs you far more than an abstention, so we design it that way.
Typed rows land in your masters and an API your systems call directly. Every value keeps a page reference back to the source, so when someone in audit asks where a number came from, the answer is a document and a page — not a prompt and a hope.
Typed schemasAPI + webhooksPage-level lineageZero egressInside your boundary nothing leaves your account; on gothink Cloud nothing trains ours. The weights, the records and the lineage stay yours.
Hover to explore
Capture tools extract values. They do not attest to them. Forge produces records that carry their own evidence — every value traceable to a file, a page and a region, every value below your threshold stopped and routed to a named steward rather than flowing through quietly. The cost of a bad extraction has moved from rework to attestation risk, and that is a different budget with a different owner.
They extract. They do not attest. A value arrives with no page reference, no confidence score tied to your risk policy, and no opinion about whether it should have been asserted at all.
Each one owns a single document family. Four families becomes four vendors, four schemas and four sets of stewards — and the reconciliation between them is still done by a person.
It reads a page. It has no schema for a trust deed, no steward workflow, no confidence gate tied to your risk policy, and no opinion about whether the value it produced is defensible. You would build all of that on top of it.
Forge is our intelligent document processing product, and Optim is the model routing and cost engine that ships inside it. Forge is a licence to a versioned system, not a project. Every customer runs the same artifact at the same version — what differs between two deployments is configuration you could read.
Five layers between a document nobody can query and a record your systems can trust: ingest, clean, classify, govern, serve. Plus the exception queue, the lineage store, the command centre and the eval harness — as one signed image.
Per-domain schemas, field definitions, confidence gates, validation rules, steward routing and a reference ground-truth set. Two ship today. Others are built as a scoped pack against a real deployment.
What is inside a packDepth on two beats maturity labels on seven. These are the packs that ship with a schema, a field list, confidence-gate defaults and a reference ground-truth set — versioned, dated and on the changelog.
Trust deeds, ACAT packs and held-away statements are re-keyed by hand today. Forge types them straight into the household master — including the held-away positions that never got structured, which is usually where the advice actually lives. Every value lands with a page reference, or it abstains into the steward queue.
Trust deedsACAT packsHeld-away statementsOnboarding packsHeld-away positions land in the same schema as custodied ones, with a page reference behind each.
Open the packBroker submissions, SOVs, loss runs and FNOL packs typed onto your policy and claims master. Underwriters see a typed risk rather than a stack of attachments, and every value on it traces back to the page it came from when the auditor asks.
Submissions & SOVsLoss runsFNOL packsMedical recordsUnderwriters see a typed risk, with every value traceable to the page it came from.
Open the packHover to explore
Everything on this site ladders to one of these. If a claim here cannot be demonstrated on demand, it should not be here.
Click a field in a live run and land on the source page. Anything below the threshold your risk function set abstains into the exception queue and waits for a named approver — because a confident error costs you far more than an abstention.
See the gate and the lineage YoursNot gothink’s. The base model is named in your contract schedule, with a warranty that its licence permits the tuned weights to transfer to you and run without us — and we will demonstrate that on your exported weights before you sign. Everyone else in this category rents you inference on a model you will never see. Three years of your experts’ corrections should end up on your side of the boundary.
Read the ownership schedule ScoredDo not take our number. Twenty-five of your own documents, labelled by your own experts, scored on five published measures. You end up with a figure we did not produce and cannot tune — and a method you can run against every other vendor on your shortlist.
Score your documentsOn gothink Cloud your data stays in your tenant and region. Inside your boundary nothing leaves at all. Nothing you run trains anything of ours. SOC 2 status, subprocessor list, DPA and the security questionnaire are all published rather than promised.
Trust Center Who approvesEvery automated action above your threshold waits, states what will change, and asks.
Autonomy & gates How it startsSchema mapping, gates, ground truth labelled by your experts, then a pilot gate you can walk away at.
How onboarding runsYou should be able to test the claims on this page, compare us against the alternatives, and work out how you would buy it — all before anyone from gothink is in the room.
Your documents, your labels, your number — plus the gate log, the exception list and a one-page summary written for a risk committee rather than for an engineer. Free, no contract, five business days.
Score your documents Compare usCitation correctness, gate precision, abstention correctness, reviewer agreement, evidence completeness. Open corpus, open harness, permissive licence — including for marketing against us.
The attestation benchmark Buy itForge runs in your own account, so buying it through the marketplace that operates that account removes a step rather than adding one. Direct, marketplace, embedded or partner-delivered — same artifact, same licence, same ownership.
Marketplaces & partners