Forge is gothink.ai’s intelligent document processing product. It reads a document estate into validated records with a page reference behind every field. Five layers, confidence gates tied to your risk policy, an exception queue with named approvers, a lineage store and the eval harness — shipped as one signed image, not as a project.
What differs between two deployments is configuration you could read. There are no customer-specific branches, and the public changelog is how you check that.
Everything below ships in the same signed image at the same version, whether you run it on gothink Cloud, in your own VPC or air-gapped. Nothing here is a separate purchase and nothing is held back to create a pricing ladder — the tiers differ by where it runs and what you own, not by what it does.
| Component | What it does | Ships as |
|---|---|---|
| The five-layer pipeline | Ingest, clean, classify, govern and serve. One owner across all five, which is the point — a chain of specialist tools leaves the reconciliation between them manual. | v1.0.0 · in the image |
| Model runtime | Executes the specialist model tuned on your estate, inside your boundary, on the model family you select. | v1.0.0 · in the image |
| Confidence gates | Per-field thresholds set by your risk function. Below threshold, a value abstains rather than asserting itself quietly. | v1.0.0 · thresholds are config |
| Exception queue | Every abstention routed to a named steward, with the reason and the source page attached. Corrections are captured as labelled examples from your own documents. | v1.0.0 · routing is config |
| Lineage store | File, page and bounding region behind every extracted value. Queryable through the API, clickable in the command centre. | v1.0.0 · in the image |
| Command centre | Run status, throughput, straight-through rate, exception backlog and cost per document family, by role. | v1.0.0 · in the image |
| Eval harness | Scores any release against a ground-truth set your experts labelled. Exposed as a product surface, which is what makes scoring us before you sign possible at all. | v1.0.0 · in the image |
| Connector library | Released, versioned source, master and target connectors. Which ones point where is set during Fitting. | v1.0.0 · wiring is config |
| Optim | The routing and cost engine underneath the pipeline: per-page model routing bounded by your accuracy floor, realtime cloud and token metering, and spend forecasting. Not separately licensed — what it does. | v1.0.0 · in the image |
| Estate packs | Schemas, field definitions, gate defaults, validation and reference ground truth for one document estate. Two ship today. | Versioned separately |
That is the entire layer. Each stage runs on gothink Cloud in a tenant isolated to you, or inside your own boundary — under your credentials, on your audit trail, either way.
Swipe to follow the pipeline →
Shared drives, mailboxes, scanning queues, earnings call audio, video exhibits, supplier portals, and archival silos.
Layout-aware OCR keeps columns, tables, stamps and signatures where they were. Skewed scans are straightened; genuinely illegible pages are flagged, not guessed at.
Document type first, then the things inside it: parties, dates, amounts, identifiers, clauses, line items — typed in context rather than pattern-matched, and bound to the schema your structured feeds already use.
Your validation rules, automated PII redaction, entity resolution, and per-field confidence scoring. Ambiguous values route directly to human stewards, feeding an active learning loop that continuously retrains extractors.
Published once as a canonical data asset: written into core transactional databases, indexed with entitlement controls for enterprise RAG, and exposed as typed tools for agentic workflows.
One document from a run shaped like yours. Nothing under the bar reaches the database on its own.
| Field → your schema | Value & source | Confidence | Outcome |
|---|---|---|---|
| contract_value | $25,000,000 · page 1, cl. 2.1 | 0.99 | PASS |
| counterparty_id | Redwood Ind. LLC → PARTY_MASTER #48812 | 0.98 | PASS |
| effective_date | 2026-04-01 · page 1, preamble | 0.97 | PASS |
| payment_term_cd | NET45 · page 6, cl. 9.3 | 0.96 | PASS |
| termination_notice | 90 days · page 11, cl. 14.2 | 0.95 | PASS |
| governing_law | ambiguous — two jurisdictions named, page 12 | 0.87 | HOLD |
Raise it and more goes to review; lower it and more goes straight through. Either way the trace records which happened, for every field — including the ones a steward confirmed by hand.
Corrections are not discarded. They feed back as labelled examples, so the extractor that got governing_law wrong on a two-jurisdiction contract gets better at exactly that shape of document.
Every deployment runs the same runtime at the same version. This is the whole list of what differs between two customers — and every item on it is a file you could open and read.
| What differs per customer | Form it takes | Changed how |
|---|---|---|
| Schemas and field definitions | Versioned data file in the estate pack, extended for you at Fitting | Config |
| Confidence thresholds | A number per field, set by your risk function | Config — change it on a Tuesday, no release |
| Validation rules | Cross-field checks and party resolution against your masters | Config |
| Steward routing | Which exception reaches which queue and which approver | Config |
| Connector wiring | Which sources, masters and targets, under which credentials | Config |
| The eval set | Ground truth your experts labelled — yours to keep | Data |
| The tuned weights | The specialist model trained on your estate — your asset, permanently | Data |
| A branch of the runtime | Does not exist, at any tier, for any customer | Never |
It goes into the runtime for everyone, or it goes 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, because a fork is a promise to maintain a second product forever in exchange for one signature. You can check that this holds by reading the public changelog: it is very hard to maintain a per-customer branch and a public changelog at the same time.
Without the layer, every consumer is its own extraction script, its own integration, and its own argument about which number is right.
Governed fields written back into ERP, CRM, or master databases.
Indexed with entitlements intact, citations down to the pixel.
Exposed by API to anything that needs to act on the record.
Every value with its source page, ready for compliance review.
K-1s, pitchbooks, earnings transcripts, and credit agreements normalized into master records.
ACORD forms, loss runs, and adjuster notes turned into deterministic payout logic.
Data room extraction, change-of-control clauses, and regulatory filings with 100% citation trails.
Complex multi-line POs, vendor terms, and service agreements validated against ERP systems.
The next consumer should be a permission grant, not another project. A reasoning engine is only ever as good as the corpus it can reach — and buying a better engine does not widen the corpus.
The five layers do not change — the runtime does not know or care what industry a document came from. What a pack adds is a schema someone has argued about, gate defaults a risk function has signed off, and a labelled ground-truth set. Wealth Management and Insurance Submissions ship at v1.0.0; the rest are built during Fitting.
Trust deeds, ACAT packs and held-away statements typed into the household master.
LPAs, PCAPs and capital notices typed onto your fund, party and security master.
Clause-level lease abstraction at portfolio volume, without an offshore queue.
Account packs, loan files, KYB packs, remittances and disputes typed onto one customer, merchant and settlement schema.
Broker submissions, SOVs, loss runs, FNOL packs and medical records typed onto your policy and claims master.
PBAs, ISDA schedules and lock-up addenda typed onto your client and financing master.
Structure documents, source-of-wealth evidence and pledge terms typed onto one client master.
The pipeline is the same in every one of these. Each industry is an estate pack — the schema, extractors, thresholds and update stream — added to your licence, then tuned on your own ground-truth set during onboarding before anything scales.
The runtime is licensed software, not a project deliverable. That is what keeps it current after onboarding ends. The licence is an annual subscription, per deployment. Estate runs on gothink Cloud; Enterprise and Sovereign put the same runtime inside your VPC or on-prem.
Quarterly feature releases and security patches, signed, with a changelog and an SBOM. Offline bundles for air-gapped estates.
Release notesSources, masters and targets maintained as versioned artefacts under the licence. New ones join the library through Extend or onboarding.
The libraryAccuracy trend, drift, steward queue, lineage coverage, release inventory and licence state — what your team sees every month. Hosted in your tenant on gothink Cloud; inside your boundary on Enterprise and Sovereign.
Business hours on Estate and Enterprise; 24×7 severity-1 with a named security contact on Sovereign.
Schema, extractors, thresholds and rules per industry, with an update stream when forms and regulations change.
Seven packsRecalibration and drift, scheduled new document types, and a signed quarterly accuracy attestation for your model-risk function.
| A year with Forge | When | What happens | Who |
|---|---|---|---|
| 01 Onboarding | weeks 1–6 | Discovery on your estate, thresholds tuned on your ground-truth set, connectors wired, one document family live into a sandbox master, then production. Fixed scope. | gothink pod + your team |
| 02 Live on the licence | week 6 | The number is proven on your documents. The licence carries the runtime from here; the pod hands to your operators and, if chosen, to gothink Operate. | your operators |
| 03 Weekly recalibration | every week | Steward corrections retrain extractors where the pipeline runs — in your tenant or inside your boundary. Drift is watched in the control plane; under Operate, gothink acts on it. | your stewards · Operate |
| 04 Quarterly release & pack updates | every quarter | A signed runtime release with a changelog. Pack updates ship when a form, a regulatory schema or a custodian layout changes. | gothink |
| 05 Quarterly attestation | every quarter | Under Assure: accuracy attestation, audit pack and lineage report, signed by gothink, ready for your model-risk and audit functions. | Assure |
| 06 Annual review | month 12 | Accuracy trend, packs added, estates covered, and the renewal and expansion conversation. If you do not renew, nothing switches off. | both |
Twenty-five documents from your own estate, typed onto a pack schema, scored against ground truth your experts labelled. Free, no contract, and you keep the report either way.
Twenty-five documents from your own estate, labelled by your own experts. Free, no contract, and you keep the report either way.
Schema mapping, confidence gates and a ground-truth set, at a fixed price with a fixed end date. Exit at the pilot gate having paid the first milestone only.
Go live on the licence with the tuned weights handed over. Releases, packs and support carry on from there.