How it works

A 13-stage pipeline. Deterministic first.

Manifesto extracts, normalizes, and verifies every row before a reviewer sees it. AI is a fallback for ambiguous cases — never the engine.

Request access

Watch the 13-stage pipeline process a real contract

Coming soon — request a live walkthrough

Watch the 13-stage pipeline process a real contract

13 stages, end to end.

Each stage is deterministic, versioned, and auditable. Failures cluster on document family — not random samples.

Foundation
01

Metadata

The pipeline opens the PDF and reads the cover page to extract the contract identifier, carrier name, shipper, effective and expiration dates, and minimum quantity commitment. This data anchors every downstream row to the correct contract version.

1 / 13
Why deterministic-first

AI is a fallback. Not the engine.

Auditability

Every row links back to the page, line, and tokens it came from.

Reproducibility

Pipeline runs are versioned. Reprocess and the diff is exact.

Explainability

Failures cluster on document family, not random samples.

Three states, made visible.

Every extracted row is scored against the PDF text that supports it. The result is one of three evidence states — surfaced directly on the row, not hidden behind a single confidence number.

This is evidence verification (does the PDF support the row?), distinct from benchmark verification (does the row match a template?) which is optional.

  • verifiedStrong PDF text support.
  • needs-reviewPartial or ambiguous support.
  • weakMinimal or no text match found.
CNSHA→USLAX$1,450
verified
CNNGB→USLGB$1,500
needs-review
CNYTN→USNYC$1,680
weak
Architecture

Built on infrastructure your security team will accept.

PDF Upload
Pipeline Workers (Render)
Neon Postgres
Supabase Auth
Vercel Frontend

Cloudflare R2 for encrypted object storage. EU residency available on request. Read the security overview →

Questions from engineering teams.

What's the LLM used for?
Fallback only — for genuinely ambiguous cases where deterministic rules can't resolve a value. Every LLM-touched row is flagged and carries a confidence score.
Can we bring our own templates?
Yes. Templates enable optional benchmark mode — comparing extraction output against your ground truth. They are not required for operational use.
How are amendments handled?
Each amendment is processed against the base contract. Row-level diffs are preserved with full version history.
How fast is a typical run?
Minutes per contract on average, bounded by PDF size and page count. Runs are async and parallelized across workers.
Can we run it on-prem?
Not today. Cloud-hosted with EU residency option on request. On-prem is on the roadmap for enterprise plans.

See the pipeline on your contract.

Bring a sample to the discovery call. We'll run it live and walk through every stage.

Request access