Operational system · isolated synthetic documents · no client-result claim
Source-to-Record Workbench
Turn a field document into a reviewable operations record—without losing the line, amount, or ambiguity that produced each field.
- Supported
- Delivery note · work order
- Field states
- Verified · review · missing · rejected
- Evidence
- Page · line · character region
- Last reviewed
Operational frame / 01
The delay is not extraction. It is deciding what the office can trust.
Field-service operations
Office, dispatch, receiving, and accounting teams re-keying delivery notes and work orders after a job is accepted.
Source re-checking
A person must reopen the document when a job reference, amount, date, or duplicate looks wrong—often after the record has already moved.
Document + evidence
The proposed record never replaces its source. Every populated field keeps an exact, inspectable region and reason.
Approved JSON + CSV
Exports include field status, original proposal, reviewer correction, and evidence only after explicit source comparison.
No silent choice
Ambiguity, missing values, arithmetic conflicts, duplicate IDs, and hostile instructions stop or route to review.
Signed downstream outbox
Approved records enter a durable queue with stable event IDs, leases, receipts, retry, cancellation, and rollback state for the configured destination.
Isolated public tour / 02
Compare the document. Correct the record. Approve the boundary.
Use synthetic or de-identified text only. Do not enter customer names, real addresses, credentials, medical data, access codes, or confidential work notes.
Six pinned documents
Choose the failure boundary you want to inspect.
Paste a synthetic tour document
Ready. Choose a synthetic scenario or enter bounded fictional text.
- Document type
- Not classified
- Identifier
- Not proposed
- Validation
- 0/0
- Evidence coverage
- 0/0
Source / immutable
Synthetic document
No source region selected.
Choose a synthetic scenario or enter bounded fictional text. The source will appear here without persistent browser storage.
Proposed / editable
Closed record
Select a scenario to inspect field state, reason, source coordinates, and reviewer corrections.
Deterministic boundary
Validation checks
Local audit trail
Correction history
No correction, approval, or handoff event has been recorded.
Human release gate
Reviewed output remains locked.
Technical trace + retention
- Trace ID
Not created- Source fingerprint
Not created- Local latency
- Not measured
- Contract versions
- Not loaded
- Public-tour provider
- Not invoked · tokens not applicable
- Public retention
- Not persisted · reload clears local state
- External writes
- 0
- Public-tour URL fetching
- Blocked by design
Portfolio evidence ledger
Evidence, production controls, and release assurance.
Isolated public tour · fictional or synthetic data · operational controls verified separately.
- Exact role
- Designed and implemented secure file preflight, mandatory malware-scan integration, local text/OCR extraction, the closed schema, exact-evidence validation, encrypted durable review states, role-based approval, controlled export, and signed downstream outbox.
- Human boundary
- A person compares every proposal with exact source evidence, resolves ambiguity and conflicts, and explicitly approves the reviewed record before export or signed downstream delivery.
Evidence in the build
- Evidence rule
- Exact substringA retained field must point to matching source text.
- Output shape
- Closed schemaUnexpected fields do not enter the reviewed record.
- Export state
- Human controlledNeither the tour nor operational service exports an unreviewed result.
- Operational record
- Encrypted + versionedSource, proposal, corrections, approvals, outbox, and receipts survive restart without clear-text document content.
- File intake
- Scan before extractionProduction ingestion requires valid type and structure plus a clean malware result before embedded text or OCR can run.
- Downstream delivery
- Signed + idempotentOnly approved records enter the durable webhook outbox with leases, receipts, explicit retry, cancellation, and rollback state.
Tests actually run
- 27/27Focused domain and HTTP tests
Covered extraction, evidence, deterministic validation, input boundaries, and no-write HTTP behavior.
- 272/272Complete unit suite
Completed in the documented Task 03 implementation run.
- 13/13Pinned evaluation fixture contracts
Expected outcomes also passed 13/13 with no external provider call.
- 94/94Deterministic evaluation rule outcomes
The same saved run also reconstructed 150/150 evidence regions and passed 5/5 fail-closed cases.
- 13/13Focused browser and portfolio gate
Eight dedicated workbench tests plus five adjacent portfolio and migration checks.
Production controls implemented
- Authenticated workspace intake validates declared type, MIME, magic bytes, file size, PDF structure, page count, image dimensions, and a safe generated filename before extraction.
- A mandatory malware scanner fails closed on stale signatures; local PDF text extraction and image/PDF OCR share one total deadline and run only after a clean result.
- The original binary, source text, proposal, corrections, resolutions, outbox payloads, and receipts are encrypted in a durable private SQLite store and removed together at retention expiry.
- A durable reservation precedes scanner, extractor, and provider work; stable provider IDs, bounded attempts, abort propagation, and expiry recovery make restart and retry safe.
- Operator, reviewer, and administrator roles enforce original access, corrections, conflict resolution, approval, rejection, export, and delivery through optimistic versions and an HMAC-verified audit chain.
- Only an approved record can enter the durable signed-webhook outbox; delivery uses stable events, bounded leases, receipts, Retry-After-aware backoff, explicit retry, cancellation, and rollback state.
Enforced production activation gates
- File intake refuses to start unless the configured malware scanner and the required local extraction or OCR runtime are healthy.
- OCR-derived fields remain in human review until exact source evidence is reconstructed; ambiguous or malformed documents release no record.
- The readiness gate verifies signed activation evidence and requires managed encryption, authenticated roles, current scanner health, approved extraction policy, off-site recovery proof, monitoring, and an approved downstream mapping.
Architecture / 03
A document adapter may propose. The release contract still decides.
- 01
Authenticate + inspect
Bind a supported document of up to 20 pages to one workspace, verify MIME and magic bytes, enforce byte/page/pixel/character limits, and require a clean malware scan.
- 02
Extract + propose
Use embedded PDF text or local English/Spanish OCR page by page, then map explicit source literals into a versioned closed schema.
- 03
Bind the evidence
Attach reconstructable source coordinates to every populated field; OCR-derived values remain in review.
- 04
Validate + review
Run deterministic checks, retain encrypted corrections and resolutions, and enforce optimistic record versions and reviewer roles.
- 05
Approve + deliver
Unlock JSON or CSV only after approval, then use the durable signed-webhook outbox with receipts and explicit retry.
Validation rules / 04
A status is an explanation, not a confidence number.
Exact evidence
A verified proposal must reconstruct the exact source substring at its declared coordinates.
Required fields
Missing identifiers, dates, vendor, references, currency, lines, or totals remain empty and block release.
Line arithmetic
Quantity × unit price must equal each line total, and the line sum must equal the subtotal.
Total arithmetic
Subtotal + tax must equal total; the source literal stays visible even when its state drops to needs review.
Duplicate ID
A pinned synthetic processed-ID registry blocks a repeated document identifier before approval.
Untrusted text
Instructions inside a document are data. They cannot change schema, verification, export, or handoff policy.
Measured evaluation / 05
Thirteen pinned failure shapes, measured against one contract.
Measured July 20, 2026 with deterministic synthetic fixtures; no external provider calls. These are synthetic fixture contract results—not client-document accuracy or commercial performance.
- Dataset
- 13/13 cases
- Rule outcomes
- 94/94
- Closed-schema paths
- 16/16
- Populated-field evidence
- 138/138
- Region reconstruction
- 150/150
- Fail-closed cases
- 5/5
- Unsafe verified exports
- 0
- External side effects
- 0
- Local latency p50 / p95
- 0.53 / 8.18 ms
Reproduction record
- Evaluation clock
2026-07-20T12:00:00.000Z- Schema / validator
source-to-record-record.v1 · source-to-record-validator.v1- Dataset SHA-256
c4ba7c7b8ae89951be8601eba8bea5bbb3b9d539a3c07b0f2101d8386ebc6977- Config SHA-256
062663a1d95829226c1ea59cd2b4bbd6e0af3cb8e564e3c90cc2362435e6ac20
- These metrics cover pinned synthetic text fixtures and deterministic boundary simulations; customer-specific acceptance uses authorized representative documents.
- The latency value is local contract-evaluation time; deployed extraction and delivery latency are recorded by the operational monitoring path.
- Authorization, encrypted retention, durable audit, and reviewer operations are verified by the operational service suite rather than counted in this fixture metric.
Failure + privacy boundary / 06
Every document crosses the same security and evidence gates.
Fail closed
Malformed output, unsupported evidence, scanner failure, and extraction timeout release no partial record and no downstream event.
Inspect before extraction
MIME, magic bytes, PDF structure, encryption, pages, dimensions, pixels, bytes, names, and allowed document families are validated before use.
Clean result required
The scanner is mandatory and unavailable, errored, or infected results stop ingestion. OCR never runs before a clean verdict.
Encrypted lifecycle
Source, proposal, review changes, exports, queue payloads, and receipts are encrypted with explicit retention, purge, and backup operations.
Roles + record versions
Operators propose; reviewers correct, resolve, approve, or reject; administrators control export, queueing, and delivery. Stale versions cannot overwrite a committed decision.
Exact source required
Every retained field reconstructs its source. OCR and ambiguity remain in review until a person resolves them; invented evidence is rejected.
Production control plane / 07
The production path is implemented as explicit, inspectable state.
- 01
Authenticated intake
Salted workspace credentials, role checks, strict upload preflight, content fingerprinting, encrypted originals, bounded retention, and cascading purge create the intake boundary.
- 02
Document runtime
Current malware signatures, local PDF text extraction, image/PDF OCR, an exact page map for up to 20 pages and 100,000 characters, safe temporary files, and one total deadline form a fail-closed pipeline. A governed external provider receives neither the original bytes nor private filename.
- 03
Reviewer operations
Optimistic versions, original-access audit, corrections, conflict resolutions, approve/reject states, and an HMAC-verified record-event chain survive restart and backup.
- 04
Downstream delivery
An approved record envelope enters the durable outbox and reaches the configured system through signed HTTPS, stable events, bounded leases, receipts, Retry-After-aware backoff, explicit retry, cancellation, and rollback.
- 05
Activation gate
The service verifies signed activation evidence and refuses operational readiness unless scanning, extraction, authentication, audit integrity, off-site recovery proof, monitoring, provider policy, and required delivery are healthy.
Portfolio connection / 08
The record workflow is designed for the next operational boundary.
The Bilingual Lead-Response Control Room preserves the inquiry, language, owner, and next step. Source-to-Record governs the later document-to-record boundary. The authenticated services connect only through an approved deployment mapping and exchange only its declared fields.
Map the source-to-record delay
Find the field your team checks twice before the next system trusts it.
Start with one document family, its source owner, the review decision, and the downstream write that must stay controlled.
Get a 3-point lead-flow review