Systems and records
Approved, rejected or superseded: status labels for a document stamp signature dataset
By SourceX Editorial · Updated
10 sections
Short answer
Label each status stamp in three layers: the mark (region, kind, legibility), the reading (the exact printed text plus a status mapped from that source's own legend) and the governance (what the stamp applies to, its date, and whether a later record replaced it). Treat superseded as a link to the replacing record, and exclude pages whose status the file and its logs cannot resolve instead of guessing.
Document stack
Approved
Key takeaways
- A detector finds stamps; a status label records what the stamp decided, what it governs and whether it still stands.
- Keep the verbatim stamp text beside a normalized status, and map each source's legend separately because codes differ between organizations.
- Superseded is a link from an older record to its replacement, backed by evidence such as a register entry, not just a word on the page.
- Exclude pages whose status cannot be resolved from the file and its linked records, and keep an exclusion code so the dataset can report what was left out.
- Measure agreement per field; approved versus approved with conditions is where guides usually break.
What does a status stamp tell a model that a signature box does not?
A signature or stamp detector answers whether a mark is present and where. A status label answers what the mark decided, what it applies to, and whether that decision still stands. Those are three different questions, and a dataset that only boxes the ink cannot teach the second and third.
Status stamps sit on submittals, shop drawings, invoices, inspection reports, purchase orders and controlled procedures. Their wording carries fine distinctions. The U.S. Army Corps of Engineers transmittal form (ENG Form 4025, May 2017 edition) separates code B, approved except as noted on drawings with no resubmission required, from code C, approved except as noted with comments attached, and both from code E, disapproved. A model that collapses B and C into approved will tell a contractor to build from a drawing the reviewer asked to see again.
This guide covers the reading and governance layers: the status text, its scope and its place in a version chain, plus the rule for leaving a page out. Detection of the marks themselves is covered in the companion guide on signature and stamp detection data.
The three layers of a stamp status label
Label every status stamp in three layers and keep them in separate fields: the mark (where it is and how readable), the reading (what it says, verbatim and normalized), and the governance (what it applies to, when, and whether something later replaced it). Separating them lets you reuse one annotation pass for detection, status classification and version reasoning, and lets a reviewer see exactly which layer failed.
The table is the field list of the annotation guide. Field names are suggestions; the split between layers is the part to keep.
| Field | Layer | Allowed values | Annotation rule |
|---|---|---|---|
| stamp_region | Mark | Box or polygon in page pixels, with render DPI | Box the ink extent, including handwritten initials or dates inside the stamp frame |
| stamp_kind | Mark | status, received, paid_entered, copy_control, reviewer_id, other | Classify by function, not by shape or color |
| legibility | Mark | clear, partial, illegible | Partial means a person can read the status word but not every field |
| status_text_verbatim | Reading | Exact printed and handwritten text, including checked boxes | Transcribe what is there; never complete a word from memory of the template |
| status_normalized | Reading | One value from the source's mapped legend (next section) | Leave unmapped when the legend is missing; do not guess |
| status_scope | Governance | whole_document, page, item_list, unknown | Item_list requires links to the item regions it governs |
| action_date | Governance | Date as written plus normalized ISO date, or blank | Use the stamp's own date, not the scan or file date |
| supersession | Governance | current, superseded, void, unresolved | Superseded requires a link to the replacing record and the evidence type |
| exclusion_code | Governance | Blank or one code from the exclusion table | Set only after the other fields are filled |
Mapping printed wording to a normalized status without losing it
Map each source's own legend to a small normalized set, and store the verbatim text beside the normalized value. The same letter or phrase can mean different things in different organizations, so a single global mapping quietly mislabels whole archives.
A practical normalized set has eight values: approved, approved_with_conditions, revise_and_resubmit, rejected, received_only, informational, superseded and void. Build the mapping from the legend printed on the form, the transmittal template or the reviewing office's procedure, and record which legend version you used. When a page shows a code letter with no legend in the file or its linked records, leave status_normalized blank and keep the letter in status_text_verbatim.
The example rows use the Corps form's published wording to show the method. They are a mapping pattern, not a universal standard.
| Printed wording (example source) | Normalized status | Common labeling error | Rule that prevents it |
|---|---|---|---|
| A, approved as submitted | approved | None; this is the easy case | Still record scope; an approval can cover only listed items |
| B, approved except as noted, resubmission not required | approved_with_conditions | Labeled plain approved, dropping the notes | Any wording with except, as noted or subject to maps to conditions |
| C, approved except as noted, refer to attached comments (resubmission required on the Corps form) | revise_and_resubmit | Labeled approved because the word approved appears | Resubmission required outranks the word approved |
| E, disapproved, refer to attached comments | rejected | Treated as the end of the item's history | Link the next submission; rejection usually starts a new revision |
| RECEIVED with a date | received_only | Read as acceptance | Receipt stamps never carry a decision unless a status box is checked |
| SUPERSEDED or VOID across the page | superseded or void | Labeled as a rejection of the content | Supersession and voiding describe currency, not quality |
Superseded is a relation between records, not a word on a page
Label supersession as a link from the older record to the record that replaced it, with the evidence that proves it. Many superseded pages carry no stamp at all: the later revision simply exists, and the archive shows it only through a revision block, a drawing register or a transmittal log.
At a glance
Superseded is a relation between records, not a word on a page
Existing vocabularies already model this. The Dublin Core term isReplacedBy describes a related resource that supplants, displaces or supersedes the described resource, and its inverse is replaces. The W3C PROV ontology treats a revision as a derivation in which the newer entity is a revised version of an earlier one, and records invalidation as a separate event with its own time. Borrowing those two ideas gives the annotation guide a clean rule: a revision link says which record came from which, and a supersession date says when the older one stopped governing.
Record the evidence type for every supersession link, because they are not equally strong:
- Stamped: a SUPERSEDED or VOID mark on the older page names or implies its replacement. Strongest when the stamp cites the new revision.
- Revision block: the newer page lists the older revision letter or number in its own revision history.
- Register or log: a drawing register, document control list or transmittal log records the issue of the newer revision.
- System record: a document management or ERP history shows the new version and the date it was released.
- Inferred: same document number, later date, nothing else. Mark as unresolved, not superseded, until another record confirms it.
Which page region and linked record hold the evidence?
The evidence for a status label usually sits in two places: the stamp region on the page, and at least one record outside the page that says what the stamp governed. Annotate both, and link them, or the label cannot be checked later.
On the page, capture the stamp region, any checked box inside it, the reviewer's handwritten notes, the item list or clouded revision areas the stamp refers to, and the legend if the form prints one. Region-to-region links are a standard feature of annotation tools; Label Studio, for example, provides a Relations tag for labeling the relationship between two labeled regions, and it works with image data. Use it to link a stamp to each item it governs and to the note that conditions it.
Outside the page, attach the identifiers that let a reviewer confirm scope and currency: the transmittal or submittal number, the revision register entry, the purchase order or invoice line, and the next revision if one exists. When OCR reads the stamp text, keep the word-level output. Tesseract's TSV output, for example, gives each recognized word a bounding box and a confidence value, which lets the guide set a rule such as sending any status word below a chosen confidence to a second reader.
- Source file name and checksum, page number, page count and render DPI
- Stamp region, legibility and any overlap with printed text
- Status text verbatim, plus the legend and its version
- Linked item regions or line identifiers the stamp governs
- Footnotes, attached comment sheets and reviewer notes referenced by the stamp
- Transmittal, register or system record that confirms the action date
- Link to the replacing revision and the supersession evidence type
- OCR engine, version and word confidences for the status text
Exclusion guide: which stamped pages to leave out of a status set
Exclude a page from status training when its status cannot be resolved from the file and its linked records; do not guess and do not silently drop it. An excluded page keeps its mark-layer labels, so it can still train detection, and it keeps an exclusion code so the dataset documentation can report how many pages were left out and why.
The table is the second half of the annotation guide. Codes are illustrative names; what matters is that each one maps to one decision.
| Situation | Decision for the status task | Why | What to record |
|---|---|---|---|
| Status word illegible and two readers disagree | Exclude (X-ILLEGIBLE) | A guessed status teaches the model to guess | Both readings and the OCR confidences |
| Code letter with no legend in the file or linked records | Exclude (X-NOLEGEND) | Letters mean different things across reviewers | Verbatim letter and the source organization |
| Two status stamps with no evidence of order | Exclude (X-CONFLICT) | The current decision cannot be known | Both stamps, dates if any, and the missing record |
| Stamp scope unclear on a multi-item page | Keep at page level with scope unknown | Page status is still true; item status is not | Which items might be governed |
| Specimen, template or training stamp | Exclude (X-SPECIMEN) | It records no decision | Where the specimen came from |
| Stamp copied from an earlier revision onto a later one | Keep, labeled carried_over | The decision belongs to the earlier revision | Link to the revision that received it |
| Stamp or notes cover personal data | Redact the personal data, or exclude (X-PRIVACY) if redaction removes the status | Status labels never need a person's identity | Redaction method and reviewer |
| Status shown only in a system field, no stamp on the page | Do not add a stamp label; keep as a record-level label | The page image does not contain the evidence | System, field and export date |
Worked example: one submittal across three revisions
Illustrative example: invented to show the method; the counts and identifiers do not describe a real project. A mechanical subcontractor's pump schedule goes through three revisions over several weeks.
Revision 0 comes back stamped with code C and two pages of comments. The annotator labels status_text_verbatim as the printed C wording, status_normalized as revise_and_resubmit, scope as whole_document, and links the stamp to the comment sheet. Revision 1 comes back with code B and one clouded note on pump P-3. It is labeled approved_with_conditions with scope whole_document and a link from the stamp to the clouded note region.
Four weeks later a design change triggers revision 2, which comes back with code A. Revision 1 carries no SUPERSEDED stamp, but the transmittal log records the issue of revision 2. Revision 1 is labeled superseded with evidence type register, and both revisions are linked with a revision relation. Revision 0 keeps its revise_and_resubmit label; it was never wrong, only replaced.
The exception: the clouded note on pump P-3 in revision 1 was carried into revision 2 by the reviewer writing see previous note. The guide keeps the revision 1 note in the dataset, linked to revision 2, because the current approval depends on it. A rule that dropped every superseded page would have deleted the condition the final approval relies on.
Across the package, that gives three status labels, two revision links, one supersession link with register evidence, and no exclusions. If the transmittal log had been missing, revision 1 would be marked unresolved rather than superseded, and a task that asks for the current approved schedule would exclude it.
Checking agreement before the guide scales to thousands of pages
Double-annotate a sample from each source before full annotation and report agreement per field, not one overall score. Mark-layer fields usually agree well; the readings and governance fields are where guides fail, and a single averaged figure hides that.
Expect the hardest confusions to be approved versus approved_with_conditions, received_only versus approved, and superseded versus unresolved. Adjudicate each disagreement against the legend and the linked records, then add the resolved case to the guide as an example. Stop and revise the guide if one field keeps disagreeing after two rounds; that usually means the legend mapping is wrong for that source.
Public sets will not supply status semantics. The DFKI StaVer collection provides 400 document images with pixel-accurate ground truth built from dummy bills with colored logos, text and various kinds of stamp, and DDI-100 includes stamp masks in its ground truth. Both are useful for finding marks; neither labels what a stamp decided or which revision it governs, so status labels have to come from operational archives with their logs intact.
Document the result the way Datasheets for Datasets recommends documenting any dataset, covering its motivation, composition, collection process and recommended uses. For a stamp-status set that means stating each source's legend and version, the normalized mapping, exclusion counts by code, the supersession evidence mix and the agreement figures by field.
Signals that a records specialist, not another annotator, should decide
Bring in a records or document-control specialist when the status depends on rules outside the page. More annotators will not fix a missing legend or an unclear control procedure.
Escalate when a source has no recoverable legend for its codes, when regulated quality records are involved and the controlled-document procedure defines what obsolete means, when a stamp may record a legal act such as notarization, recording or certification rather than a business review, or when pages are under a legal hold. Narrow the scope instead of stretching the guide: a dataset limited to one source's submittals with a known legend is more useful than a mixed archive with guessed codes.
Status stamps also sit on records that belong to more than one company. A reviewer's stamp on a contractor's drawing involves the reviewer, the contractor and often the owner, so who may license the page is a rights question for the holder and counsel, not something the annotation guide can settle.
Requesting stamp-status records with their logs intact
A stamp-status dataset is only as good as the records that travel with the pages: legends, transmittal logs, revision registers and system histories. When you request this kind of data, ask for those linked records by name, along with the normalized mapping, the exclusion counts and the agreement figures described above.
SourceX does not stock stamped archives; each set is tracked down at U.S. businesses after a buyer's request-data form lays out scope, volume, language, format, timing and licensing needs. Because the archive holder signs an agreement fixing scope and terms before handing anything over, linked logs and legends can be written into the scope from the start.
Frequently asked questions
Is a stamp status label the same as a stamp detection label?
No. A detection label marks where a stamp is and what kind of mark it is. A status label records what the stamp decided, what it applies to and whether a later record replaced it, which needs the legend and linked records as well as the image.
Should superseded pages be removed from a document stamp dataset?
Not by default. Keep them, labeled superseded with a link to the replacing record, and let each training task decide. A task about the current approved version excludes them; a task about review history or carried-forward conditions needs them.
What if the code letters on the stamps have no legend?
Keep the letter verbatim and leave the normalized status blank. Look for the legend in the transmittal template, the reviewing office's procedure or the contract specification; if none can be found, exclude the page from status training and record why.
Can OCR read status stamps reliably?
Often, but not always. Stamps are rotated, colored, faded or printed over body text, so keep word-level confidences and send low-confidence status words to a second reader rather than accepting the OCR text as the label.
Do approvals recorded only in a document system count as stamps?
They are valid status evidence but not stamp labels. Store them as record-level labels tied to the system, field and export date, and do not draw a stamp region that the page image does not contain.
Sources
- The U.S. Army Corps of Engineers transmittal form (ENG Form 4025, May 2017) defines action codes including B, approved except as noted with resubmission not required; C, approved except as noted with comments attached; and E, disapproved. Source
- DCMI's isReplacedBy term describes a related resource that supplants, displaces or supersedes the described resource; replaces is its inverse. Source
- The W3C PROV ontology treats a revision as a derivation in which the newer entity is a revised version of an original, and models invalidation of an entity as a separate relation. Source
- Label Studio's Relations tag creates labeled relations between labeled regions and works with image data, among other types. Source
- Tesseract's TSV output lists each recognized word with its bounding box and a confidence value. Source
- DFKI's StaVer dataset consists of 400 document images with pixel-accurate ground truth, built from dummy bills with colored logos, colored text and various kinds of stamp. Source
- The DDI-100 document dataset includes images with stamps, and its ground truth contains the corresponding stamp masks. Source
- Datasheets for Datasets proposes that every dataset be accompanied by a datasheet documenting its motivation, composition, collection process and recommended uses. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.