Algorithm Updates

Changelog

Latest updates and improvements to HTPBE?

31.10.2026Latestv2.48.3

Corrected one check that could report an unmodified document as modified. Some documents that an earlier release reported as modified can now be reported as inconclusive. Results issued before this release are kept as they were issued and are not re-assessed; to obtain a result under this release, re-submit the document. A finding describes the file’s history – not the document’s content, and not the person who submitted it.
This is a correctness change, not a measured detection improvement, and the evaluation corpus does not cover every document. An inconclusive result is not a clean result: it does not certify a document as an untouched original. No public marker identifier was added, renamed or removed – the marker contract is unchanged.

30.10.2026v2.48.2

Corrected how some documents are assessed. Some documents that an earlier release reported as modified can now be reported as intact or inconclusive. If a result issued before this release matters to you, re-submit the document. A finding describes the file’s history – not the document’s content, and not the person who submitted it.
Some documents that an earlier release reported as intact can now be reported as inconclusive: they are of a kind whose results cannot be confirmed as intact. An inconclusive result is not a clean result: it does not certify a document as an untouched original.
Switched off one part of the analysis that could report an unmodified original as modified. Some documents that an earlier release reported as modified can now be reported as intact or inconclusive. Results issued before this release are kept as they were issued and are not re-assessed; to obtain a result under this release, re-submit the document.
Corrected further cases affecting how some documents are assessed. Some documents that an earlier release reported as modified can now be reported as intact or inconclusive. Results issued before this release are kept as they were issued and are not re-assessed; to obtain a result under this release, re-submit the document.
These are correctness changes, not a measured detection improvement, and the evaluation corpus does not cover every document. No public marker identifier was added, renamed or removed – the marker contract is unchanged.

29.10.2026v2.48.1

Broadened an existing check. Some documents that an earlier release reported as intact can now be reported as modified, with a finding of a kind already listed in the public catalogue of checks; the verdict only ever moves to a stricter result. If a result issued before this release matters to you, re-submit the document. A finding describes the file’s history – not the document’s content, and not the person who submitted it.
Measured across the evaluation corpus, the change reached only a small number of documents, all of them already reported as modified. Little independent evaluation evidence stands behind the broadened check, so this change is not presented as a measured detection improvement, and the corpus does not cover every document. An inconclusive result is not a clean result: it does not certify a document as an untouched original. No public marker identifier was added, renamed or removed – the marker contract is unchanged.

28.10.2026v2.48.0

Broadened an existing check. Some documents that an earlier release reported as intact or inconclusive can now be reported as modified, with a finding of a kind already listed in the public catalogue of checks; the verdict only ever moves to a stricter result. If a result issued before this release matters to you, re-submit the document. A finding describes the file’s history – not the document’s content, and not the person who submitted it.
This change is not presented as a measured detection improvement, and the evaluation corpus does not cover every document. An inconclusive result is not a clean result: it does not certify a document as an untouched original. No public marker identifier was added, renamed or removed – the marker contract is unchanged.

27.10.2026v2.47.1

Corrected how some documents are assessed. Their findings and verdict can differ from those an earlier release reported; a document an earlier release reported as modified can now be reported as intact, and a document reported as intact can now be reported as modified or inconclusive. If a result issued before this release matters to you, re-submit the document. A finding describes the file’s history – not the document’s content, and not the person who submitted it.
This is a correctness change, not a measured detection improvement. The evaluation corpus does not cover every document, so the change described above can still reach documents outside it. An inconclusive result is not a clean result: it does not certify a document as an untouched original. No public marker identifier was added, renamed or removed – the marker contract is unchanged.

18.10.2026v2.47.0

Some documents can return a result that differs from the one an earlier release returned: the findings listed with the verdict, the confidence level reported with it, or the status_reason accompanying an inconclusive verdict can each differ, and the verdict can differ with them. If a result issued before this release matters to you, re-submit the document. A finding describes the file’s history – not the document’s content, and not the person who submitted it.
The analysis is now more robust against malformed files. Some documents that an earlier release could not analyse, ending in an error, now return a result. Some files whose structure cannot be read now end in the existing invalid_pdf error instead; no result is stored and no credit or quota is used. An error is not a result: it certifies nothing about the document, so treat a document that could not be analysed as unverified, not as clean.
An analysis that cannot be completed in full is never counted as clean. It either ends in the existing invalid_pdf error, with no result stored and nothing charged, or returns a result that is never intact: modified where the evidence supports it, otherwise inconclusive with status_reason analysis_incomplete. Some genuine documents that an earlier release reported as intact can now come back inconclusive.
Corrected several cases affecting how some documents are assessed. Their findings and verdict can differ from those an earlier release reported, in either direction. An inconclusive result is not a clean result: it does not certify a document as an untouched original.
Measured across the full evaluation corpus against version 2.44.7, the last release in service before this one, a small number of documents changed their verdict, in either direction. Among documents still reported as modified, some list different findings – most of them fewer – and some report a different confidence level, nearly all of them lower. Some documents that were already inconclusive now carry status_reason analysis_incomplete instead of another value. The corpus does not cover every document, so the changes described above can still reach documents outside it. An inconclusive result is not a clean result: it does not certify a document as an untouched original.
API contract change: an analysis that does not finish within its time limit now ends with the new error code analysis_timeout instead of invalid_pdf. The HTTP status stays 422 and the error body keeps its shape (error, code, details); the code value and its message are what change. No result is stored and no credit or quota is used. Unlike invalid_pdf, it can succeed on retry: wait a minute and retry with backoff. An integration that recognised a timeout by the invalid_pdf code needs to handle analysis_timeout instead, and any integration should treat an error code it does not recognise as a failed request rather than relying on a fixed list of codes. Like any error, a timeout certifies nothing about the document: if a document keeps returning it, treat it as unverified, not as clean.
API contract change: xref_count in GET /api/v1/result/{id}, and has_incremental_updates in both GET /api/v1/result/{id} and GET /api/v1/checks, can now be null. Null means the value could not be established for that document, or was not recorded for an older check; an earlier release returned 0 or false in those cases, and results of earlier checks read after this release can show null too. An integration that does not accept null for these fields needs to. A null value is not a finding and says nothing about whether the document was changed.
Requests with a URL containing credentials are now rejected with 400 invalid_url_format; 5xx error bodies include a request_id; 429 server_at_capacity is also returned when an account already holds its share of the analysis queue. See the migration guide.
Apart from the API contract changes above, these are correctness and robustness changes, not a measured detection improvement. Compared with version 2.44.7, two status_reason values, visible_signature_origin and analysis_incomplete, can now be returned alongside an inconclusive verdict; an inconclusive result is not a clean result. No public marker identifier was added, renamed or removed – the marker contract is unchanged.

08.10.2026v2.46.8

A result returned as inconclusive with status_reason analysis_incomplete now uses a credit or quota like any other delivered result. Submitting a file again is a new check and is charged under the same rules; reading an existing result again, or replaying a request with its original Idempotency-Key, is not.
An analysis that a fault on our side stops from completing ends in the existing invalid_pdf error, ‘Analysis was interrupted’. No result is stored and no credit or quota is used.
Some documents that an earlier release could not analyse, ending in an error, now return a result. Where the analysis cannot be completed in full, the result is never intact: modified where the evidence supports it, otherwise inconclusive with status_reason analysis_incomplete. An inconclusive result is not a clean result: it does not certify a document as an untouched original. Either result is charged like any other delivered result.
Some documents can return a result that differs from the one an earlier release returned: the verdict, the findings listed with it, or the status_reason accompanying an inconclusive verdict can each differ. If a result issued before this release matters to you, re-submit the document.
These are billing, correctness and reporting changes, not a measured detection improvement. No new status_reason value was introduced, and no public marker identifier was added, renamed or removed – the marker contract is unchanged.
The creator and producer values returned with a result can now be shortened copies. Two new fields, creator_truncated and producer_truncated, are true when the returned value is the beginning of a longer one. Only the returned copy is shortened.

07.10.2026v2.46.7

Some documents can return a result that differs from the one an earlier release returned: the verdict, the findings listed with it, or the status_reason accompanying an inconclusive verdict can each differ. If a result issued before this release matters to you, re-submit the document. A finding describes the file’s history – not the document’s content, and not the person who submitted it.
Measured across the full evaluation corpus, every document kept the verdict, the findings and the confidence level it had, and no document was newly certified as an untouched original. The corpus does not cover every document, so the change described above can still reach documents outside it. An inconclusive result is not a clean result: it does not certify a document as an untouched original. These changes are not presented as a measured detection improvement. No new status_reason value was introduced, and no public marker identifier was added, renamed or removed – the marker contract is unchanged.

06.10.2026v2.46.6

Some documents can return a result that differs from the one an earlier release returned: the verdict, the confidence level reported with it, the findings listed with it, or the status_reason accompanying an inconclusive verdict can each differ. If a result issued before this release matters to you, re-submit the document. A finding describes the file’s history – not the document’s content, and not the person who submitted it.
Some documents that an earlier release reported as inconclusive now come back modified. A modified verdict reports that the file was changed after it was first produced – a document whose content is genuine can still receive it, and it says nothing about the person who submitted it.
Some documents that an earlier release reported as untouched originals can now come back with an inconclusive verdict instead – a result that reports the question could not be answered for that document. An inconclusive result is not a clean result: it does not certify a document as an untouched original. If a result issued before this release matters to you, re-submit the document.
Corrected a case where a document could wrongly be left reported as inconclusive. Such a document can now receive the verdict its analysis supports, which can be intact. This does not mean every inconclusive result issued before this release would now change.
Corrected cases where some documents could be assessed incorrectly. Such a document can now return a different verdict, as its analysis supports.
Corrected a further case. The findings a document reports can differ from those an earlier release reported, and its verdict can differ with them.
Measured across the full evaluation corpus, a very small number of documents changed their verdict. One document already reported as modified no longer lists one of its findings. No other document changed its verdict or its findings; in particular, no document in the corpus moved away from an intact verdict. The corpus does not cover every document, so the changes described above can still reach documents outside it. An inconclusive result is not a clean result: it does not certify a document as an untouched original. These changes are not presented as a measured detection improvement. No new status_reason value was introduced, and no public marker identifier was added, renamed or removed – the marker contract is unchanged.
Many forensic checks now carry reworded public descriptions. This rewording on its own changes no identifier, no API field and no result.

02.10.2026v2.46.5

Some documents can now report a finding that an earlier release did not report, and the verdict can move with it – only ever to a stricter result. If a result issued before this release matters to you, re-submit the document. A finding describes the file’s history – not the document’s content, and not the person who submitted it.
Measured across the full evaluation corpus, every document already in it kept the verdict and the confidence level it had, and no document was newly certified as an untouched original; a very small number of documents already reported as modified now list one further finding of a kind they could already carry. An inconclusive result is not a clean result: it does not certify a document as an untouched original. The change broadens an existing check and is not a measured detection improvement. No new status_reason value was introduced, and no public marker identifier was added, renamed or removed – the marker contract is unchanged.

01.10.2026v2.46.4

Some documents can return a result that differs from the one an earlier release returned: the verdict, the findings listed with it, or the status_reason accompanying an inconclusive verdict can each differ. If a result issued before this release matters to you, re-submit the document. A finding describes the file’s history – not the document’s content, and not the person who submitted it.
Measured across the full evaluation corpus, every document already in it kept the verdict, the findings and the confidence level it had, and no document was newly certified as an untouched original. An inconclusive result is not a clean result: it does not certify a document as an untouched original. This is a correctness fix, not a measured detection improvement. No new status_reason value was introduced, and no public marker identifier was added, renamed or removed – the marker contract is unchanged.

29.09.2026v2.46.3

Some documents now return a result that differs from the one an earlier release returned: the verdict, the findings listed with it, or the status_reason accompanying an inconclusive verdict can each differ. If a result issued before this release matters to you, re-submit the document. A finding describes the file’s history – not the document’s content, and not the person who submitted it.
Measured across the full evaluation corpus, a very small number of documents return a different verdict and every verdict that changed moved to a stricter result; no document was newly certified as an untouched original. A document that is itself genuine can also come back modified where an earlier release did not report it that way. This is a correctness fix, not a measured detection improvement. No new status_reason value was introduced, and no public marker identifier was added, renamed or removed – the marker contract is unchanged.

28.09.2026v2.46.2

Some documents now return a result that differs from the one an earlier release returned: the verdict, the findings listed with it, or the status_reason accompanying an inconclusive verdict can each differ. If a result issued before this release matters to you, re-submit the document. A finding describes the file’s history – not the document’s content, and not the person who submitted it.
Measured across the full evaluation corpus, no document was newly certified as an untouched original. A very small number return a different verdict, a very small number no longer report one of the findings they reported before, and a very small number carry a different status_reason; an inconclusive result is not a clean result and does not certify a document as an untouched original. This is a correctness fix, not a measured detection improvement. No new status_reason value was introduced, and no public marker identifier was added, renamed or removed – the marker contract is unchanged.

25.09.2026v2.46.1

Some documents can now report findings that an earlier release did not report. If a result issued before this release matters to you, re-submit the document. A finding describes the file’s history – not the document’s content, and not the person who submitted it.
Measured across the full evaluation corpus, no document was newly certified as an untouched original and no verdict changed; the change in reported findings stayed confined to a very small part of it. This is a correctness fix, not a measured detection improvement. No new status_reason value was introduced, and no public marker identifier was added, renamed or removed – the marker contract is unchanged.

23.09.2026v2.45.2–2.46.0

A document carrying a signature can now report findings that an earlier release did not report, and can come back modified or inconclusive where an earlier release reported it as unmodified. If a result issued before this release matters to you, re-submit the document. A finding describes the file’s history – not the document’s content, not the validity of any signature or the identity of a signer, and not the person who submitted it.
A further status_reason value, visible_signature_origin, can now be returned alongside an inconclusive verdict. An inconclusive result is not a clean result: it does not certify a document as an untouched original. Clients that branch on status_reason should keep a default branch for values they do not recognise; the full set is listed in the API reference.
Corrected cases where a document could be reported as modified although the available evidence did not support that finding. If a modified result issued before this release matters to you, re-submit the document.
A document whose structure cannot be read is now rejected as an unprocessable file rather than failing with a server error. The existing error code is used and no new one was introduced; no result is stored and no credit is consumed.
Measured across the full evaluation corpus, no document was newly certified as an untouched original, and a small number no longer carry one of the markers they previously reported. These are correctness changes, not a measured detection improvement. No public marker identifier was added, renamed or removed – the marker contract is unchanged.

21.09.2026v2.45.1

Results that record an incomplete check are now returned more consistently. A result of this kind is never a clean one: it states that the document was not fully checked, and it does not certify the document as unmodified. If a clean result issued before this release matters to you, re-submit the document.
No new detection was added in this release, and no public marker identifier was added, renamed or removed – the marker contract is unchanged. No new status_reason value was introduced either.
Measured across the full evaluation corpus, every document already in it kept the verdict and findings it had. This is a change to how an unfinished check is reported, not a measured detection improvement.

20.09.2026v2.45.0

Some documents that previously returned an error with no result now return a result instead. Such a result is never a clean one – it records that the check could not be completed in full, and it does not certify the document as unmodified. Treat it as a document that was not fully checked rather than one that passed, and do not rely on it; a document that returned an error before this release can be re-submitted.
A further status_reason value, analysis_incomplete, can now be returned alongside an inconclusive verdict. No new detection was added in this release, and no public marker identifier was added, renamed or removed – the marker contract is unchanged. Clients that branch on status_reason should keep a default branch for values they do not recognise.
Measured across the full evaluation corpus, every document already in it kept the verdict and findings it had. This is a change to how an unfinished check is reported, not a measured detection improvement.

18.09.2026v2.44.8

Documents changed after they were signed are now reported as modified in cases where an earlier release reported them as unmodified. Affected documents can also carry named markers they did not report before. This was confirmed on internal test documents; it is a correctness change, not a measured improvement, and no document already in the evaluation corpus changed verdict.
A genuine document can also be reported as modified where an earlier release reported it as unmodified. If a result issued before this release matters to you, re-check the document. A modified result describes the file’s history – not the document’s content, and not the person who submitted it.
No public marker identifier was added, renamed or removed in this release – the marker contract is unchanged.

15.09.2026v2.44.7

Corrected a case where a document could be reported with a finding the available evidence did not support.
Measured across the full evaluation corpus, every document already in it kept the verdict and findings it had.

05.09.2026v2.44.6

Extended origin handling to a further class of documents. Documents in this class are no longer certified as untouched originals; they can still be reported as modified. This is a design change, not a measured improvement.
Measured across the full evaluation corpus, every document already in it kept the verdict it had, and none was newly reported as modified.
Two forensic checks now carry public descriptions that state more precisely what each check reports. Their identifiers, the API contract and detection behaviour are unchanged: the same documents are flagged, with the same markers and the same verdicts.
No public marker identifier was added, renamed or removed in this release – the marker contract is unchanged.

27.08.2026v2.44.5

Corrected a defect that prevented one forensic check from reporting a concern the available evidence did support. A small number of documents now carry that concern among their named markers, alongside the markers they already reported. Measured across the full evaluation corpus, no document changed verdict.
No public marker identifier was added, renamed or removed in this release – the marker contract is unchanged.

26.08.2026v2.44.4

Corrected a defect that let one forensic check report a concern the available evidence did not support. A small number of documents no longer carry that concern among their named markers.
Measured across the full evaluation corpus, the change stayed confined to that same small group, each document in it still reported as modified on other, independent grounds. No document changed verdict as a result.
No public marker identifier was added, renamed or removed in this release – the marker contract is unchanged.

14.08.2026v2.44.3

Continued tuning of the forensic checks so that a document is reported as modified only where the evidence supports that conclusion. A small number of documents previously reported as modified now return inconclusive. An inconclusive result is not a clean result: such documents can still be reported as modified, and it never certifies a document as an untouched original.
Measured across the full evaluation corpus, that reclassification stayed confined to the small group described above; no other document changed verdict as a result of it.
A further class of documents whose origin cannot be established firmly enough to certify them as an untouched original now routes to human review instead of returning a clean, untouched-original result. This is a tightening – no document becomes easier to certify.
No public marker identifier was added, renamed or removed in this release – the marker contract is unchanged.

13.08.2026v2.44.2

Corrected a case where a document could be reported as modified without evidence of an edit. Such documents are still analysed in full.
Refined one forensic check so that it no longer reports a concern the available evidence cannot support. Measured across the full evaluation corpus, no document changed verdict as a result.
No public marker identifier was added, renamed or removed in this release – the marker contract is unchanged.

12.08.2026v2.44.1

Fixed a defect in how previously stored results were returned: a document checked in an earlier release could come back with fewer of the named markers it was issued with. Those results again report exactly what they reported when they were issued. Analysis of newly submitted documents was never affected.
Three forensic checks now carry public descriptions that state more precisely what each check reports. Their identifiers, the API contract and detection behaviour are unchanged: the same documents are flagged, with the same markers and the same verdicts.
Corrected the example API responses shown on this site so that every example matches a response the API can actually return, including the field that explains what to do with a result. The API itself is unchanged – no integration needs updating.
No public marker identifier was added, renamed or removed in this release – the marker contract is unchanged.

10.08.2026v2.44.0

Broadened detection of documents that were re-worked after they were first produced and previously passed as unmodified. Such documents are now reported as modified.
Extended origin handling to a further class of documents whose origin cannot be established firmly enough to certify them as an untouched original: files of that class that previously returned a clean, untouched-original result now return inconclusive and route to human review. An inconclusive result is not a clean result: such documents can still be reported as modified.
No public marker identifier was added, renamed or removed in this release – the marker contract is unchanged.

09.08.2026v2.43.2

Fewer genuine documents of a certain type are mistakenly reported as modified. Documents of that type that were altered after creation are still reported as modified, and detection of that class has been broadened.
Corrected a defect that could report a document as modified without sufficient basis. Fewer documents are affected this way.

05.08.2026v2.43.1

Verdicts shown in list, dashboard and summary views now match the verdict returned for the same document on its own. Some documents that previously appeared in a list as an untouched original, or as modified, now show there the inconclusive result they already returned when retrieved individually, and some older records that carry no stored result now show no verdict rather than a clean one. Individual document results are unchanged, and detection itself is unchanged.

31.07.2026v2.43.0

Added a new check that catches a further class of documents whose stated origin does not hold up. Such documents are now reported as modified – a class that previously passed as unmodified or inconclusive.
Extended origin handling to a further class of documents whose origin cannot be established firmly enough to certify them as an untouched original: some files that previously returned a clean, untouched-original result now correctly return inconclusive and route to human review. An inconclusive result is not a clean result.
Broadened detection of documents whose pages were assembled from multiple separate sources, catching a further class that previously passed as unmodified or inconclusive. Such documents are now reported as modified: we do not certify a multi-source document as an untouched original. Existing verdicts across our regression corpus were unchanged by this broadening.

30.07.2026v2.42.1

Fewer genuine documents of a certain type are mistakenly reported as modified.
Modified verdicts that the evidence does not fully support now report at a lower confidence level instead of as a certain result.
Fewer genuine digitally-signed documents are mistakenly reported as modified.

29.07.2026v2.42.0

Added a new check that catches documents whose page content was substituted after the document was created. Such files are now flagged as modified – a class that previously passed as unmodified.
Reduced over-certification on a class of documents whose origin cannot be established firmly enough to certify them as an untouched original: some files that previously returned a clean, untouched-original result now correctly return inconclusive. An inconclusive result is not a clean result.

28.07.2026v2.40.0–2.41.0

Added a new check that catches a further class of fabricated documents. Such documents are now flagged as modified – a class that previously passed as unmodified. Existing verdicts across our regression corpus were unchanged by this new check.
Reduced false positives on a class of genuine documents that were occasionally reported as modified. Their verdict is never raised to unmodified.
Broadened detection of documents that were altered after their original creation, and of documents that merely reproduce a document rather than being the authentic original file. In each case a further sub-class that previously returned an inconclusive result is now reported as modified.

27.07.2026v2.39.0

Added a new check that catches documents whose stated scan origin is not genuine. Such documents are now flagged as modified – a class that previously passed as unmodified.
Added a new check that catches documents that were altered after their claimed original creation even though the file is presented as an untouched original. Such documents are now flagged as modified – a class that previously passed as unmodified.

23.07.2026v2.38.3

Reduced false positives on a class of legitimately produced documents that were occasionally reported as modified without evidence of an actual edit. Documents in this class are no longer reported as modified on that basis. They are not certified as untouched originals.

22.07.2026v2.38.2

Reduced false positives on a class of legitimately digitally-signed documents that were occasionally reported as modified even though their content had not been changed. Documents in this class are no longer reported as modified on that basis.
Strengthened detection of genuine post-creation content edits, so that more documents whose figures or text were substituted after the original was created are correctly flagged as modified rather than left inconclusive. Existing verdicts across our regression corpus were otherwise unchanged.

21.07.2026v2.38.1

Broadened detection of documents whose visible content was altered after the original was produced, including altered financial statements. A range of such documents that previously passed as unmodified are now reported as modified, regardless of the intent behind the edit.
Reduced false positives on a class of legitimate documents that were occasionally reported as modified without evidence of an actual edit. Documents in this class are no longer reported as modified on that basis.
Reduced false positives on a further class of legitimate documents that were occasionally misread. Documents in this class are no longer reported as modified on that basis.

20.07.2026v2.38.0

Added a new check that catches documents whose stated scanner origin is not genuine. Such documents are now flagged as modified.
Reduced false positives on a class of genuine documents that were previously flagged as modified. These are now returned as inconclusive rather than flagged, and are never reported as untouched originals.

19.07.2026v2.37.5

Reduced false positives on a class of legitimately produced documents that were reported as modified without evidence of an actual edit. Documents in this class are no longer reported as modified on that basis. They are not certified as untouched originals.
Extended the same relief to another class of legitimately produced documents that were reported as modified without evidence of an actual edit. They are not certified as untouched originals.

18.07.2026v2.37.4

Reduced false positives on a class of legitimately generated documents that were reported as modified without evidence of an actual edit. Documents in this class are no longer reported as modified on that basis. They are not certified as untouched originals.

17.07.2026v2.37.3

Some documents whose origin cannot be reliably established were being treated as untouched originals. They are now returned as inconclusive instead of passing clean. An inconclusive result is not a clean result.
Closed a narrow case where a document carrying strong evidence of tampering could still be under-reported. Such documents are now flagged as modified.
Reduced false positives on a class of genuine documents that were reported as modified without evidence of an actual edit. Documents in this class are no longer reported as modified on that basis. They are not certified as untouched originals.

16.07.2026v2.37.2

Reduced false positives on a class of genuine documents that were reported as modified without evidence of an actual edit. Documents in this class are no longer reported as modified on that basis. They are not certified as untouched originals.

15.07.2026v2.37.1

Closed a route where a document carrying strong evidence of tampering could be under-reported. When such hard evidence is present, the document is now flagged as modified.

14.07.2026v2.37.0

Added a new detection that catches a further class of documents which were edited after they were first produced and present an untouched origin they do not actually have. Documents in this class are now flagged as modified.
Reduced false positives on a class of legitimately-generated documents that were previously flagged in error. Documents in this class no longer receive a false modified flag. No document passes as an untouched original on this basis alone.

13.07.2026v2.36.1

Reduced false positives on a legitimate class of documents that were reported as modified without evidence of an actual edit. Documents in this class are no longer reported as modified on that basis. They are not certified as untouched originals.

12.07.2026v2.36.0

Extended detection of a class of documents that were altered after issuing, closing a laundering route that previously let some edited files pass as originals. Documents in this class are now flagged as modified.

10.07.2026v2.35.0

Added a new detection for a further class of documents whose visible content was altered after the original was produced. Documents in this class are now flagged as modified.

09.07.2026v2.34.4

Refined origin handling so a class of legitimate documents is no longer automatically held at inconclusive. Such documents are not certified as untouched originals on this basis.

07.07.2026v2.34.3

Broadened detection to a further class of documents that were edited and then made to look like a scan. Files in this class are now correctly flagged as modified.
Refined how the result report names a class of form-based documents so each is labelled accurately. This is a reporting-clarity improvement only – no document changes verdict as a result.

06.07.2026v2.34.2

Tightened origin handling for a further class of documents that cannot be certified as untouched originals – these now correctly settle at inconclusive rather than passing as untouched originals. An inconclusive result is not a clean result.

05.07.2026v2.34.1

Tightened origin handling for a further class of documents whose origin cannot be established firmly enough to certify them as untouched originals. Such files previously passed as untouched originals and now correctly cap at inconclusive.

04.07.2026v2.34.0

Added a new detection for documents whose recorded origin was falsified after the file was changed following its original creation. Files in this class previously passed as untouched originals and are now correctly flagged as modified.
Fixed a false positive in our signature analysis on a class of correctly signed documents that were previously reported as modified.
Extended an existing check to catch a further class of scanned documents whose original values were replaced. This class previously passed as an untouched original and is now correctly flagged as modified.
Added a new detection for a class of documents whose declared origin does not hold up – indicating it was rebuilt and re-labelled after its original creation. Files in this class previously passed as untouched originals and are now correctly flagged as modified.
Reduced false positives on a class of legitimate statements that were previously flagged in error.

03.07.2026v2.32.0

Improved classification accuracy on scanned and image-based documents, cutting false modification flags on a class of legitimate exports that were previously misread as edited.
Recovered detection on a further class of documents that were altered after their original creation – closing a case where some such files previously slipped through.

02.07.2026v2.31.3

Improved accuracy on a class of legitimate invoices and statements whose normal design was previously misread. These now correctly pass.
Reduced false positives on a class of legitimate documents that were previously misread as assembled from multiple sources. These now correctly pass.
Tightened origin handling for a further class of documents whose origin cannot be established firmly enough to certify them as untouched originals. Such files previously passed as untouched originals and now correctly cap at inconclusive.
Reviewed the published check catalogue so every published entry maps to one distinct outcome. No document changes verdict as a result; this is a clarity pass on how our coverage is presented, not a change to what we detect.
Reduced false positives on a class of legitimate documents produced by some generators, whose ordinary output was previously misread as evidence of editing. These now correctly pass.

01.07.2026v2.31.2

Reduced false positives in signature analysis on a class of genuine digitally-signed documents that were previously reported as modified.
Tightened origin handling for a further class of documents that cannot be certified as untouched originals – these now correctly settle at inconclusive rather than passing as originals. An inconclusive result is not a clean result.

30.06.2026v2.31.1

Broadened detection to catch a further class of documents that were rebuilt after their original creation but still passed off as untouched originals. Files in this class previously slipped through and are now correctly flagged as modified.
Tightened origin handling for a further class of documents whose origin cannot be established firmly enough to certify them as untouched originals – these now correctly cap at inconclusive instead of passing as untouched originals.

29.06.2026v2.31.0

New detection class for a further class of documents that were digitally rendered and then presented as a scan. Such files previously passed as inconclusive and are now correctly recognised as modified.

27.06.2026v2.30.0

Expanded detection of documents that were edited after their original creation – closing a case where some such files previously slipped through.
Recognised a further class of scanned documents that previously passed as untouched originals – they now return inconclusive.
Made verdicts consistent across servers, so the same document no longer receives a different verdict depending on which server analysed it.
A further class of documents whose origin cannot be firmly established now returns inconclusive instead of passing as an untouched original.
Recovered detection on a class of statements that were rebuilt after the issuer originally produced them. These are now correctly flagged as modified.

24.06.2026v2.29.0

New detection class for a further class of documents that were reworked in an editor after their original creation. Files reworked this way previously passed as untouched; they are now recognised as edited after the fact, surfaced under the same edited-in-an-editor outcome as our other content-editing checks.
Reduced false positives on a class of genuine bills and statements that were previously flagged as modified and now pass.
Broadened our edited-in-an-editor detection to recover a further class of documents reworked after their original creation. This class is surfaced under the existing edited-in-an-editor outcome.
Reduced false positives on a class of genuine financial statements that were wrongly flagged as modified.

23.06.2026v2.28.0

Added a new proprietary integrity signal to the detection suite. We are publicly acknowledging that this signal exists, but we are deliberately not disclosing how it works – the method is held back on purpose so that forgers cannot read a description of it and engineer their way around it. When it fires, it contributes evidence that the file is a modified derivative rather than an untouched original.

22.06.2026v2.27.1

Reduced false positives on a class of genuine documents that were wrongly flagged for concealed content.

19.06.2026v2.27.0

Reduced false positives on a class of genuine forms that were previously mistaken for modified files. Forms of this kind are no longer reported as modified on that basis.
New detection class for a further class of reassembled documents.
New detection class for documents whose recorded origin was rewritten after the original authoring.

18.06.2026v2.26.1

Broadened detection of documents assembled by merging pages from more than one source file – recovering a class of merged documents that previously slipped through certified as a single original, including documents with extra pages glued on after the fact.

17.06.2026v2.26.0

Retired a check that was producing false positives on genuine documents without improving coverage – outcomes previously reported by it remain covered.
New detection class for a further class of documents that were re-written after their original creation and presented as untouched originals. Such documents are reported as modified.
Restored detection of documents assembled by merging pages from more than one source file, including merges that were subsequently rebuilt to look like a single original.
New detection class for documents whose page content was substituted after the original was produced to alter what the page appears to say.

16.06.2026v2.25.3

Reduced false positives on a class of genuine documents that were previously being misread as an imitation of a scan – authentic scans of this kind now correctly settle at the not-certifiable scan-origin ceiling instead of being flagged as modified.

15.06.2026v2.25.2

Widened the class of documents whose origin cannot be established firmly enough to certify them as unmodified; such files no longer pass as clean.
Recovered scanned-document recognition on a class of scanned files that were previously slipping through certified as clean – a scanned document can never be confirmed unmodified, so this class now correctly returns the not-certifiable ceiling rather than an intact verdict.

13.06.2026v2.25.1

Reduced false positives in our generator-origin check on a class of legitimate documents.

10.06.2026v2.25.0

New detection class for documents that misrepresent which tool produced them – flagging origin that was rewritten.
Broadened generator-identity-forgery detection to a further class of documents whose stated origin does not hold up.
Reduced false positives on a class of legitimate documents.

09.06.2026v2.24.2

Reduced false positives in our metadata analysis on a class of legitimate documents.
Reduced false positives in our template-assembly detection on a class of legitimate documents that were previously misread as a pieced-together template.

05.06.2026v2.24.1

Reduced false positives in our template-assembly detection on a class of legitimate documents that were previously being mistaken for pieced-together templates.
Refined how the page-assembly analysis reads the document so a legitimate layout is no longer misread as reused template scaffolding.

04.06.2026v2.24.0

New detection class for a kind of document reassembled from an original.
Hardened structural-integrity checks on rebuilt documents.

27.05.2026v2.23.3

Recovered tamper detection on a class of edited documents that previously slipped through.
Reduced false positives on a class of legitimate documents that were previously mistaken for tampered files.

26.05.2026v2.23.2

Reduced false positives on a class of legitimate documents that were previously mistaken for tampered files.
Hardened analysis so a class of legitimate files is no longer misread as tampered.
Broadened the class of documents whose origin cannot be established firmly enough to certify them as unmodified.

23.05.2026v2.23.1

Withheld the integrity guarantee on a class of documents whose origin cannot be corroborated.
Brought the algorithm into line with our existing position that some document origins cannot, on their own, support certifying a file as unmodified.
Withheld the integrity guarantee on a further class of outputs whose origin cannot be corroborated.

22.05.2026v2.23.0

New detection class for a further class of documents rebuilt and presented as a genuine original. Files in this class are now flagged as modified.
Hardened this check so the claimed origin of a document alone no longer certifies it. Files whose claimed origin does not hold up are correctly flagged as modified rather than passing as untouched originals.

21.05.2026v2.22.1

Reduced false positives on a class of genuine documents that were reported as modified without evidence of an actual edit. Documents in this class are no longer reported as modified on that basis. They are not certified as untouched originals.
Extended the same relief to a further class of genuine documents. They are not certified as untouched originals.

20.05.2026v2.22.0

New detection class for a further pattern characteristic of non-genuine documents
A further class of documents whose origin cannot be corroborated is no longer reported as an untouched original
Hardened detection against reassembled documents that misrepresent their origin
Recognised a further class of scanned documents – recovers correct origin classification for them
New detection class for synthetic edit-history patterns

19.05.2026v2.21.1

Withheld the integrity guarantee on a class of documents whose authenticity cannot be corroborated.
Recovered correct origin classification on a class of legitimately-generated documents that the previous heuristic was over-flagging.

18.05.2026v2.21.0

Added detection of a further class of documents that were edited and then re-issued – such documents are now reported as modified
Added a detector for documents whose declared origin does not hold up
Broadened detection of documents whose page content was substituted after the original was produced
Recovered a structural check that was previously reporting too little
Removed detector branches that did not produce unique findings – outcomes previously reported by them remain covered
Improved handling of digitally signed documents so edits made after signing are reported reliably
Recovered detection on classes of documents whose recorded origin does not hold up, and which previously passed as unmodified
Reduced false positives on an additional class of legitimate authoring pipelines
Reduced false positives on a further class of reports.

17.05.2026v2.19.3

Reduced false positives on a class of documents
Reduced false positives on a further class of reports
Reduced false positives on a further class of statements

16.05.2026v2.18.1–2.19.2

Recovered detection on further classes of edited documents
Broadened coverage so more documents can no longer be certified as untouched originals
Reduced false flags on genuine documents, including by retiring checks whose outcomes remain covered by other checks

15.05.2026v2.18.0

Added detection of documents whose earlier edit history was erased
Added detection of programmatically generated pages presented as a captured scan
Added detection of a further class of documents with content added after creation
Improved origin classification accuracy for a further class of image-derived documents

14.05.2026v2.17.1

Improved coverage of editing-tool analysis on additional document layouts
Reduced false positives in embedded image analysis for certain rendering pipelines

13.05.2026v2.17.0

Added detection of a further class of pseudo-scanned documents
Reduced false positives on a class of image-based documents
Added detection of a further class of rebuilt documents
Improved origin classification for a further class of generators
Improved scan classification in additional ambiguous cases

12.05.2026v2.16.0

Reduced false positives in metadata analysis on a further class of documents
Reduced false positives on a class of machine-printed labels
Added detection of a further class of edited form-based documents
Added detection of documents that carry content added after the original was produced
Reduced false positives on a further class of generators
Reduced false positives on genuine machine-generated documents whose normal output was previously misread as evidence of an edit
Further reduced false positives in metadata analysis on a class of generated reports

11.05.2026v2.15.1

Reduced false positives in metadata analysis on a class of generated reports
Broadened metadata-tampering detection to cover an additional class of tools
Reduced false positives on a class of statements

09.05.2026v2.15.0

Added a new metadata check
Improved analysis precision to reduce false positives on legitimate documents
Reduced false positives on documents produced by certain processing tools

08.05.2026v2.14.0

Expanded multi-source assembly detection across additional structural layers
Reduced false positives on a class of generators
Strengthened synthetic-scan detection against an additional class of forgeries
Added detection of a further metadata anomaly
Added detection of additional post-processing patterns
Added detection of additional tampering patterns
Added detection of further fraud-kit metadata fingerprints
Added additional integrity checks

07.05.2026v2.13.25

Tightened verdict semantics for additional editor-workflow patterns
Reduced false positives on certain authoring tools
Improved origin classification for documents touched by an additional PDF editor
Strengthened synthetic-scan detection
Reduced false positives on additional library variants

06.05.2026v2.13.24

Improved metadata-date parsing for documents from certain authoring tools
Improved reliability of metadata extraction for documents from certain processing tools
Broadened detection of documents whose page content was substituted after the original was produced

05.05.2026v2.13.23

Added new detections targeting additional classes of post-creation content modification
Strengthened metadata analysis
Improved multi-source assembly detection
Reduced false positives in metadata analysis on documents from certain generators
Reduced false positives in metadata analysis on outputs from a further class of generators
Detection extended to documents that have passed through certain PDF editing and conversion tools
Reduced false positives on a class of genuine digitally-signed documents
Detection extended to programmatic PDF processing libraries
Reduced false positives in metadata analysis on a further class of documents
Added detection of additional multi-source page assembly patterns

04.05.2026v2.13.20–2.13.22

Improved accuracy of page-content analysis on further classes of documents
Added detection of a further document-fabrication pattern
Reduced false flags on further classes of genuine documents

03.05.2026v2.13.19

Added detection of a further page-content alteration pattern
Reduced false positives in mixed-origin page-assembly analysis on additional rendering pipelines
Reporting no longer lists overlapping findings for the same issue
Reduced false positives on additional rendering pipelines
Improved multi-source assembly detection to cover a previously-undetected pattern
Extended origin coverage to additional page-assembly software

01.05.2026v2.13.18

Added detection of additional metadata-tampering tool markers
Added detection of additional post-processing tool artifacts
Added detection of additional script-injection patterns
Reduced false positives in structural page-content analysis on exports from certain rendering pipelines
Reduced false positives in date analysis
Reduced false positives on a further class of legitimate documents
Reduced false positives on a further class of legitimate exports
Refined verdict semantics for documents whose state cannot be authenticated structurally

30.04.2026v2.13.17

Reduced false positives in structural page-content analysis on a further class of legitimate documents
Reduced false positives on documents from certain generators
Reduced false positives in metadata analysis on documents from certain document-processing pipelines
Reduced false positives on a class of genuine completed and signed documents

29.04.2026v2.13.16

Reduced false positives in multi-source page assembly detection on documents from certain rendering pipelines
Reduced false positives on a further class of legitimate documents
Reduced false positives on a class of generated reports
Improved accuracy of signature analysis on genuine documents that were previously misjudged
Improved accuracy of structural page-content analysis on a further class of documents
Improved accuracy of hidden-content analysis for a further class of documents
Improved origin classification for documents from additional form-editing platforms

28.04.2026v2.13.15

Refined verdict semantics for a class of documents whose structural origin cannot establish authenticity on its own – the API response now includes actionable guidance to verify such documents with the issuing organisation

27.04.2026v2.13.14

Reduced false positives on a further class of genuine digitally-signed documents
Reduced false positives in multi-source page assembly detection for certain generator output
Reduced false positives on certain generator output

26.04.2026v2.13.13

Reduced false positives on documents produced by a class of conversion pipelines
Reduced false positives for documents produced by an additional generator whose normal output was previously misidentified

25.04.2026v2.13.12

Expanded recognition of tools used for document re-processing
Improved generator-origin detection
Added a further metadata check
Reduced false positives in scan classification for certain generators
Expanded detection to cover an additional family of PDF editing tools – documents edited by tools in this family are now correctly identified as modified

24.04.2026v2.13.10–2.13.11

Expanded detection of documents assembled from more than one source, of changes made after creation, and of timestamp anomalies
Added detection of a further redaction-related pattern, and expanded detection of changes that do not alter how a document looks
Further classes of documents now correctly return an inconclusive result – documents in one of those classes can no longer be certified as untouched originals
Reduced false flags on a further class of genuine documents

23.04.2026v2.13.9

Recognised a further class of scanned documents – such documents are now correctly reported as inconclusive
Reduced false positives on a further class of legitimate documents

22.04.2026v2.13.8

Reduced false positives for documents from certain generators, which no longer incorrectly signal post-creation modification
Reduced false positives on a class of programmatically generated documents whose normal output was misread as evidence of post-creation editing
Improved robustness of origin detection on certain page layouts
Reduced false positives on a class of documents that were reported as modified without evidence of an actual edit. They are not certified as untouched originals
Reduced false positives on a further class of documents

21.04.2026v2.13.7

Reduced false positives in metadata analysis on documents from certain generators
Extended the above fix to related generators
Reduced false positives in metadata analysis for certain generators
Improved coverage of structural page-content analysis on a further class of documents
Extended detection of documents rebuilt from prior templates

20.04.2026v2.13.6

Significantly expanded software recognition – documents from a broader range of software now correctly return inconclusive instead of intact
Added detection for additional document editing tools
Improved our internal quality-sampling process

19.04.2026v2.13.5

Improved detection of documents assembled from pages of different origins – more cases are now correctly identified as modified
Improved detection of a class of documents pieced together from multiple images
Improved detection of documents rebuilt from scratch by editing tools

17.04.2026v2.13.4

Improved classification of scanned documents
Improved detection of documents with evidence of post-creation text editing
Improved detection of documents where identifying metadata has been replaced

16.04.2026v2.13.3

Refined verdict semantics for a class of documents with insufficient evidence for analysis
Corrected origin classification for a class of documents that were previously held at inconclusive without cause
Improved detection of additional metadata-tampering patterns
Improved detection of further selective metadata edits

15.04.2026v2.13.2

Reduced false positives in structural page-content analysis on a class of legitimate exports
Reduced false positives in structural page-content analysis on a class of legitimate documents
Reduced false positives in metadata analysis on a further class of documents
Reduced false positives for documents generated by certain generation pipelines

14.04.2026v2.13.1

Reduced false positives in structural page-content analysis on a class of documents

12.04.2026v2.13.0

Added structural detection of scanned documents
Added detection of a class of hidden content
Added detection of content edited in a document editor after initial generation
Added detection of byte-level structural manipulation in document files
Added detection of an additional class of post-modification structural anomalies

11.04.2026v2.12.0

Added detection of image substitution in scanned documents

10.04.2026v2.11.9

Improved detection of a class of generated documents
Improved parsing of non-standard date formats in document metadata

08.04.2026v2.11.8

Improved tool identity checks

08.04.2026v2.11.7

Expanded software recognition

08.04.2026v2.11.6

Improved detection of metadata anomalies

06.04.2026v2.11.5

Reduced false positives for a class of documents
API is now available at the dedicated subdomain api.htpbe.tech/v1
The previous base URL (htpbe.tech/api/v1) continues to work – no migration required

03.04.2026v2.11.4

Expanded software recognition

02.04.2026v2.11.3

A class of documents is now correctly classified by origin

01.04.2026v2.11.2

Reduced false positives on a class of legitimate generators

01.04.2026v2.11.1

Expanded software recognition – a class of documents now correctly returns inconclusive instead of intact

31.03.2026v2.11.0

Added detection of a further class of metadata date anomalies
Added detection of a further class of post-creation edits
Improved detection reliability for modern PDF formats
Broadened analysis of a class of digitally-signed documents
Reduced false positives on legitimately signed documents
Documents we cannot open for analysis now return a clear error instead of an unreliable result
Previously analyzed files may benefit from re-analysis

26.03.2026v2.10.0

Corrected origin classification for a class of documents that were previously held at inconclusive without cause
Improved origin classification for a class of generation pipelines
Previously analyzed files from affected pipelines may benefit from re-analysis

26.03.2026v2.9.0

Introduced detection of documents assembled from pages of more than one origin
Documents whose pages come from more than one source are now flagged as modified
Previously analyzed files may benefit from re-analysis

25.03.2026v2.8.0

Added detection of an additional modification pattern
Previously analyzed files may benefit from re-analysis

25.03.2026v2.7.0

Added detection of an additional mixed-origin page-assembly forgery pattern
Includes guards against false positives on genuine documents
Previously analyzed files may benefit from re-analysis

24.03.2026v2.6.1

Improved scan classification reliability – fixed a false-negative where certain scanned PDFs were not returned as inconclusive.

24.03.2026v2.6.0

Introduced detection of an additional class of document assembly
Previously analyzed files may benefit from re-analysis

22.03.2026v2.5.3

Reduced false positives for a class of documents

17.03.2026v2.5.2

Reduced false positives for a class of documents

17.03.2026v2.5.1

Improved parser robustness so a class of multi-page documents is analysed correctly instead of passing as intact
Previously analyzed assembled documents may benefit from re-analysis

16.03.2026v2.5.0

Improved parser robustness so a class of documents is analysed correctly instead of bypassing analysis
Added detection of an additional PDF editing tool
Introduced detection of documents assembled from multiple independently imported pages
Previously analyzed files may benefit from re-analysis

15.03.2026v2.4.0–2.4.2

Fixed false ‘scanned document’ classification for a class of modern PDFs
Removed a false-positive basis in image analysis – affected PDFs are no longer incorrectly flagged
Recovered signature analysis on a class of documents that previously passed as unmodified
Expanded origin recognition to include a class of editors
Previously analyzed files may benefit from re-analysis

15.03.2026v2.3.0

Improved analysis consistency: all detection signals are now evaluated uniformly
Structural anomalies are now included in the analysis output
Reduced false positives for a class of PDFs
Previously analyzed files may benefit from re-analysis

12.03.2026v2.2.1

Fixed detection reliability for a class of modern PDFs
Resolved edge cases in document parsing
Previously analyzed files may benefit from re-analysis

11.03.2026v2.2.0

Introduced detection of template-assembly document forgeries
Expanded coverage to identify a class of composite documents
Previously analyzed files may benefit from re-analysis

10.03.2026v2.1.7–2.1.8

Improved detection accuracy: fixed rare false negatives where a real modification marker was missed
Improved post-signature tampering detection
Fixed false positives for a class of PDFs

08.03.2026v2.1.6

Added ‘Inconclusive’ result for a class of PDFs
New origin detection: API now returns origin.type and origin.software fields
New primary status field in API: ‘intact’, ‘modified’, or ‘inconclusive’
Result page now shows a grey ‘Inconclusive’ badge and explanation for these PDFs

25.02.2026v2.1.5

Improved detection accuracy for a class of documents
Previously analyzed files may benefit from re-analysis

17.02.2026v2.1.2–2.1.4

Improved detection accuracy for a class of documents
Reduced false positives for a class of PDF creation tools
Previously analyzed files may benefit from re-analysis

16.02.2026v2.1.1

Improved compatibility with modern PDF formats
Enhanced verification for digitally signed documents; significantly reduced false positives on legitimately signed PDFs

12.02.2026v2.1.0

Redesigned the detection engine with a new approach to identifying document modifications
Improved accuracy with a more robust analysis method
Introduced detection of a set of PDF editing tools
Previously analyzed files may benefit from re-analysis

11.02.2026v2.0.4

Fixed timezone bug in date display (dates no longer shown in the future)
Improved UTC timestamp handling for accurate relative time display

09.02.2026v2.0.3

Fixed false positive bug in PDF modification detection
Improved accuracy for metadata analysis

08.02.2026v2.0.2

Added user dashboard with API key management
Implemented passwordless authentication (Google, GitHub, Magic Links)
Created billing and subscription management interface

07.02.2026v2.0.1

Enhanced API infrastructure with monthly quota management
Improved typography and visual consistency
Updated documentation for PDF metadata analysis

06.02.2026v2.0.0

Major platform update – rebuilt application infrastructure
Migrated to Turso database with Drizzle ORM for improved reliability
Comprehensive UX improvements across the application
Enhanced error handling and stability

11.10.2024v1.0.0

Initial public release
Core PDF analysis features