Status and release notes

The version this page runs, how the live site is checked, and what changed in each release. Nothing is sent from this page.

This version

Analysis engine build (SHA-256 of core.wasm): 321de5e40d00314e473d0446fc460dbfd6abbf908f7149119825759a5cf63479

How the live site is checked

Every day a scheduled check loads the live site and confirms its security headers and that every file it serves is the file that was built, and checks the analysis engine's dependencies for published security advisories.

Security advisories

None published. A vulnerability is reported as /.well-known/security.txt describes.

Release notes

The notes are published in English, as the records are.

v0.7.27 (2026-11-10)

  • The camera check now asks every match the question it already asked of a picture tried as a piece cut from the enrolled camera's picture: if the same picture scores nearly as high against the profile flipped into its mirror image, which no camera produces, the result reads inconclusive instead of a match. A picture taken in the dark, with little in it but a few lit areas, can score high against a profile it shares nothing with; on the reference photographs the check withholds none of the right camera's matches, and on a further set with very dark photographs it withholds 53 of 84 matches by other cameras and 7 of the right camera's 65. The error-rate page shows what it changes.

v0.7.26 (2026-11-03)

  • On a computer, the Workbench now runs the camera check at the same time as the checks of the picture's pixels, each on an engine of its own, so an analysis with a camera profile chosen finishes sooner. A phone or tablet still runs them one after the other, within its memory.
  • The signal-to-noise check reads the picture a band of rows at a time, so the picture checks need less memory and run faster; their results are the same, bit for bit.
  • The camera check does the same work with fewer steps in two of its calculations; its results are the same, bit for bit.
  • Checking the Content Credentials of a JPEG picture that carries none, as most photographs do, no longer copies the picture in memory: the page answers that none are present, as the verifier would, without reading the picture again.
  • Every file chooser in the app can now be reached and opened from the keyboard, on the Workbench, the report page, the verification page and in the Sensor Vault. Before, none could: each was hidden in a way that also hid it from the Tab key.
  • Under a dark system theme the pages keep their light colours, so the drop areas and the notes stay readable; before, the drop areas' text could be white on a near-white ground. The drop areas' edges are darker, so they stand out from the page.
  • A screen reader now hears when the work on a page ends: the Workbench's analysis, the signed report, each check on the verification page, and each step in the Sensor Vault, from frames read to a picture tested. The Workbench's overlay and re-save buttons say which one is on.
  • The Workbench's maps use colours that stand apart from their background and from each other, and the areas that could not be measured are hatched instead of grey. Each map is named and described for a screen reader.
  • Each table that scrolls sideways is named after what it holds, so a screen reader moving between them can tell them apart.
  • Each page's title now names the page, in English or Arabic as the page is shown ("Workbench · PixVeritas"). Before, every page was titled "PixVeritas" alone, so a screen reader arriving on a page, a browser tab and the history list could not tell the pages apart.
  • On the verification page, a declared source type given as a long web address now wraps on a narrow or zoomed screen, instead of making the page scroll sideways.
  • The verification page reads Content Credentials only in the formats it offers (JPEG, PNG, WebP, AVIF, HEIC and TIFF pictures, MP4 video and MP3 audio) and says so, in English and Arabic, for any other file before reading it. A .c2pa file, which holds the Content Credentials of another file, is checked against that file when the two are given together.
  • A file given to the verification page's signed-report check that is not a signed report is now said to be one in the page's own words, in English and Arabic, instead of in the verifier's internal error text.
  • Reading a file's Content Credentials, and opening a profile saved before version 2 of the profile format, now take their memory with them when they finish, instead of leaving it with the page for every later file.
  • The Sensor Vault keeps each profile's sealed contents apart from its entry in the list and reads them only when a step needs them, so a page that lists the vault no longer loads every profile in it. A vault stored by an earlier version is moved over the first time a page opens it; an earlier version of PixVeritas opened afterwards in the same browser shows the vault empty until the page is reloaded on this one. Exported .pvault files are unchanged.
  • If the engine fails to start when a page first loads after an update, the page now starts it again by itself, and asks for a reload only if that keeps failing.
  • The report and verification pages show any correction published for the engine build that signed a report, under the report's build line.

v0.7.25 (2026-10-24)

  • An AVIF or HEIC file lists where its parts are stored, and may describe each place in no bytes at all. The engine went through such a list place by place, so a crafted file of a few hundred kilobytes kept the checks of the file's format, its metadata and how it was captured, and the signed report, busy until each stopped at its time limit and read not tested. The engine now reads each part's place once, and such a file is answered at once. The release's own fuzz testing found it.
  • The web framework the site is built with is updated again, to a release that fixes security issues published against it on 24 October. They concern server features this site does not use (image optimisation, cached server functions and server actions), since it serves pages built in advance, so none of them could be reached here.
  • A constant-time helper library under the engine's encryption is updated to the release that fixes a published weakness (RUSTSEC-2026-0385): some of its masks, and some comparisons of 128-bit values on 32-bit platforms such as the browser's engine, could take a time that depends on the values. The engine reaches that library only to check that a profile file, the reference data or the hand-over between its own workers was not altered, through a routine the weakness does not concern, so it could not be reached here.
  • The error-rate page's figures for profiles made through a 1024 × 1024 or 2048 × 2048 window now come from the camera check comparing each test picture as the product does: the whole picture, with what its own file records. Before, each picture was cut to the profile's window and compared at the camera's own scale, including pictures from another camera that the product compares as resized copies. No other camera matched any of these comparisons.
  • The whole-frame figures come from the same comparisons: one reference set reads no false match in 14,268 where it read two, which the measuring tool had made, and the other one in 11,265 where it read none in 11,244. That one is a match the camera check makes, which the measuring tool's trimming of the picture had hidden.
  • One reference set's rate over 20 photographs a camera reads 4.55 % where it read 4.09 % (5.88 % where 5.29 % inside the standard operating conditions): its sample now takes every photograph of one phone into account, those stored turned or at another size too, and so holds others.
  • The strength of the evidence given beside a camera match is recalculated on these comparisons. It moves in its fourth significant figure, or by three at its upper limit.
  • The table of how the camera check copes with resized copies now gives each copy what its original file records, as the camera check itself does: a photograph whose file records a digital zoom is tried at its zoom, and one the camera check would not judge on a miss, such as a photograph of unknown history or one made from several exposures, is counted as not judged. One reference set's half-size row reads 13.76 % where it read 14.09 %, and the other's quarter-size row 60.00 % where it read 60.56 %.
  • For a profile made through a 1024 × 1024 or 2048 × 2048 window, or a whole-frame profile of a camera of up to 4.6 megapixels, a camera match at the camera's own scale is now reported as inconclusive when 83 % or more of its score comes from the finest detail, where the limit was 78 %. Fitted again on the reference sets' training photographs as the camera check compares them, one phone's own ordinary photograph matched with 78 % of its score in the finest detail, and the limit keeps its margin above the camera's own photographs; another camera's pictures that matched that way had 93 % or more.
  • The error-rate page's table for another camera's photograph of the profile's size now counts the comparisons the camera check makes at that size, as it makes them: a photograph stored turned against the profile is left out, since the check compares it turned, on its search. The size thresholds that table fits move to 24.18 and 22.00, still below every camera's decision threshold, so each camera's own threshold applies as before.

v0.7.24 (2026-10-20)

  • If a second picture is opened while the first is still being read, each picture's results always come from its own analysis, on the Workbench and the report page.
  • Text in a file that can reorder or hide what is read, such as a right-to-left override in a signer's name or a file's name, is shown as codes in the summary sentences too, not only in the detail tables.
  • If PixVeritas is updated in another tab, a tab still on the earlier version says so and offers a reload, instead of retrying its analysis without end.
  • A report file whose signature would check out whatever the report said is now shown as not verifying.
  • When a picture's Content Credentials include another picture that declares itself AI-generated in its own credentials, the verification page says that was said of the included picture, instead of presenting it as this file's own signed declaration.
  • Content Credentials from a signer that has been withdrawn, and whose certificate authority also reports it revoked, now read as revoked instead of as a file the verifier could not read.
  • The engine build row on the report and verification pages now says it compares with the engine this page is running, which an installed copy keeps until it is reloaded, rather than the engine the site runs now.
  • Content Credentials signed with a certificate meant only for time-stamping, or one issued under a time-stamping authority, are no longer shown as signed by a trusted signer.
  • Checking Content Credentials can no longer leave a page waiting for good: a check that runs too long is stopped and says so, and the verification page refuses files over 200 MB before reading them.
  • An installed copy of PixVeritas now offers the update when a release changes only the site's security settings, so it takes them up too.
  • The machine-readable analysis record no longer gives calibrated weight to the two checks that measure without concluding, and no longer states a grid size for a check that never chose one.
  • A portrait or multi-frame marker that one phone maker writes no longer changes the camera check on a picture from another maker.
  • Notes on the pages, the error-rate page and the machine-readable analysis record no longer refer to sections of an internal planning document.
  • A signed report now records which Sensor Vault profile its camera check used, by the SHA-256 of the profile's .pvault file, and the pixels its findings read, by their size and SHA-256; the report and verification pages show both. Pixels that are not the size of the file being reported on are refused rather than read.
  • An enrolment whose results were altered between its steps is refused: every figure the engine works out from the measurements, the admission decision included, is worked out again at each step.
  • Level C, source-device consistency, is not offered for now. A camera match stops at Level B, model consistency, until the published error rates show Level C holding for other units of the same camera model and for computational-photography phones. The result says so, and the error-rates page shows where that stands.
  • The web framework the site is built with is updated to a release that fixes security issues published against it. They concern server features this site does not use, since it serves pages built in advance, so none of them could be reached here.
  • The error-rate page also reports the camera check on bright photographs: in the reference sets' bright photographs no other camera matched, and the page says it is a measurement of the photographs those sets hold.
  • A flaw in the compression library the engine uses to open saved profiles is fixed: a crafted profile file, opened with its passphrase, could make the page run out of memory.
  • A camera check now says when the picture comes from another output pipeline than the profile's frames: another file format, encoder settings, size or camera software. A camera's signature carries less from one pipeline to another, so a miss there says less than usual.
  • The Sensor Vault leaves out an enrolment frame that an editing application saved, and says so: some editors add a pattern of their own that can make pictures from different cameras look alike, so a profile made from such files could match another camera's.
  • When a profile's signature repeats itself across the frame (a pattern the phone's processing writes into every picture, as Samsung's Exynos phones do), the Sensor Vault now says so when the enrolment closes: a picture tested against such a profile reads inconclusive, never a match.
  • HEIC pictures are now analysed in browsers that can decode them, with every pixel result marked as decoded by the browser and not reproducible elsewhere.
  • A signed report is now produced for a picture the browser cannot decode, covering the file's structure and metadata, and it says the pixel checks did not run. Before, no report was signed.
  • The Sensor Vault marks results on AVIF and HEIC pictures, and on profiles made from them, as decoded by the browser and not reproducible, and advises taking frames in the phone's JPEG mode.
  • On a phone or tablet the app runs only the checks its browser has the memory for: the Sensor Vault enrols through a 2048 × 2048 or 1024 × 1024 window, the camera check runs against such profiles, and a picture above 16 megapixels gets its file checks with a note to use a desktop computer for the rest.
  • The error-rate page shows how the camera check holds up when a picture is changed: saved again as JPEG down to quality 25, reduced to a quarter of its size, or turned slightly. No other camera matched any changed picture, and the check never reported a changed picture's own camera as not matching: it said it could not test it, or, eight times, that the result was inconclusive (six of them at a quarter of the size, by this release's rule for a score carried by the finest detail).
  • A picture smaller than the enrolled camera's own, enlarged after it was taken and saved without recompression, can score above the threshold against another camera's profile through its finest detail rather than the camera's signature. On that route too, a score carried mostly by the finest detail is now reported as inconclusive, with its share and the scale it was compared at. Copies of the camera's own pictures from messaging apps and social platforms are tested as before; the error-rate page says how many of the camera's own reduced pictures it withholds.
  • The error-rate page's figures for the search that follows resized copies now come from the engine itself, as the product compares the pictures, rather than from the measuring tool's earlier search: on the same comparisons another camera's picture matched once in 12,004 on the reference sets' training devices and never on the held-out ones, and how many of the camera's own WhatsApp and Facebook copies are found now comes from the Vault's own test.
  • The notice before a WhatsApp or Facebook copy is analysed now gives how many such copies the camera check finds over every camera measured: about 6 in 7 WhatsApp copies and 2 in 3 high-quality Facebook copies against a whole-frame profile, measured on copies those platforms made in 2016–17 from photographs of 2014–16 phones, as the notice now says. It had quoted the held-out cameras alone (19 in 20 and 4 in 5). The Sensor Vault's note gives a 1024 × 1024 profile's figures per platform: about a third of WhatsApp copies and a fifth of Facebook ones. The error-rate page no longer says such copies are not tested: they are tested as copies, and a miss reads not tested, never a different camera.
  • The error-rate page now gives the camera check's rates over the pictures inside its standard operating conditions as well as over every test picture, and again with every natural photograph of each validation camera. On the recent phones' reference set, the camera check misses about 11 % of a phone's own photographs at the 1024 × 1024 window, not the 4 % that 20 photographs a camera showed, and the page names the phones the misses gather on.
  • A camera check that reads NO MATCH now says that a camera's own photograph can miss too, more often on some recent phones whose processing leaves little of the camera's signature in a photograph, so a NO MATCH alone does not show that another camera took the picture. The error-rate page gives the camera check's rates for whole-frame profiles with every natural photograph of each validation camera, and names the phones the misses gather on.
  • A camera check that reads MATCH now says how strongly its score supports the enrolled camera having taken the picture rather than another camera: a likelihood ratio, with its order of magnitude and a grade in words, calibrated on the reference sets' validation photographs for each profile size, and "at least" at its upper limit. The error-rate page shows the calibration and how it holds on the reference sets' own comparisons.
  • A Sensor Vault profile saved before 27 September 2026, as a file or in this browser, is now converted inside the engine, so the page never decodes its fingerprint: it reads the file's header and hands the rest to the engine, and wipes its copy of the file once the engine has sealed the profile. A vault last opened before that date is still converted the next time it is unlocked. That conversion will be removed in a later release, no sooner than a year after this one, so open such a vault before then.
  • A picture the camera recorded a digital zoom for is now left out of a Sensor Vault enrolment, with a message: it is the middle of the sensor enlarged, so its signature does not line up with the other frames', and it would weaken the profile. Before, an enrolment of flat fields that held one was refused at its last step, and one of ordinary photos took it in unseen.
  • A Sensor Vault profile now keeps how its camera scored on the ordinary photos it was checked on when it was made, and a camera check that reads NO MATCH against it says how many of them also scored below the threshold. When some did, the camera does not leave enough of its signature in every photograph it takes, and the NO MATCH says even less than usual about which camera took the picture. Photos made from several exposures or in portrait mode are scored but not counted, since a miss on them is never read as NO MATCH. Give 10 to 25 ordinary photos, without large bright areas, when you make a profile. A profile made before this release says that enrolling the camera again adds this count.
  • The camera check's likelihood ratio for profiles enrolled through a 1024 × 1024 or 2048 × 2048 window is calibrated again, on the comparisons the check actually makes at the camera's own scale: another camera's narrower picture is compared at another scale and gets no likelihood ratio, so it no longer counts in the calibration. Its upper limit is lower (about 2,200 rather than 3,900 at 1024 × 1024, and 1,600 rather than 2,200 at 2048 × 2048), and at that limit "at least" is followed by a figure rounded down, never up.
  • Results on pixels the browser decoded with its own HEIC or AVIF decoder, and matches against a profile made from such pictures, carry no likelihood ratio, as results the published rates do not describe.
  • The error-rate page also gives the likelihood ratio of the camera check's decision itself, at each operating point, beside the score's; and it states the one security advisory against earlier releases (the compression library fixed in this release).
  • The error-rate page no longer says its tooling and records are public: they are kept in the project's repository, which is not public yet. It also attributes the inpainting test sets it reports on, and says that Level C needs a reference set showing its miss rate holding as well.
  • A profile's check no longer counts photographs that are copies, or whose history the file cannot tell (no camera data), since the camera check never reads their miss as NO MATCH. The count beside a NO MATCH now starts a sentence of its own.
  • The Sensor Vault leaves out an enrolment frame whose header the engine could not read, with a message, rather than taking it in without checking it for an editor's export or a recorded zoom. Frames cannot be added while a batch is still being read, a profile file over 200 MB is refused before it is read, and a profile converted from the earlier format is listed as not admitted, as the engine treats it.
  • A camera check whose comparison produced no score now reads inconclusive. Before, a profile with values the engine cannot score, which the conversion of an earlier-format file could make, could read NO MATCH; such a profile file is now refused when imported.
  • A camera check against a profile converted from the earlier format no longer records that profile's old quality figures in the signed report as if the engine had measured them.
  • The Workbench no longer says the browser decoded a HEIC picture it could not decode.
  • The report and verification pages say when the pixels a report names were refused because they are not the size of the file's frame, rather than calling them the pixels read.
  • If the analysis engine or the Content Credentials checker fails while reading a file, nothing more runs on it: the checks still to come for that file say they did not run, and a fresh one is started for the next.
  • The Sensor Vault refuses to create or import a profile anywhere but on app.pixveritas.com, where profiles are kept; an existing vault can still be unlocked and exported elsewhere.
  • The report page no longer leaves the signed report in the browser's session storage.
  • A check that finishes within its time budget is no longer stopped because another check on the same page ran long.
  • The error-rate page states as a limit that no check here finds an object a diffusion model has inpainted into a photograph, with its measurement.
  • The error-rate page says that its 1,024 × 1,024 and 2,048 × 2,048 rows compare every picture at the camera's own scale, where the camera check compares another camera's narrower picture on the searched route instead, and counts how many of their false-match comparisons that is.

v0.7.23 (2026-10-11)

  • A picture enlarged or zoomed after it was taken can score above the threshold against another camera's profile, through its finest detail rather than the camera's signature. At the camera's own scale, a score carried mostly by that detail is now reported INCONCLUSIVE, with the reason, instead of MATCH. In the measurements behind it, no photograph as the camera saved it was affected.
  • When a picture's file names no lens, the Vault chooses the phone's profile for the camera that took it by the focal length the file records, which many phones write when they name nothing else.
  • The stability a profile enrolled from ordinary photographs must reach is now measured on recent phones as well as older ones, and falls from 2,776 to 1,491. Recent phones make far weaker profiles from ordinary photographs than from flat fields, so flat fields remain the way to enrol.
  • The Resampling Signature row no longer says a picture was rescaled or rotated "since capture": some phones resample every photograph they save, so the row now says the frame was resampled after capture or by the camera itself.
  • The Resampling Signature row reports a whole-frame resample only from a stronger pattern, measured on recent phones as well as older ones: untouched photographs from recent phones are reported as resampled about a third as often, while about nine in ten of the rescaled pictures found before are still found.
  • The error rates page shows the spatial checks' false alarms on recent phones beside older ones, including on photographs as the phones saved them, and the camera check on the reference sets' dark photographs.
  • Spatial Noise Continuity and Compression Grid Continuity now report what they measured without concluding anything ("measured, not concluded"): on recent phones neither tells edited photographs from untouched ones reliably enough to report a finding either way. The Noise field overlay still shows the noise map.
  • Content Credentials signed by a signer on the frozen interim trust list are now reported as valid under that legacy trust model. They were reported as coming from an unknown signer, because the list was never consulted.
  • When a file's own signer is on that list but a manifest it was made from was signed by an unlisted certificate, the Content Credentials row now says so, instead of saying the file's signer is on no list.
  • When the Compression History row finds no second save, it now says that this is not proof of a single save: in testing it found about one camera original in four saved again at a lower quality, the commonest history of a shared photograph. The error rates page lists its rates for every kind of re-save measured.
  • The pages now enforce Trusted Types, a browser protection: a page can start scripts only from this site's own files, and no text can be inserted into it as markup.
  • The camera check and enrolment use less of the browser's memory on large pictures: from 12 to 48 megapixels the heaviest step needs 7 to 12% less, leaving more of the budget free.
  • Content Credentials whose signer's certificate was revoked, as a response from the certificate's issuer stapled in the file shows, are now reported as revoked. They were reported as a file the check could not read.
  • A JPEG whose data ends early, such as a file cut off in transfer, no longer gets a compression-history reading: the row reports that the file is damaged or truncated. The reading had been made partly over picture data the file does not hold.
  • A picture whose file says an editing application saved it is no longer treated as straight from the camera: its capture class is EDITED EXPORT, and the camera check does not run on it, because some editors add a pattern of their own that can make pictures from different cameras look alike.
  • For an AVIF picture, the results read from its pixels now say that this browser decoded them: another browser or system may decode them differently, so they cannot be reproduced elsewhere.
  • Each signed report now records the build of the analysis engine that produced it (the SHA-256 of the engine file the page loaded, which the browser checks before running it), and the report and verification pages show it.
  • Pictures of up to 100 megapixels are analysed, among them the full-resolution photographs of 64-megapixel phone cameras, which were refused above 64 megapixels; above 24 megapixels the camera check and enrolment still compare the middle 24 megapixels of the picture.
  • The report and verification pages now say whether the engine build a report records is the one this site runs, a build of an earlier release in the signed engine build registry, or an unknown build.
  • Content Credentials signed in a Nikon Z6III with one of the certificates Nikon withdrew in September 2025 now read as revoked, not as valid under the older trust list.
  • Content Credentials signed by a signer on the C2PA Trust List now read as valid and trusted; signers that are only on the older interim list still read as valid under the legacy trust model, and the verification page names the dated set of trust lists it used.
  • Photographs taken in portrait mode on Samsung, Huawei and Apple phones, and iPhone HDR photographs, are no longer reported as not matching the enrolled camera when the camera check falls short: the blur and the merging the phone applies weaken the signature the check relies on, so such a result now reads not tested. A match still counts.
  • The Duplicated Region Search now also finds an area copied within a picture after it was turned, resized or mirrored, and finds more of the areas copied unchanged. The row says how far the copy was turned and resized, or that it is a mirror image. In testing it found about seven in ten turned copies and six in ten resized or mirrored ones, and reported a copy on about four in a thousand untouched photographs.
  • The Duplicated Region Search now finds smaller copied areas: in testing, two in three areas covering 1% of the picture, where it found about half before, with no more untouched photographs reported. The error rates page shows how often it finds copies of each size.
  • The Duplicated Region Search no longer reports a scene's own mirror symmetry, such as a symmetric building front, as an area copied as its mirror image. In testing this removed about a third of its false reports on untouched photographs, without missing any copy it had found.
  • The Duplicated Region Search now says what a match is consistent with: an area copied within the picture, or an object that appears twice in the scene, such as two copies of the same sign.
  • The Workbench, the signed report and each analysis record now say what PixVeritas does not examine: what the picture shows, such as its lighting, shadows and perspective, which checking whether a picture is authentic also needs. They also say that one picture can hide an edit, so related pictures should be examined too.
  • The Duplicated Region Search no longer reports a repeating pattern matched onto itself, such as a row of windows or a tiled floor, as an area copied within the picture. In testing this removed most of its remaining false reports on untouched photographs from one of the two reference sets, without missing any copy, including copies pasted over part of the area they were copied from.
  • The Compression Grid Continuity row measures in larger areas, 192 pixels square instead of 128, chosen again on clean data: in testing its map marked more of the pasted areas and fewer untouched photographs. The row still reports what it measured without concluding anything, and no longer measures pictures smaller than 576 pixels on their short side.
  • The Sensor Vault gives the precautions for enrolling a device that is itself part of a case, and records the device's system version and the capture settings you enter with its profile.
  • The links on the error rates archive page are now large enough to tap easily.
  • A damaged or deliberately malformed PNG or WebP file no longer stops the analysis in the browser, and a malformed report file no longer stops the verification page: they are reported as files the checks cannot read.
  • The Workbench's Quantization Variance Heatmap always belongs to the file on screen: a heatmap still being drawn for the previous file is no longer shown under the next one.
  • Importing a profile file from an early version of the Sensor Vault no longer risks making the Vault page fail to open, and such a profile is shown as not admitted, as the camera check treats it.

v0.7.22 (2026-10-07)

  • Pictures above 24 megapixels, such as a 48 MP phone photograph, are analysed within the browser's memory budget: the camera check compares the middle 24 megapixels of the picture and says so, and a 48 MP camera can be enrolled in the browser.
  • A picture enlarged after it was taken, or zoomed by the camera without recording the zoom, is no longer reported as not matching its own camera: the result is NOT TESTED, with the reason.
  • Every error message names its code (PV-E and four characters), and the Error codes page lists them all, in English and Arabic.
  • The Workbench offers "Copy diagnostic report": stage timings, status codes, the picture's size and memory, copied to the clipboard and never sent.
  • The analysis engine checks itself on a built-in test picture when it starts; a report is offered only when that check passes, and every signed report records it.
  • The error rates are published on the site, with the reference sets' attribution and the signing-key fingerprints.
  • Every page shows an egress meter, which counts what the page sends and should read zero, and the analysis engine's build hash.
  • The Status page shows this version, how the live site is checked every day, and these notes.
  • On Safari, a return visit to the Workbench starts the analysis engine again; once the installed app was active it could stop at "reference data unavailable".
  • Characters in a file's text that are invisible or change the direction of text are shown as codes such as <U+202E>, so they cannot change what is read.
  • A picture larger than 64 megapixels is refused before the browser decodes it, and says so; a JPEG with more than 100 progressive scans is not read for its compression history.
  • Cameras that write their names in Exif 3.0's UTF-8 text are read as naming the camera.
  • Easier to read and to use from the keyboard: grey text is darker, and every table that scrolls sideways can be scrolled from the keyboard.
  • Each release's error-rate page is kept as it was published, under "Error-rate pages of earlier releases".
  • A browser that cannot run the analysis is told what it lacks and which browsers work.
  • Signing keys can now be withdrawn from a date, so a leaked key cannot sign reference data the engine accepts.

v0.7.21 (2026-10-06)

  • The reference data is signed on a hardware security key kept offline; nothing the build service holds can sign data the engine accepts.

v0.7.20 (2026-10-05)

  • A photograph the camera zoomed digitally is checked at its zoom, and a miss is NOT TESTED rather than "no match". This also closes a case where a zoomed picture could match another camera.
  • A picture is tested against the profile made with the lens that took it, when the camera's entry has one.
  • A new enrolment is compared with the other profiles in the Vault, which says when two profiles share a pattern.

v0.7.19 (2026-10-04)

  • Copies from current messaging and social apps get the camera check, which reports only a match; square and 4:5 crops of a photograph are followed.
  • The notice for messenger copies says what was measured and how to ask for the original.

v0.7.18 (2026-10-03)

  • Enrolment checks each frame against the others, and Level C needs that agreement. Profiles enrolled earlier answer at Level B until the camera is enrolled again.
  • The Vault notices a picture taken with another of the camera's lenses.

v0.7.17 (2026-10-03)

  • The Redistribution Trace: which platform's settings the last save of a JPEG matches, if any.
  • The interface in English and Arabic.
  • The app can be installed and works offline.

v0.7.16 (2026-10-03)

  • A faster engine build. Signed reports name the right engine version again; reports signed by v0.7.15 named an older one.

v0.7.15 (2026-10-03)

  • A picture checked as a crop must also score clearly above the profile's mirror image, which removes matches driven by textured scenes.

v0.7.14 (2026-10-02)

  • The camera check scores at the alignment the picture's geometry implies, and its thresholds follow the profile's size: fewer matches against another camera's photographs with whole-frame profiles.

v0.7.13 (2026-10-02)

  • The threshold for rescaled copies was corrected after a measurement found another camera's smaller photographs matching too often.

v0.7.12 (2026-10-02)

  • A picture cut from a camera's frame without resizing is checked as a crop of it.

v0.7.11 (2026-10-01)

  • A shape the camera check cannot follow is reported as NOT TESTED instead of being searched over scales, a search that had produced matches only against other cameras.

v0.7.10 (2026-10-01)

  • The local re-save check needs a larger region before it reports one. The camera check's calibration was measured again on evenly spread samples.

v0.7.9 (2026-09-29)

  • Result tables fit a phone screen.

v0.7.8 (2026-09-29)

  • The reference data was rebuilt without any device kept back for the final evaluation.

v0.7.7 (2026-09-29)

  • A Samsung photograph fused from several exposures can still match, but a miss is never reported as "no match".

v0.7.6 (2026-09-29)

  • Clearer wording for the affected Samsung phones.

v0.7.5 (2026-09-28)

  • Samsung phones whose own processing writes a shared pattern: a would-be match is reported as inconclusive rather than attributed to the phone.

v0.7.4 (2026-09-28)

  • A crafted JPEG can no longer exhaust memory in the compression-history check.

v0.7.3 (2026-09-27)

  • Updated reference data from two reference sets.

v0.7.2 (2026-09-27)

  • Spatial Integrity on the Workbench: the noise map, the resize trace and the duplicated-region search.

v0.5.0 (2026-09-15)

  • The Sensor Vault: enrol a camera you own in the browser, and test whether a photograph came from it.

v0.1.3 (2026-09-14)

  • The first public deployment.

v0.1.0 (2026-09-14)

  • Content Credentials, metadata, encoding history and a signed analysis report, all in the browser.

Analysis engine build (SHA-256 of core.wasm): 321de5e40d00314e473d0446fc460dbfd6abbf908f7149119825759a5cf63479