loupe · a worked example for archivists · September 2026

One photograph, and its provenance

A photographer's finished look, carried in an open form, from a real archive, with its fidelity measured. This page walks the three points of a comment on FADGI's provenance draft (TCR4CAP v0.2) beside the pictures and records that back them. The files themselves, checksummed, are in the repository: github.com/streamrD/one-photograph-and-its-provenance.

What is in the bag

One photograph, every form it took

Six files under data/, in the order they came into being. The comment calls the first "the capture" and the last "the work". Every file has its SHA-256 fingerprint in manifest-sha256.txt.

L1000787.RWL
The capture. Raw sensor data in Leica's private format.
the camera · 23 Nov 2009, 15:54:55
L1000787.camera.jpg
The camera maker's own rendering of the same exposure.
the camera's firmware · same moment
L1000787.dng
The same sensor data in the open container. Names its source inside.
Capture One 4 · 25 Nov 2009
L1000787.RWL.cos
Capture One's settings: a five-star rating, everything else at default. Chosen here, finished elsewhere.
Capture One 4 · 25 Nov 2009
L1000787.settings.xmp
The finishing settings, verbatim, as Photoshop wrote them into the finished picture.
Camera Raw 4.6 in Photoshop CS3 · 28 Nov 2009
L1000787.jpg
The finished picture. Saved 28 Nov 2009, untouched since. Byte for byte the file loupe shows.
the photographer · 28 Nov 2009, 20:24:48

Beside them: provenance/ holds the PREMIS record and the C2PA-signed copy of the finished picture; look/ holds the documented development, the color table in CLF and .cube, the fidelity report and a comparison picture. README.md and LICENSE.md (CC BY 4.0) sit at the top.

Point 1

The record cannot tell an artist's finishing from a tampering

What the comment says
An archive that cannot tell them apart will one day treat an artist's print as a tampered file.
Two very different things produce exactly that record. A photographer developing their own raw file into the picture they intend, which is the making of the work. And a later hand, human or AI, altering a finished picture.

Proposed change: a row for the relationship of the agent to the object, and a rule that institutions name which file is the work.

What the bag holds
The camera's own JPEG: darker, cooler, the fog gray-green
The camera's picture. L1000787.camera.jpg. The maker's rendering. Not the work.
The photographer's finished picture: lighter, warmer, the fog pale and open
The work. L1000787.jpg. What the photographer made from the same capture.
The draft's record, for either
PREMIS eventType image modification C2PA action c2pa.adjustedColor softwareAgent Adobe Camera Raw 4.6 source → result L1000787.dng → L1000787.jpg

This is the draft's crosswalk example applied to the finishing. It would read the same for an AI that altered the finished picture later. Nothing in it says the maker made it.

The bag's record
premis.xml eventType creation linkingAgentRole creator (Todd Stabley) linkingObjectRole source: L1000787.dng · outcome: L1000787.jpg significantProperties L1000787.jpg = "THE WORK" c2pa.jpg action c2pa.adjustedColor description "…the making of the work by its creator, not a modification of it by another party." CreativeWork.author Todd Stabley ingredient L1000787.dng (parentOf)
Point 2

Record the how, not only that it happened

What the comment says
The crosswalk records that color was adjusted, when, and by what. It does not record how, so nobody can repeat or check it.
A settings file's meaning is bound to the program that wrote it, so the rendered result must be kept as pixels as well.

Proposed change: name the minimum: the settings file by hash, the software and version, and an open transform where one exists.

What the bag holds
The settings, from the finished picture's own metadata
L1000787.settings.xmp crs:Version 4.6 · crs:RawFileName L1000787.dng WhiteBalance As Shot · Tint +40 · Exposure +0.50 · Brightness +50 · Contrast +25 ToneCurve "Medium Contrast" · HueAdjustmentYellow +18 · HueAdjustmentGreen +23 Sharpness 25 · ColorNoiseReduction 25 · HasCrop False sha256 5b1f0701…0fae6f33 software Adobe Camera Raw 4.6 in Photoshop CS3 (Windows)

Anyone can read these numbers. Only Camera Raw 4.6 knows what "Brightness +50" did to a pixel, and the versions since define it differently. That is the comment's second sentence, in one file.

The open transform, and what it does
The base development: flat and dark, color uncorrected
The base. The raw developed by the written-down free method in look/L1000787.develop.json.
The base with the open color table applied: light and warm, nearly the finished picture
Base + table. look/L1000787.clf applied. Free software, no Adobe file.
The photographer's finished picture
The work. What the table was measured against.
13.7
mean color difference, base alone vs the work
2.5
base + table vs the work, every pixel at full size
1.6
the same with texture set aside (a 2 px blur); 1.0 is about the smallest difference an eye notices

These are the figures in look/L1000787.fit-report.json. What remains between 1.6 and the pixels is the 2009 sharpening and noise reduction, which is texture, not color.

At 100 %: where the difference lives
Detail of the porch balusters in the base development
Base.
Detail of the porch balusters, base with table applied: same color as the work, softer edges
Base + table. The color is there; the edges are softer.
Detail of the porch balusters in the finished picture: crisper edges from the 2009 sharpening
The work. Crisper from Camera Raw's sharpening.
A map of the remaining difference: pale over the fog and lawn, darker along branches, balusters and in the shadows
The remaining difference, mapped. Pale where base + table matches the work, darker where it differs. It gathers along the branches and balusters and in the shadows, which is where sharpening and noise reduction act. Over the fog and the lawn, the color is the work's.
Point 3

A born-digital photograph as a worked example

What the comment says
A proprietary camera raw, a sidecar of settings, and the photographer's rendering, which is the work. The archive has to keep all three and record the relationship between them.
FADGI's own Technical Guidelines note that keeping raw and adjusted versions "would need additional metadata elements". These are those elements.

Proposed change: a second worked example in their appendix, raw + sidecar → rendering, by the creator.

What the bag holds

This bag is that example. The three things the comment names, and how the two records tie them:

The comment namesIn the bagPREMIS saysC2PA says
a proprietary camera rawL1000787.RWL (and its open twin, L1000787.dng)object · is source of the DNG; the DNG is source of the workthe DNG is the parent ingredient of the signed picture
a sidecar of settingsL1000787.settings.xmp (+ L1000787.RWL.cos)outcome of the creation event; in its detail extension by hash, with the softwarein the action's parameters, by hash, with the software
the photographer's rendering, which is the workL1000787.jpgoutcome of creation by the creator; significant property "THE WORK"the signed asset; author Todd Stabley
the additional metadata elementsprovenance/premis.xml, provenance/L1000787.c2pa.jpg, look/which file is the work · who the agent was to it · the settings by hash with software and version · an open transform with its measured fidelity · the finished pixels kept

The dates are the files' own: capture 23 November 2009 at 15:54:55, conversion 25 November, finishing 28 November at 20:24:48, all local time. The finished picture's metadata date and the file's last-modified time agree, which is how the readme can say "untouched since".

Sentence by sentence

Every claim, and where to check it

The comment claimsWhere to check it
The crosswalk's example is "adjusted color in a TIFF" with those PREMIS and C2PA termsThe draft itself, Appendix, first row after the header✓
PREMIS carries the agent's role in linkingAgentRolePREMIS 3 dictionary; used in premis.xml on every event✓
C2PA carries digitalSourceType, IPTC's vocabularyC2PA 2.4 specification; the bag's manifest definition uses it✓
The Library's Recommended Formats Statement asks for a file "equal to the master copy"Its 2025–2026 edition, section A, "Faithful representation"✓
The Common LUT Format is SMPTE ST 2136-1:2026, published this February as a free standardpub.smpte.org, approved 26 February 2026; the bag's look/L1000787.clf is one✓
C2PA's action parameters is an open map; PREMIS has eventDetailExtensionBoth specifications; both used in the bag✓
A settings file's meaning is bound to the program that wrote itL1000787.settings.xmp: Camera Raw 4.6's vocabulary, seventeen years old✓
FADGI's Technical Guidelines say raw and adjusted versions "would need additional metadata elements"Third edition (2023), the raw-files passage✓
The bag holds every file from the raw to the finished picture, both records, the CLF transform with measured fidelity, checksummed, with a readme naming the workThe repository, 23 files, verified against the drive✓

Two things the readme says plainly: the C2PA signature is a test certificate, and the color table carries color and tone, not texture or local edits. Both are in the readme's "What is not claimed".

The files

github.com/streamrD/one-photograph-and-its-provenance

Public since the evening of 26 September 2026. Opens on the readme. The photograph and everything with it: © 2009–2026 Todd Stabley, CC BY 4.0. Contact in the bag: todd@loupe.pictures.