Turnstone

3D records of cultural objects, from every side.

▶ Try it live — ethical-tech-colab.github.io/turnstone

Runs in any browser, works offline, installs to a home screen.

Capture	Record an object. Coverage dial, one instruction at a time, and a session that survives a reload.
Quality demo	Watch a session score itself, and see what each dimension actually measures.

Every number is computed from actual pixels — nothing is simulated.

---

A tool for photographing a statue, vessel, mask or carved fragment so that the resulting 3D model is worth something afterwards — carrying a machine-readable statement of how good it is and what it may be relied upon for.

A turnstone is a shorebird that feeds by flipping stones over to see what is underneath — the most-skipped part of object capture, and the reason a capture rubric needs angular coverage and surface completeness as separate dimensions at all.

---

Status

✅ Research	92 Tavily queries, licences verified at source (research/)
✅ Concept + decisions	PHOTOGRAMMETRY-CONCEPT.md
✅ Specification	PHOTOGRAMMETRY-SPEC.md — 42 requirements, 14 conformance tests
✅ Phase 0 · rubric as data	rubric/heritage-v1.json — DPA's ten dimensions, transcribed mechanically
✅ Phase 0 · scoring engine	lib/rubric.js — worst-dimension class, progress score, limiting dimension
✅ Phase 0 · orbit geometry	lib/overlap.js — distance cancels, so guidance needs no depth sensor
✅ Phase 1 · metric extractors	lib/metrics.js — the gap DPA names in its own §10
✅ Phase 1 · guidance loop	lib/guidance.js — one instruction at a time, with hysteresis
✅ Phase 1 · live demo	web/ — the extractors running on real pixels
✅ Phase 2 · camera control	lib/camera.js — full-res stills with a fallback chain, optics locking
✅ Phase 2 · persistent sessions	lib/store.js — IndexedDB; frames survive reload, phone call, flat battery
✅ Phase 2 · auto-capture	lib/steady.js — one-shot on stability, so the tap never blurs the shot
✅ Phase 2 · offline client	web/capture.html — installable, boots with no network
⬜ Phase 3	Printable target: scale bar, colour patches, fiducials, masking
⬜ Phase 4	Reconstruction server (Docker, AliceVision)

node --test test/*.test.mjs           # 108 passing
node tools/validate-rubrics.mjs       # rubric files load and score
node tools/demo-probe.mjs             # drives the demo in headless Chrome
node tools/capture-probe.mjs          # drives the capture client, incl. offline

---

What it measures, and how

Dimension	Method	Live?
Sharpness	Variance of the Laplacian, plus a within-session relative rule	✅
Exposure	Clipping fraction per channel, worst reported	✅
Image overlap	Angular: 1 − sin(Δθ/2)/tan(φ/2)	✅
Ground sample distance	From the scale bar, cross-checked against optics	✅
Angular coverage	24 azimuth bins × 4 elevation bands	✅
Surface completeness	Silhouette sectors — an estimate, see OQ-1	~
Lighting consistency	IQR of luma and chromaticity across frames	✅
Scale / colour reference	Target detection	✅
Device metadata	Fraction of frames retaining focal length, sensor, pose	❌ deferred

Six things that are easy to get wrong

Sharpness is content-dependent. A carved, weathered surface outscores a smooth glazed one at identical focus. An absolute threshold alone marks every smooth object blurred; a relative rule alone cannot notice that every frame was soft. Both are required.

The silhouette edge must not leak into a masked measurement. An object against its background is an occlusion boundary — the strongest gradient in the frame, and strong whether or not the surface is in focus. A mask that only tests the centre pixel lets the kernel read across it, so a blurred object scores as sharp, and scores differently against a different backdrop.

Distance cancels in the overlap geometry. Overlap depends only on the angular step and how much of the frame the object fills, so live guidance works with no depth sensor. A 10–15° step gives 64–86% overlap depending on framing, which brackets CIPA's 60–80% for close-range heritage work.

A deferred dimension must not cap a live class. Device metadata cannot be known during a session. Scoring it live reports a contributor as insufficient for a reason they can neither observe nor act on — a post-mortem, not guidance. It is listed as deferred, the class is flagged provisional, and it still gates the final record.

Autofocus invalidates the lens profile. It silently changes the effective focal length, so a session shot with autofocus active has a different camera model in every frame and the reconstruction has no way to know. Locking optics is a metrology feature, not a convenience — and lens_profile_valid is false whenever it could not be done.

A session must survive the device. Frames are written to IndexedDB as they are captured, not accumulated and saved at the end, because "save at the end" makes the failure mode losing everything rather than losing the last frame. The measurements are written before the image: if storage runs out, the record of what was measured survives, and that is the part that cannot be recreated from the object afterwards.

---

Why this exists

Ethical-Tech-CoLab/DPA defines a ten-dimension capture rubric and states plainly in its own §10 that there are "no metric extractors" and "no mobile client". Turnstone is those extractors and that client.

The rubric is data, with every threshold in one file so the argument can be had against specific numbers rather than against a vibe. Turnstone measures against whatever rubric it is handed — DPA supplies heritage-v1, another programme supplies its own, and neither forks this tool.

What it is not

• Not Plumb. Plumb measures flat facades you cannot reach; this handles solid objects you can walk around. Shared philosophy, almost no shared geometry.
• Not a claim of lawful ownership. A flawless capture of a looted object is a flawless capture of a looted object. attests / doesNotAttest are mandatory fields.
• Not an input to provenance confidence. Otherwise a museum raises an object's provenance score by buying a better camera.

Licences

Code Apache-2.0 · docs CC BY 4.0 · rubrics and schemas CC0, so a rubric can be cited, forked and argued with freely.

See CONTRIBUTING.md for the rules that are not up for negotiation, and SECURITY.md for what counts as a security issue in a project that makes claims about objects.