Research

What the evidence supports, and what it warns about

Sākṣī claims that ordinary visitors, given a fixed viewpoint and an honest instrument, can produce a monitoring record a conservator will use. That breaks into four separable claims, and the literature answers each of them differently.

Read the research and market reportPDF · 5 pages · 37 sources

Foundations

The method is inherited, not invented

Repeat photography, returning to a fixed viewpoint and taking the same frame again, has been used in landscape and heritage science for well over a century, and there is now a heritage-specific review literature covering the whole chain: image registration, archival organisation, presentation of the comparison, and the conservation of rephotographic collections in their own right.

The question every institution asks first is whether non-experts produce data worth having. It has been answered quantitatively, most durably by the Snapshot Serengeti analysis: aggregated volunteer classifications agreed with expert-verified data on 98% of images, and 90% of images were correctly classified with just five volunteers each. Accuracy fell for rare cases, however, which carried higher false-positive and false-negative rates.

98%
Volunteer consensus agreement with expert-verified dataSnapshot Serengeti
5
Volunteers per image needed for 90% correct classification
2M+
Registered volunteers across the Zooniverse platform
~100k
Volunteer classifications submitted per day, platform-wide

Three design consequences follow, and all three are in the product: redundancy at the vantage, because several observers per viewpoint beat one expert; uncertainty carried on the record, because volunteer data is reliable in aggregate and has to be able to say when it is not; and attention to the rare case, because the unusual condition, the one actually worth reporting, is exactly where non-expert accuracy is weakest, and therefore the one place a custodian’s confirmation is mandatory rather than optional.

Recent work

Where the technical risk actually is

The area

Damage detection on heritage surfaces
A 2026 systematic review covers 26 papers from 2020–2025; attention-augmented U-Net variants and the YOLO family are the settled architectures, with on-device deployments reported at Italian and Cappadocian sites. The open problem is generalisation across materials: brick, coursed stone, lime plaster and eroded sandstone do not share a crack morphology. That is the strongest argument for the rule that the detector proposes and a person disposes.
Digital twins and HBIM
A scoping review of 204 studies finds the field settling on laser scanning, UAV photogrammetry, BIM and sensor networks: a capital-intensive stack that produces exquisite geometry at intervals of years. Sākṣī is the sampling layer beneath a twin, not a competitor to it: dense, dated, pose-tagged visitor photographs are exactly what photogrammetry-driven change monitoring consumes.
Grounded generation
Faithfulness became measurable between 2024 and 2026: RAGTruth supplies ~18,000 span-labelled examples, and public leaderboards now track hallucination rates directly. The uncomfortable finding is that models still introduce unsupported statements even with the correct context in front of them, which is why retrieving is not the same as being grounded, and why the verify-then-refuse step is not optional.
Running models on a phone
Reviews of mobile inference converge on the same conclusion the build reached in practice, and models in the 1–4 billion parameter range are now credible for constrained tasks, which is exactly the task assigned here: rephrasing a retrieved passage, never asserting a fact.

The field

Four neighbours, and the gap between them

There is no direct competitor. There are four well-populated adjacent categories, each of which owns one link in the chain that runs from a visitor's attention, through positioned evidence, to institutional action, and each of which stops before the next.

Photo-monitoring citizen science

Chronolog · rePhotoSA · Zooniverse · iNaturalist

Visitors really do produce usable image series at scale.

The gapEnvironmental rather than heritage, aligned by a fixed bracket rather than a sensor, and nothing downstream: the output is a time-lapse, not a report that reaches whoever is responsible for the site.

Heritage inventory platforms

Arches (Getty Conservation Institute and World Monuments Fund) · CollectiveAccess · CollectionSpace

The institutional record of a site lives here, and Arches is very good at it.

The gapThese systems do not acquire. Arches waits for a professional survey to be entered into it. This is a partner, not a rival: the right long-run posture is that Sākṣī feeds it.

Field data capture and mobile GIS

Esri Field Maps and Survey123 · Fulcrum · KoboToolbox · ODK

Capable, general-purpose tools a heritage office would otherwise buy.

The gapPriced per seat, which structurally forbids the public from contributing, and with no concept of a vantage: they record where you were, not whether you stood where the last photograph was taken from.

Visitor engagement and interpretation

Smartify · Bloomberg Connects · Google Arts & Culture · CyArk

Where the funding in this space actually is, and at real scale.

The gapInterpretation flows one way. The visitor is an audience, and the site gets no information about its own condition from having been visited.

Nothing in the survey spans all three links. The two capabilities that make spanning them possible, sensor-verified alignment to a stored vantage, and a custodian surface that acknowledges and closes a report, are precisely the two the adjacent categories each lack. The defensible position is not a feature; it is the loop.

Urgency

The demand side is no longer speculative

80%
of World Cultural Heritage sites under climate stress2025 assessment
98%
have faced at least one climate-related extreme since 2000
~73%
at high risk from water-related hazardsUNESCO, July 2025
19%
built substantially of threatened materials such as stone and wood

UNESCO has since launched a live monitoring platform integrating more than forty datasets with near-real-time alerts. That is the single most encouraging signal in this research, and it is worth stating plainly: the institutions have built the dashboard and are short of the ground truth to fill it. Satellite and climate-model data can say a site is exposed. Only someone standing in front of the wall can say whether the wall has moved.