N1: studying ‘Where’ in physical inference

Oceanir uses evidence to work out what’s true about the physical world. We call this physical inference. We’re starting with a question that gives an observation its spatial context: Where?

For a photo, that means working out where it may have been taken and how precisely the available evidence supports locating it. Some scenes suggest a part of the world while giving very little away about the street, building, or spot where they were captured.

N1: a proposed model for the experiment

We propose N1 as a research model for this experimental work. Its intended role within an Oceanir system is to connect a visual observation to candidate locations that can be examined against further evidence.

The research question is how to move from a plausible area toward a more precise location, while preserving alternatives and recognising when the evidence remains inconclusive. The experiments in this article study that problem and inform the N1 proposal.

What ‘Where’ contributes to physical inference

A location gives other observations a place to be checked. It can help someone decide which area to examine next, whether two images might concern the same place, or where a claim about an event needs to be verified. A candidate location is a hypothesis to investigate; establishing what happened requires additional evidence.

N1’s proposed contribution is this spatial part of the system: possible places for an observation, with a useful distinction between a broad area and a close location. Our design goal is for the wider system to carry that uncertainty into the next stage of review.

Oceanir’s broader direction is to understand places and events from photos, video, and eventually live sensor feeds. The measurements here concern photo-location estimates. They are one research step toward that direction.

Figure 01 / Oceanir research

Where N1 would fit

Starting with difficult photos to locate

We gathered 68 private photos from some of the most difficult cases we had tried to locate. We wanted to understand what happens when the obvious clues are missing or the first answer is still too broad to be useful.

On October 31, 2026, we tested two versions of an experimental system on that collection. This is an early research update for upcoming Oceanir releases. The collection was deliberately selected for difficulty, rather than sampled at random from all uploads.

Why a photo can be hard to locate

A row of buildings, a stretch of road, or a coastline may resemble places thousands of kilometres apart. An image can suggest a climate or architectural style without containing a clue that identifies one particular place.

Researchers call this localizability: how precisely the visual information supports locating a photo. Work such as Around the World in 80 Timesteps explicitly models this ambiguity, allowing several plausible areas instead of forcing every photo into one confident point.

The test itself matters, too. OpenStreetView-5M separates training and test locations to examine whether a model has learned geographical features beyond memorising familiar places. A strong score on one collection does not tell us how a system will handle every difficult upload.

These are reasons to measure errors at several distances. They are also reasons to keep a broad estimate visibly broad.

What we set out to learn

To investigate the ‘Where’ problem behind N1, we assembled an experimental system and tested a practical question: can an updated version bring the estimated location closer to the recorded location on the same difficult photos?

We measured both versions at several distance thresholds. This lets us see whether a change helps at close range, only at broader distances, or comes with a regression elsewhere. The results below describe those outcomes.

Results on the 68 difficult photos

The updated experiment placed 25% of these photos within 25 km of their recorded coordinates and 48.5% within 200 km. Its median error was 247.5 km.

Figure 02 / Oceanir research

The gap between a broad area and a close location

Figure 03 / Oceanir research

Two summary measures, two separate scales

In both versions, 51.5% of estimates were still more than 200 km from the recorded location. That is the gap this collection exposes. A high result at a large distance threshold can coexist with a much lower result when the task requires a close location.

The within-1-km figure also means within a kilometre of the reference, rather than confirmation of the exact camera position. Even that tighter threshold needs to be interpreted carefully.

Both versions were evaluated on the same 68-photo collection. Lower median error is better; higher GeoScore is better. Distances are relative to the recorded coordinates.

What changed on the difficult photos

The updated experiment had a lower median error and a higher GeoScore. But its within-25-km result fell from 27.9% to 25.0%, while results within 200 km and 750 km were unchanged. The increase within 1 km corresponds to roughly one additional photo out of 68.

This study does not establish an overall improvement. It tells us where an apparent gain needs closer examination, where performance went backwards, and how much work remains before a broad estimate becomes a close location.

The same approach, tested on 123,345 images

We also tested the same overall approach in a separate, larger experiment with 123,345 images. On that collection, the updated experiment improved the regional results: the share within 25 km rose from 36.30% to 41.27%, and the share within 200 km rose from 65.59% to 70.26%.

Median error fell from 73.1 km to 46.8 km. That means half the estimates were within 46.8 km of their reference locations, compared with 73.1 km before.

Figure 04 / Oceanir research

Regional gains, with a close-range tradeoff

Swipe the table to compare both versions.

The gain within 25 km was about five percentage points. Close-range performance moved the other way: the share within 1 km fell from 6.59% to 6.19%. This was a diagnostic experiment with a regional improvement and a precision tradeoff.

These figures belong to the larger collection. They cannot be transferred to the 68 difficult photos, where the results were smaller and mixed. Other evaluations also showed regressions, so this is evidence to guide further work rather than proof of a release-wide improvement.

Figure 05 / Oceanir research

Lower median error in both collections

What we still need to verify

For the 68-photo collection, the reference coordinates came from EXIF and were not independently verified. It was selected from difficult cases, without a documented difficulty ranking. It cannot establish average performance across Oceanir uploads or rank modern location AI systems against one another.

Only aggregate results were retained from the 68-photo run. Its per-photo predictions are unavailable, so we cannot audit which individual photos improved or regressed. The chart shows recorded outcomes; it cannot establish the cause of each error.

Both experiments inform the proposed N1 research model. N1 remains at the proposal and experimental stage. These measurements do not establish the performance of N1 as a released product, the current Orca app, or a released Manta version.

How this informs upcoming releases

For the next evaluation, we need independently checked reference locations, retained per-photo outputs, and a documented selection of test photos. That would let us inspect an apparent gain image by image and understand which kinds of difficult photos still need work.

For N1, the aim is practical: make the ‘Where’ part of physical inference more useful. That means helping someone locate a photo more precisely when the evidence supports it, showing plausible alternatives, and making unresolved uncertainty understandable. We’ll share release-specific results as that work is validated. This study does not set a launch date.

Tell us what you need to locate