What the monitor flags
The example follows pigs leaving a lorry, crossing a ramp and entering a corridor. The simulator assigns flags to selected event types, such as crowding at a door. Other clips remain unflagged. We ask whether some observations might need a description that the existing categories do not provide.
In this simulated example, – clips are flagged per 100 pig-minutes. A pig-minute is one pig observed for one minute. Rates account for differences in observation time; raw counts alone do not.
Basis: in the public Sentient Futures recording, operators describe a move from random footage review to system-selected footage. The transcript was checked against the video. Reporting rates per unit of exposure addresses a limitation noted for Čobanović et al. (2021): observation time differed between groups.
What goes unreviewed
Each dot represents a simulated clip of one pig, positioned by its longest stop and number of turns back per minute. Amber marks clips selected by the simulated monitor. Other clips are unflagged; the review layer can sample them later.
The review layer adds grey-to-teal shading for flag distance: the distance from each unflagged clip to the nearest simulator-flagged clip, using the monitor’s two features and a scale from 0 to 1. This distance is inspired by open-set recognition, which allows for inputs outside known classes. It is not an estimate of open-space risk, calibrated uncertainty, confidence in a welfare judgement or a validated measure. The simulator assigns the flags; no trained classifier is used. The simulated monitor leaves both grey and teal clips unflagged. Hover over a dot to see its feature values.
A model can assign high confidence to an incorrect prediction. Distance from training examples and predictive uncertainty are different quantities, so selecting only low-confidence predictions can miss confident errors.
Basis: Boult et al. 2019 (uncertain is not unknown); Lakkaraju et al. 2017 (confident errors cluster where training data was thin). In this page there is no trained classifier: flags are simulator-assigned, and the distance is to simulated flagged examples. Gómez et al. 2021 found 93% of 111 validations of pig monitoring technologies were internal, tested on the population used to build them.
Candidate groups and their descriptions
The review layer uses density-based clustering to group simulated clips with similar feature values. The groups depend on the selected features and settings. In this prototype, a lookup supplies the welfare descriptions; a real review would need to assess whether they fit. A description is a proposal, not a finding.
Features available to the review layer (2 of 8 selected; changing them reruns the clustering)
Start with the monitor’s two features, then add posture and look at group E. The clips stay the same; the available features change. In this constructed example, the selected features do not clearly distinguish the planted group. Adding posture makes a separate cluster visible. This illustrates a possible representation limit: discovery depends on having informative features, an assumption discussed by Lakkaraju et al. (2017). It does not show that every method would fail, that the group has welfare meaning or that a new category exists. The random slot can draw an E clip even when clustering does not separate the group. Encountering a clip and recognising a distinct group are different things.
Rates are per 100 simulated pig-minutes. The grouping step is inspired by Lakkaraju et al. (2017). This page uses DBSCAN, a density-based clustering method, on the selected features; the outlines show its output. The published method and this implementation differ. Here, matching a cluster to a published definition is an authored lookup. In a real review it would require judgement. All clips are simulated.
How sure are we, and about what?
This display is under review. Its numerical policy, interval labels and authored fit judgements are unresolved. It does not validate a welfare interpretation.
Three questions, kept apart: how often a pattern occurs, how unfamiliar it is to the monitor, and whether our own estimate can be trusted.
| Group | Unfamiliar to the monitor | Description fits | Can we trust the interval? |
|---|
Notice where group E lands: once small counts are pulled towards the background, it looks entirely ordinary. Ranked by frequency alone, it would never reach a reviewer. And on the monitor’s own features it is not even unfamiliar: its flag distance is as low as a typical clip’s. Neither frequency nor unfamiliarity would single it out.
Group E is the case open-world methods exist for, and also the case where they reach their limit. Its interval is not the main problem: any interval assumes we know what we are counting, and for E we do not. The honest statement is that we cannot yet say how uncertain we are, because the features in which we could say it are not in the monitor. The next step is a different observation, not more counting. In this page that is a constructed illustration, not a proof.
Borrowed from drug-safety signal detection: compare each pattern with what exposure would lead you to expect, and pull small counts towards the background before ranking them (the empirical Bayes idea behind EBGM). These methods suit known kinds of events at unexpected rates and are weak for genuinely new ones, which is where group E sits. Unfamiliarity follows open-set recognition (Boult et al. 2019); description fit follows Lakkaraju et al. 2017. Background rate and intervals are invented.
A review budget of twelve clips
Review time is limited. This example allocates eleven clips across eligible clusters and any unplaced unflagged clips. Each eligible group receives one clip first, then remaining slots follow distance-based ordering. Within each arm, or selection group, clips with the lowest local density are drawn first. A separate seeded draw selects the twelfth clip from all unflagged clips.
| Arm | Clips | Why this share, and which clips |
|---|
A future review workflow could let reviewers change the allocation and record their reasons. This prototype does not yet provide that control.
Basis: Lakkaraju et al. 2017 allocate a limited labelling budget across groups with an explore-exploit bandit, returning to productive groups while still sampling the rest. The page displays a feedback-based allocation inspired by an Upper Confidence Bound rule. It uses the latest visitor judgement for each arm in the current partition, with an exploration bonus. It does not accumulate reward means or pull counts across repeated rounds. Within each arm, the page selects clips with the lowest local density first. Unplaced unflagged clips form a separate arm, providing a route to review for clips that do not form a cluster. This is inspired by rare-category discovery (Pelleg and Moore, 2004). Reward is high when a reviewer judges a group to be something the monitor misses, scaled by their credence. The twelve-clip budget is invented.
Review a cluster, then compare an authored response
Review the simulated clips selected from each cluster. After you save, the page shows an invented second response. It is authored content, not a human review. Each judgement is saved as a dated event with its clips and partition. Saving again adds a revision and keeps the earlier event.
Why two reviewers: when observers scored fear-related behaviour at unloading, agreement between them was low to moderate (rank correlations about 0.25 to 0.43 across 100 groups; Dalmau et al. 2010). A single expert label is not ground truth for these behaviours.
What could change, and who could act
Each answer has a suggested next step. None is automatic, and most require someone with access to agree to act.
Two limits apply to every option.
Changing a category can change recorded counts without changing the animals’ experience. A new flag can make a problem appear larger or smaller even when nothing has changed for the pigs. The copy cites a related effect of scoring rules in Veissier et al. (2011).
Fewer adverse events do not, by themselves, establish positive welfare. Assessing positive welfare requires observations suited to it, beyond the problems this monitor is designed to flag (Rault et al. 2025).
From observed patterns to wider questions
The four definitions in this diagram remain under review. They do not describe an automatic progression or a validated route to establishing moral status.
Group E raises a question about the descriptions available in this example. It is not a welfare finding. The following diagram places that question alongside broader questions about observation, welfare and moral consideration.
The wider research question is how observations that challenge existing descriptions can reach people able to examine them. This prototype records judgements and proposals. It does not automatically create a category, retrain a monitor or establish a being’s moral status.
For pigs in this example, we ask whether a pattern matters for welfare. Other contexts may also raise questions about sentience and moral consideration. Those questions require their own evidence and methods; this prototype does not answer them.
A question for future work is how monitoring could preserve routes for examining observations and subjects outside its existing categories.
This prototype explores a bounded case using simulated pig observations. It does not establish a general method for identifying sentience or deciding moral status. Applying the wider idea elsewhere would require separate investigation.