# Methods How the reports in this repository are produced, and what they can and cannot support. ## The pipeline Public feeds are captured continuously, archived, run through detection models, and rendered as reports. Capture and inference are separate: nothing that detects runs inside the process that records, so a model change cannot cost recording time. | Feed | Model | What it decides | |---|---|---| | Orcasound hydrophones | OrcaHello | Whether a killer whale call is present in a window | | Orcasound hydrophones | Ecotype classifier | Which population a call is more consistent with | | Mount Rainier webcams | MegaDetector | Whether an animal, person or vehicle is in a frame | | Underwater video | YOLO-Fish with a YOLO-World second pass | Whether a fish or a larger animal is in a frame | ## AI agents in this pipeline The models in the table above are machine learning models, and most of this document is about what their outputs can support. Separately, and worth stating plainly: the code that runs them, the code that renders these reports, and the prose on these pages including this document are written with AI coding agents, under human direction and reviewed before anything is published. This is disclosed because it changes what a reader should check. A language model can write a fluent sentence about a number it did not compute, and the failure mode is a page that reads as authoritative while describing something the run never produced. Three things limit that, and a reader can verify all three: - **Site-level numbers are built from manifests, never from the pages.** The front page and `site-data.json` are generated from each report's `report.json`. Nothing scrapes a rendered page for a fact, so a claim that exists only in prose cannot become a site-level figure. - **The rights and redaction gates are code, not judgement.** A feed whose terms do not permit redistribution is refused at publish time and there is no override flag. - **The pipeline carries a test suite**, including regression guards for failures that have already happened here once. None of that makes the prose correct. It means the figures are checkable against each report's `report.json`, and where prose and manifest disagree, the manifest is what the run actually produced. ## What a detection is A model output. Nothing on these pages is a confirmed identification unless it says a person reviewed it, and reviewed cards are marked and kept visually distinct from model output. OrcaHello is a **call-presence detector**. It answers whether a killer whale call is in the window. It does not identify pod, individual, or call type, and no report here should be read as saying it does. ## False-positive floors A detection is only shown if its confidence clears the false-positive floor **measured at that specific site for that specific label**. Floors are measured, not assumed: windows where the label should not appear are run through the model, and the rate at which it fires anyway becomes the floor. A label with no measured floor gets no card at all, however confident the model was. That rule exists because this project once displayed a humpback label at 0.97 confidence when humpback had no measured floor whatsoever. Floors are **per site**. Pooling them across hydrophones lets a site with heavy traffic clear a threshold on its own volume and then license claims at a quiet site that never earned one. That happened here and was corrected. ## Review queues Some pages here are not results. They are **queues**: model output published before anyone has judged it, so the queue itself can be checked and so changes to the pipeline can be seen from outside. A queue page is marked `model candidate` and says so at the top. Nothing on one is an identification. The floors above do not license it, and cannot: the underwater video and fixed-camera models have no measured false-positive floor at all. That is precisely why these are published as queues rather than as detections. Where a floor exists, a detection clears it or is not shown. Where none has been measured, the honest choices are to publish nothing or to publish the output plainly labelled as unreviewed, and publishing it is what lets a reader disagree with a specific card. Those disagreements are the material a floor is eventually built from. Queue pages carry a verdict layer. Judging a card records it in your own browser and the export button hands back a file; nothing is transmitted. A queue is never counted alongside reviewed evidence: every finding on the front page carries its status, and the reviewed count is separate from the candidate count. ## Presence is plotted per day Counts alone cannot distinguish 446 detections spread over eighteen days from 30 in a single afternoon. The first is a claim about every day the site was listening; the second is one encounter. Reports plot presence per day for that reason. ## Coverage Every report states the hours of usable signal it covers. Feeds fail, nodes go offline, and hardware degrades in ways that leave a recording running while the signal is dead. A gap in detections may be a gap in animals or a gap in listening, and only the coverage figure separates them. "Usable" is stricter than "recording". One hydrophone in this network kept producing perfectly well-formed audio files for five days after its acoustic signal died, and every liveness check called it healthy. ## Negative controls A second model agreeing with the first is not corroboration until it has been tested on material where it should disagree. A general-purpose whale classifier was evaluated as a second opinion here and rejected. It ranked killer whale first on 77.5% of windows the primary detector had marked negative, against 79.3% on windows dense with genuine calls. That near-identical rate is a baseline prior of an out-of-distribution model, not agreement, and without the negative control it would have read as confirmation. Any second-opinion model used in a report here has been through the same test. ## Locations and a deliberate choice Site coordinates are published at the approximate precision the feed operators themselves publish, and never at a finer resolution. Reports cover **past windows only**. There is no live feed on this site and no real-time position for any animal. Publishing a detection at a location is a different act from the underlying feed being public, and for a species subject to disturbance the distinction matters. Approximate positions plus historical windows keep the reports scientifically useful without making them a tracking tool. ## Limits - Model outputs, not sightings, except where a card says a person confirmed it. - Call-presence detection does not identify pod, individual, or call type. - Ecotype classification is a comparison of consistency, not a determination. - Coverage is uneven. Feeds fail and are not always restored quickly. - Absence of detections is never evidence of absence of animals. - A queue page is unreviewed model output, not a result, and its counts are not findings. - Code and prose here are written with AI agents under human review. Figures are checkable against each report's `report.json`, which is the source of truth. ## Sources Feeds, their operators and their terms are listed in NOTICE, generated from the feed registry. Each report additionally records its own sources, models, model versions, run identifiers and effective licence in its `report.json`.