Archive Genocide’s Gaza Node Map makes a large collection of footage easier to search as connected records. Its graph can lead from an event cluster to clips, captions, people, facilities, places, and the links between them. The interface also marks some links as leads, distinguishes publication dates from stated event dates, and publishes reasons for rejecting casualty figures.
That is useful work. It is also exactly where a reader can lose the difference between a record and a finding. A label in a database does not finish the reporting. It tells you what to check next.
For a guided search of individual clip records, use the Archive Genocide API reader on Deceit Media. It sends searches directly to Archive Genocide and shows text and source links without embedding the footage.
Two ways into the archive: clips and the event graph
Archive Genocide now publishes a read-only REST API for individual clip records, alongside the Node Map and its downloadable event-graph files. The API is a useful way to search captions, dates, source collections, media types, and location availability. The Node Map remains the better place to explore how the archive connects footage into event clusters and links those clusters to people, facilities, places, and other events.
The API needs no key or account. Its documented endpoints are /v1/clips, /v1/clips/{id}, /v1/archives, /v1/withdrawn, and /v1/meta. The OpenAPI specification is available for software that can import an API schema. The graph guide separately lists events.geojson.gz for QGIS, ArcGIS, or GeoPandas, and graph.graphml.gz for Gephi, igraph, NetworkX, or Cytoscape, alongside CSV tables. Use the API for filtered record lookup; use the graph files for bulk analysis and relationships.
For example, this request finds clips whose English or original-language caption contains all the words “al shifa,” limited to videos with map locations:
https://api.archivegenocide.com/v1/clips?q=al%20shifa&type=video&has_geo=1&per_page=25The API returns a page of records. per_page can be set from 1 to 100; use page to continue. Each search is capped at 10,000 results, so split broad searches into smaller date ranges. The service allows 120 requests per minute and returns Retry-After when a caller needs to slow down. For bulk footage, use the signed torrents on the archive’s Downloads page rather than fetching media one record at a time.
The API reports date as the time the source posted the clip, not necessarily when the event happened. confidence describes the archive’s review status, not whether an event is true. A geo value can be an area anchor or a named site; read precision_level and exact before interpreting it. The caption field may be translated, while caption_original preserves the source wording. Each API record links back to a page on Archive Genocide, and may include the original post URL.
The API’s /v1/withdrawn endpoint lists IDs removed from the archive. If you keep a local copy of records, check that list and remove withdrawn records as the API documentation requests. The API is a finding aid over source material, not a verdict system.
We expect to use Archive Genocide’s API across our research: to find clips by caption, narrow results by posting date, source collection, media type, or location, and follow each record back to its archive page and original post when one is available. When we retain record details in our notes, we will check the withdrawn list before relying on them again. Our reader sends searches directly to Archive Genocide and gives each result a path back to its record. We recommend the API to journalists and researchers who need a searchable route into the archive. It makes source material easier to find. It does not make a claim true. Credit Archive Genocide and keep the record link beside the claim.
The live search accepts examples such as cat:press tier:A, place:rafah, killed:>10, clips:>3, and date:2024-01..2024-03. The map can show events, people, actors, places, facilities, and dates. It can filter corroboration tiers and categories, include or hide possible links, and show mapped or inferred locations. That is a useful interface for discovery. The downloadable files are better when you need to examine many records, join tables, or keep a specific version for your notes.
The project describes the current graph as a pilot on partial extraction. That phrase belongs beside every chart made from it. A graph can be internally consistent and still cover only the portion of the underlying record that has been extracted.
The related datasets are not interchangeable
Archive Genocide publishes more than the clip API and Node Map. The Recorded statements of intent page is a research record with source links, a downloadable statements.csv, a separate candidate file from zionism.observer, and the score.py code used for its provisional index. That index records how the page coded public statements. It is not a legal finding about genocide or intent. The page’s source register and the original court, government, or public statement remain the route for checking a claim.
The Who Funded Genocide database is a separate public site on the archivegenocide.com subdomain. Its methodology says the political-finance records come from FEC filings, while the wider database also includes lobby disclosures and charity records. Do not collapse those categories into one kind of money. A contribution records a filed payment to a committee. It does not establish a donor’s beliefs, knowledge, or intent. Independent expenditures are spending on a race, not money received by the candidate. The FEC filing remains the authority, and the database’s snapshot date belongs beside any number quoted from it.
The topic index and source index are discovery pages. They can help a reader find a collection or route into the archive, but their visible counts are page snapshots and do not independently verify every record. A reader using any of these resources should name the resource, build or snapshot date, source checked, and remaining uncertainty.
A node is a cluster of clips
The dataset calls an event a cluster of footage. It does not claim that each cluster equals one independently verified incident. The published guide says nine out of ten events contain a single clip. Clips are grouped using signals the project says can be recomputed: matching video fingerprints, the same original post URL, or a shared date and resolved place. The Node Map’s methodology explains that a fingerprint match can tolerate re-encoding, while a shared date and place only groups clips that claim the same day and resolved location.
This design choice errs against merging distinct events into one. It also means separate clusters may still refer to the same incident, especially when captions are sparse or dates and place names do not line up. A count of graph nodes is therefore not a count of attacks, massacres, or any other kind of event in the world.
The graph’s links carry their own grades. “Established” means a deterministic signal connected the records, such as a matched fingerprint, shared post URL, resolved gazetteer ID, or a name in the text. “Possible” means the link is a lead, such as a shared surname or a weaker fingerprint match. The map draws possible links with dashed lines and labels them unconfirmed. Those lines are invitations to investigate, not evidence that the people or events are connected.
That distinction matters when a graph is screenshotted. A dashed line can disappear in a crop. A possible relationship can become a sentence with a subject and a verb. Keep the grade and the reason attached when you export or quote a relationship.
Follow a record back to its source
The event graph is most useful when you move through it in layers.
Start with an event ID and the fields that define why it appeared in your search. Then join that event to event_clips.csv.gz using event_id. Each clip row can include the archive entry ID, media path, media type, duration, verbatim caption, original post URL when recorded, source archive, publication timestamp, and any date stated in the caption. The path points to the archived file. The post URL, where present, points closer to the original publication.
Next, inspect the relationship tables. event_entities.csv.gz records whether an event is linked to a place, a struck site, a person named in the text, or a unit described as carrying out or stating something. event_relations.csv.gz connects event records and includes the reason for the link. The people and actors tables are separate. That separation is important: a person mentioned in a caption is not automatically the same thing as an actor the caption says carried out an act. The facilities table has a resolved field. A zero means the facility name has not been reconciled to a canonical place entry, so spelling differences can leave one site looking like several.
The archive’s own guide says to open the clips before citing an event. The Berkeley Protocol makes a related point for digital open-source investigations: assess the item and its context, try to identify the original source, and preserve a clear record of collection and handling. A table can speed up that work. It cannot do it in your place.
When a clip matters to a published claim, follow the chain all the way back: What does the original post say? Is the footage itself available? Does the visible material support the caption? Is the publication timestamp being mistaken for the event date? Can an independent source establish the location or timing? Which part remains unknown?
A map coordinate is not a strike coordinate
The Node Map distinguishes locations stated in captions from locations inferred by the extraction process. Its data guide says coordinates are place anchors: the gazetteer position for a named town, camp, or building. They are not surveyed strike locations. The map says an inferred place can be shown as a hollow marker, and its documentation says events whose locations could only be inferred are excluded from the downloadable event GeoJSON.
In the earlier event-table snapshot analyzed for this article, 52,786 rows had mapped=1 and 62,369 had mapped=0. Those counts describe that build’s coordinate coverage, not the share of real incidents that occurred at a known place, and should not be read as current totals.
That is a valuable warning to preserve. A point on a map carries a precision the underlying record may not have. The location could identify the town named in a caption without identifying the building, street, or precise point where an event occurred. A visual dot can look more exact than the text that produced it.
For analysis, keep the map’s location type with the coordinate. Do not turn a town anchor into an exact attack site. If the question depends on the exact location, the anchor is not enough.
Dates and numbers need their own labels
The event table includes date_source. The guide distinguishes a date stated in text from post_proxy, a publication timestamp used as a proxy, and retrospective_unknown, a reference to an earlier event without a recoverable date. A social post dated October 7 may describe something filmed earlier, summarize several events, or mark an anniversary. Sorting by date without checking that field can turn a posting timeline into an incident timeline.
Casualty fields need similar care. killed_min and killed_max hold figures stated about an event. The separate cumulative_min and cumulative_max fields hold running totals. The archive explicitly warns that event figures should not be added across rows: separately published clips can place the same incident in more than one event cluster.
The map also publishes figures it rejected and the reasons for rejecting them. That is one of its strongest accountability features. A dataset that exposes only accepted entries asks the reader to trust its editorial decisions. A rejected-figures ledger lets a reader inspect some of the work the archive chose not to include. It still does not make the remaining figures independently verified.
The CSV has a figures_disputed flag for event rows where sources in the archive disagree on casualty figures. It was set on 714 rows in the earlier build 1f0fb73f0b62. That flag points to a place where the source record conflicts with itself. It does not tell us which figure to select, and that count is not a current total.
A letter only means what this version says it means
The live map and downloadable graph guide do not use the same tier vocabulary. The downloadable guide defines four labels, A through D. It describes A as independent footage in separate archives and D as a single clip. The live map’s methodology lists A through E. It describes C as a single source with an original-post link, D as archive-only with a stated date, and E as weak because the date is the post timestamp.
The earlier event CSV snapshot analyzed for the original draft, build 1f0fb73f0b62, contained A through D, with no E rows. Its verification_tier field labeled 108,359 of 115,155 rows D. The guide said nine in ten clusters contain one clip; counting that file gave 108,333 rows with n_clips equal to one. The two counts are close, but the labels are not identical measures. A row is not automatically Tier D just because it contains one clip. These counts describe that earlier build only.
This distinction matters if a screenshot of the live interface is compared with a CSV analysis. A letter can survive while its definition changes. Name the source file or interface version when quoting a tier, and do not translate labels between them without checking the build.
The tiers can help prioritize review. They do not certify the whole event, determine responsibility, or settle a legal question.
The distinction between sourcing and authority laundering is useful here. Authority laundering occurs when a claim borrows institutional weight without showing the underlying authority or evidence. A named dataset with its method, source links, and caveats visible is not automatically authority laundering. But if a writer turns “the archive labels this Tier A” into “the event is proven,” the borrowed label is doing more work than the cited records.
Patterns are clues, not verdicts. A strong corroboration tier can tell you that the archive holds multiple pieces of footage it treats as independently filmed and separately sourced. The next questions still belong to the reader: What do the clips show? Are the sources genuinely independent? Do their captions agree? Is there outside corroboration? What would change the conclusion?
Version the data before quoting it
Counts change as records are added, corrected, or clustered differently. On November 2, 2026, the live Node Map displayed 118,293 events and 157,227 linked records. Its current data guide identifies build 32f98e7ff312 and describes 118,293 events assembled from 135,775 clips. The API’s /v1/meta response, built at 2026-11-02T11:23:17Z, reported 186,628 clip records and 33,993 with a map location. These figures describe different data products and should not be added together or treated as interchangeable totals. Name the interface or endpoint, build time, and retrieval date when quoting them.
Counting the date_source field in that earlier build’s 115,155 CSV rows gave 99,742 labeled post_proxy, 13,917 labeled stated, and 1,496 labeled retrospective_unknown. That made 86.6 percent of the rows post proxies in that snapshot. It is a count of records and their date labels, not a count of incidents or a measure of how much footage is true. It shows why a date filter needs the date field beside it. These figures are historical counts, not current API totals.
The data guide says the graph files are covered by a SHA-256 manifest and detached GPG signature. I checked the manifest signature with the public key linked from the archive’s verification page, confirmed its fingerprint against the fingerprint published there, and verified that the downloaded event CSV matches the signed manifest. GPG reported a good signature. This supports the provenance of the manifest and the integrity of the CSV against it. It does not establish that any caption, date, identity, location, or interpretation is accurate.
If a finding depends on a dataset snapshot, keep the build ID, retrieval date, manifest, and relevant rows with your research notes. A URL to a changing page is not a snapshot of the evidence you used.
The people in the records are not just fields
The dataset includes named people, units, facilities, and identified victims. Names in captions can be incomplete, transliterated several ways, or shared by different people. The people table’s in_victims_list flag records a match to the ministry register, while victim_ambiguous records cases where several register entries match. The guide explicitly calls an ambiguous match a lead, not a confirmation. A name match should not be published as an identity without independent checking. The map’s own design notes say it keeps people and perpetrators in separate classes; researchers should preserve that distinction in prose and in any derivative dataset.
The archive warns that its footage includes extreme violence and children. WITNESS’s video ethics guidance asks people using eyewitness footage to consider consent, intended audience, safety, dignity, privacy, credit, and context. That applies even when the footage is already public or preserved in an archive. Do not use a child’s face or a dead person’s image to make a point the record can support without it. A warning label does not settle whether publication is necessary.
The project says its graph data is CC BY 4.0 and asks users to credit Archive Genocide. Its download page separately says the code is MIT and the footage is available to download, mirror, and share. Those statements do not establish item-by-item rights for every media file. Reuse the licensed data with attribution; review rights and ethical necessity separately before republishing footage.
When this method is wrong
Treating every label as provisional can become an excuse to dismiss evidence that is inconvenient. A single clip can still be important evidence. A survivor’s testimony does not become worthless because it has not been corroborated by a second archive. The discipline here is not to demand impossible certainty before believing anything. It is to say precisely what each record supports, what it does not, and what further evidence could change the assessment.
The Node Map does valuable work because it leaves paths back to the source and marks some of the uncertainty in its own structure. Use it to find the clip, inspect the caption, test the relationship, and keep the limits beside the result. Before the next sentence goes into print, someone should be able to open the clip and read the caption beside it.
