How OdyseeWatchdog Collects Data

The primary scanner queries Odysee's public Lighthouse search API (the same API the site's own search box uses) with a curated keyword list grouped by violation category: terrorism, hate speech / neo-nazi, harassment / doxxing / threats, CSAM / child safety, incitement to violence, drug trafficking, illegal weapons, financial fraud, cybercrime, human trafficking, animal cruelty, self-harm promotion, conspiracy-driven violence.

For every result returned, the scanner records: id, title, channel, url, tipsLBC (the highest effective stake — creator stake plus audience tips — seen on that claim; never lowered, see Peak vs. live stake), flags (categories the title matched), and lastSeen (date the scanner last confirmed the URL was live). The output is written atomically to /data/flagged.json.

ToS-compliant operation

• 8-second polite delay between requests (well above any documented Odysee API rate limit).
• Public metadata only. Titles, channel handles, effective stake. The scanner does not view, download, or mirror video content. It does not access private channels or bypass any access control.
• Atomic writes. JSON output is written to a temp file and renamed; no partial-write states reach the deployed dataset.
• Transparent diff. Every scanner run that changes the dataset is committed to the public git repository under a bot identity. The diff is inspectable; nothing is silently revised.

tipsLBC is the highest stake our scanner has seen on an item. It is never lowered: if the stake on a claim is later withdrawn, the recorded value stays at its peak. A sum of tipsLBC across items (the estimatedTipsLBC total) is therefore a sum of peaks, not a measure of what is staked today.

That matters because some channels move a single pot of LBC from video to video: they withdraw the stake and then re-stake the same coins on their newest upload. A peak sum counts that one pot once for every video it has sat on.

From October 11, 2026 the scanner re-checks live stake (the claim's current effective amount) on every run for every item recorded at 1,000 LBC or more, plus a rotating sample of smaller items; items under 100 LBC are not re-checked. A "not found" result is only recorded after two consecutive failed checks. Headline figures switch to live stake once a re-check covers at least 60% of recorded stake and is less than 72 hours old, and otherwise show the peak figure. Items that have not yet been re-checked are counted at their peak, so the live headline is an estimate and an upper bound. Where a page shows both, the peak is labelled "peak recorded placements". Per-item tipsLBC remains the peak; the live value, when checked, is written alongside it as liveLBC.

Effect on earlier reporting. The "inflow" figures in Collapse Watch №1 to №4 were recorded stake placements, not new money. Those articles carry corrections.

A keyword scanner pulls in false positives by design. We maintain a public blocklist of channels that match the keyword traps but produce clearly off-topic content (educational science channels, consumer-rights advocates, gaming creators, mainstream science explainers). Any item from a blocklisted channel is dropped from the dataset before publication.

We also maintain a title-pattern blocklist for benign substrings that incidentally match a violating keyword (e.g., "keyboard" drops hardware-review videos that match the "board" substring of certain hate-symbol queries). Both blocklists are public.

The remaining items are then classified into category flags by re-running the title against the same keyword groups (a stricter pass than the initial search query). Items whose titles don't match any category cleanly are tagged "Under Review" and surfaced separately on the Stats page — they are reachable via a violating search query, but the title alone isn't enough to assert a category. We hold them under review rather than over-claim. Every category count we publish is therefore a conservative floor, not a ceiling.

Six additional scanners run on demand and produce specialised datasets — plus a diff tool added in May 2026 for the deep- sweep round. All read-only, all public-metadata-only:

The primary scanner's default invocation runs at MAX_PAGES=10 against the standard 218-keyword set — polite, scheduled, conservative. Approximately quarterly we run a deep sweep with MAX_PAGES=50 against a keyword set expanded with topical 2026 additions (named perpetrators, sanctioned organisations, current war-propaganda terms, foreign-language extremism). The first deep sweep committed on 2026-06-02 at 12:16 UTC; it surfaced 5,506 new flagged items including +334 Terrorism, +224 Hate Speech, +46 Threats, +31 Incitement, +1 CSAM (referred to NCMEC per standing SOP) — without dropping a single item from the previous baseline.

The full deep-sweep walkthrough — what the deeper pages contained, what the new keywords yielded — is at What The Deeper Sweep Surfaced. Reproducible from the source repo at commit 514ca5a.

The headline dataset counts items the scanner has observed live within the last 180 days. An item the sweep has not re-surfaced in that window moves to an append-only archive (nothing is ever deleted): it may well still be live, but we no longer assert it. This makes the public numbers strictly conservative — every counted item was seen on Odysee, by us, within the window. Each item also carries a firstSeen date, so "flagged N days ago, still live" claims are anchored to the first observation, not the most recent one. The archive total is published alongside the active count as allTimeFlagged.

Counting flagged items is one thing; counting responses is another. The takedown register is a hand-maintained list of specific URLs we have referred to Odysee (with statute citations) and to the appropriate regulator (Coimisiún na Meán, ARCOM, BNetzA, IWF, NCMEC, FBI IC3, depending on jurisdiction). For each case we record: URL, channel, reason, date reported, who it was reported to, and current status.

The register is the empirical answer to the question "Does Odysee act when you specifically ask?" To date, status of every entry: still_live. Not most. Every. See /takedown-tests for the live list.

Anything in our datasets you can verify yourself:

1. Pull the relevant JSON from /data/<name>.json.
2. Pick any claim — say, @neohumaneve's 495,000 LBC self-stake on "One Lie to Rule Them All".
3. Open the matching Odysee URL listed in the JSON. Inspect the channel page yourself.
4. Compare the LBC figure shown on Odysee with the figure in our JSON. tipsLBC is the peak we recorded, so the live figure can be lower if the stake has since been withdrawn; where present, liveLBC is the re-checked live value.

For the methodology code itself, clone the repo and read scripts/scan_odysee.py and scripts/scan_stake_concentration.py. Both are well under 1,000 lines and have no third-party ML dependencies.

If you believe we've misclassified content, mis-quoted data, or that a channel we've named is a false positive we should add to our blocklist: press@odyseewatchdog.com or submit via the form. We will review and either correct, expand, or document why the call stands. Either way the response is logged in the public commit history.