DisputeDB — UDRP legal research
Three tools over one corpus of UDRP (Uniform Domain-Name Dispute-Resolution Policy) material:
| Tool | Answers |
|---|---|
datapulse_disputedb_search | Reasoning: bad faith, legitimate interests, panel consensus, policy. Semantic search over decision text and the reference documents; returns passages, never statistics |
datapulse_disputedb_lookup | Facts: one case, the disputes over one domain, or one party’s or panelist’s record with outcome counts and win rate |
datapulse_disputedb_stats | Counts: outcomes and the complainant win rate across the whole corpus, per provider and per year |
“What percentage of UDRP complaints succeed” is a stats question; “how often does Microsoft win” is a lookup question. Semantic search cannot reliably answer “has anyone disputed example.com”, because that depends on the string appearing in some chunk; lookup with domain answers it exactly.
Citation integrity
Every search result and every decision carries a pre-formatted citation naming its source exactly, e.g. WIPO Case No. D2024-0001 (example-brand.com) — transfer. UDRP case numbers that do not appear in these results are unverified: LLM training data contains incorrect and fabricated case numbers that either do not exist or involve different parties. A number this server does not hold answers found: false (for example D2099-9999). A search with no relevant result supports no citation.
The corpus
| Source type | Contents | Chunks | Items |
|---|---|---|---|
decision | WIPO and Forum (NAF) panel decisions, 1999–2026 | ~946,800 | ~100,900 with full text |
overview_section | WIPO Overview 3.0, panel consensus views | 101 | 64 sections (1.1–1.15, 2.1–2.15, 3.1–3.12, 4.1–4.22) |
guide | WIPO Guide to the UDRP, Q&A explainers | 52 | 52 items (A01–I01) |
policy | ICANN UDRP Policy | 35 | 35 paragraphs |
policy | ICANN UDRP Rules | 107 | 106 paragraphs (one split across two chunks) |
- Decisions: 118,249 (74,023 WIPO, 44,226 Forum), dated 1999-10-10 to 2026-10-16, measured 2026-10-19. About 100,900 have full text and are searchable; the other ~17,300 are metadata-only records of cases that were terminated, withdrawn or never published in full. 17,183 have no decision date.
- Chunks: ~947,000, all embedded with BGE-M3 (1024 dimensions) and searchable (measured 2026-10-14); the 295 curated chunks are the overview, guide and policy rows above. Decisions average about 9 chunks.
- Refresh: decisions arrive through periodic discovery and ingestion runs and join the index at the next embedding run. The policy, rules, overview and guide are static documents, re-ingested when the source changes.
- Coverage is WIPO and Forum only. CAC, ADNDRC, CIIDRC and the other UDRP providers are not in this corpus.
Documents are split into chunks of about 2,000 characters with a 200-character overlap, at paragraph breaks and then sentence boundaries. chunk_index 0 is the start of a document; most guide and policy items are one chunk.
Disputed domains are not the complainant’s domains
Search results carry the disputed domains as domain_names (strings); lookup decisions carry them as domains (objects, see “Disputed domains”). Either way they are the domains a case was brought against. A brand owner’s own domain is absent because nobody disputed it: searching it for microsoft.com finds nothing, correctly, even though decisions routinely cite the complainant’s own domain as evidence of its rights. A company’s enforcement history is keyed on the complainant, which only datapulse_disputedb_lookup with complainant can answer; search results carry no complainant field.
datapulse_disputedb_search
| Parameter | Type | Required | Description |
|---|---|---|---|
query | string | Yes | 1–1000 characters; whitespace-only is refused |
limit | integer | No | 1–12; absent or 0 is the default of 10; above 12 is capped to 12. A result is 1–5 KB (a passage of up to 2,000 characters plus its fields), so the default page is under about 46 KB |
min_score | number | No | 0–1; default 0.55; 0 disables the filter; outside 0–1 is refused |
source_type | array of strings | No | Any of decision, overview_section, guide, policy; absent, every kind is searched |
include_all_cited_cases | boolean | No | Every cited_cases entry instead of the first 25 (default false) |
Response
{
"query": "bad faith registration",
"returned": 10,
"min_score": 0.55,
"results": [
{
"id": "uuid",
"source_type": "decision",
"source_id": "uuid",
"chunk_index": 3,
"content": "The Panel finds that the Respondent registered the domain...",
"metadata": {"source": "wipo"},
"score": 0.7142,
"case_number": "D2024-0001",
"case_uid": "wipo:D2024-0001",
"domain_names": ["example-brand.com"],
"outcome": "transfer",
"citation": "WIPO Case No. D2024-0001 (example-brand.com) — transfer"
}
]
}
returned is how many results came back. A similarity search has no total (every chunk matches to some degree, and the limit is the answer). The response echoes min_score and, when one was passed, source_type; an empty result carries a note explaining it in terms of the threshold, so “nothing scored above 0.97” is not read as “the corpus has nothing on this subject”.
Decision results carry case_number, case_uid, domain_names and outcome at the top level, read from the decision record at query time; metadata then holds only what the top level does not (source). domain_names lists the first 25 disputed domains: a longer list (one Forum case disputes 295) carries domain_names_total and domain_names_truncated: true. The citation names the first domain and adds “et al.” when there are more.
Metadata by source type
decision (the top-level fields, and source in metadata)
| Field | Example | Description |
|---|---|---|
case_number | D2024-0001 / 1733616 | Provider’s own number: WIPO D<year>-<seq>, Forum a bare integer with no provider marker |
case_uid | wipo:D2024-0001 | Provider-qualified, unambiguous across providers. Split on the first colon: WIPO numbers contain hyphens |
source | wipo / forum | Provider (in metadata) |
outcome | transfer | transfer, denied, cancelled, terminated, withdrawn, split decision; case active is a pending case |
domain_names | ["example.com"] | Disputed domains only, the first 25; domain_names_total and domain_names_truncated mark a cut list |
overview_section: section_number (2.1), title, source (wipo_overview), element (first, second, third, procedural) and cited_cases. cited_cases is a hand-curated citation graph: the leading cases on a point of law come from the Overview section for it, without semantic search. The list is cut to its first 25 entries by default, and a cut list carries cited_cases_total and cited_cases_truncated: true; include_all_cited_cases: true returns every entry. The median section cites 21 cases; 3.1 cites 113 and 3.2 cites 89 (1,521 citations across all 64 sections), so the cap applies to 20 sections (counted 2026-10-06).
guide: question_id (A04), question, section (A–I), section_heading, source (wipo_guide).
policy: document (udrp_policy or udrp_rules), paragraph (4(a)(i)), title, source (icann). The key paragraphs: 4(a) the three elements, 4(a)(i) identical or confusingly similar, 4(a)(ii) no rights or legitimate interests, 4(a)(iii) bad faith registration and use, 4(b) bad-faith factors, 4(c) legitimate interests, 4(k) court proceedings.
Scores
BGE-M3 cosine similarity between the query and a chunk. BGE-M3 scores run lower than some models: on this corpus the ceiling is about 0.75–0.80, and only paraphrase-style queries reach ~0.81.
| Score | Meaning |
|---|---|
| 0.72+ | Strong match, directly relevant |
| 0.65–0.72 | Relevant |
| 0.55–0.65 | Weak, possibly tangential |
| below 0.55 | Poor (filtered by default) |
Finding the rule itself
The 295 curated chunks compete with ~947,000 decision chunks in one index, so a question phrased differently from a curated heading is usually answered by decisions quoting the rule. Restricting source_type makes a policy lookup deterministic:
datapulse_disputedb_search(query="UDRP rules response filing deadline",
source_type=["policy","overview_section","guide"], limit=5)
datapulse_disputedb_lookup
Exactly one selector:
case_number: "D2000-0003" one decision; retired identifiers still resolve
domain: "example.com" disputes brought AGAINST that domain
complainant: "Microsoft" what that party has filed
respondent: "Fundacion …" what was brought against that party
panelist: "Neil Anthony Brown" cases that panelist DECIDED
A null or whitespace-only selector counts as absent. Each party mode reads its own records, so a name on the other side of a dispute does not contaminate the answer.
Matching
domainis an exact match. A bare label is refused (equifaxis not a domain name). The name is canonicalized asdatapulse_help(topic="conventions")describes; a leadingwww.is sent as part of the name (the response’squeryreadsdomain:www.tesla.com) and the database matches it away, sowww.tesla.comfinds the same decisions astesla.com.complainant,respondentandpanelistmatch by words. Every whitespace-separated word must appear in the recorded name, in any order, as a case-insensitive substring:panelist: "Debrett Lyons"finds “Debrett G. Lyons” and “Mr. Debrett Gordon Lyons”. A partial word matches (LyonfindsLyons); there is no stemming, fuzzy or phonetic matching (Newmannever findsNeuman). On zero results the surname alone or a shorter distinctive word is the next try.- Party queries need a word of at least 3 characters: every word is matched on its own, so the longest one counts (
G. Lyonsis accepted,G.anda bare not);complainant: "HP"is refused rather than answered with an arbitrary sample.HP Incworks. - Maxima:
case_number64 characters,domain253, party names 200.
Lists and paging
domain, complainant, respondent and panelist return a list sorted by decision_date descending, then case_number descending, so a capped list is the most recent matches.
limit: default 20, max 35 (capped); 0 is the default. A row is about 1.4 KB, up to 3.5 KB.- A list row’s
complainants,respondentsanddomainsstop at 10 entries each; when a list is cut,<field>_totalgives its full length and<field>_truncatedistrue. Thecase_numberlookup returns every entry. offset(party and panelist lookups): default 0, max 100,000 (larger values are capped to it); a negative offset is refused.offset: 0, limit: 20is rows 1–20 andoffset: 20is rows 21–40.- The response carries the true
totalbesidereturned, theoffset, and, when rows exist beyond this page,truncated: true,next_offsetand anote. A page withtruncated: trueis a partial history, not the full one; a page past the end says so.case_numberanddomainlookups ignoreoffsetwith a note: no domain has more than a handful of disputes.
Outcome counts, win rate and matched names
A party or panelist lookup also describes its whole match set:
"total": 458,
"counts": {"transfer": 428, "terminated": 26, "denied": 4},
"complainant_win_rate": {"won": 428, "decided": 432, "rate_pct": 99.1, "excluded": 26,
"definition": "complainant wins (transfer + cancelled) divided by decided cases ..."},
"matched_names": [{"name": "Microsoft Corporation", "n": 428},
{"name": "Microsoft Corporation and Skype", "n": 22}],
"matched_names_total": 5,
"matched_names_note": "The counts cover all 5 party names listed. ..."
counts sums to total, so one call answers “how often does Microsoft win”. The win rate is defined in “Win rate”. For a respondent it is how often complaints against them succeeded; for a panelist, how often their panels found for the complainant.
matched_names lists the recorded spellings the query matched, most frequent first, 25 at most (matched_names_truncated and matched_names_total mark more). Word matching is loose, and every matched party’s cases are in the counts: complainant: "apple" matches 21 spellings, Apple Inc. and Apple Computer but also Applebee’s, Apple Vacations, Apple Corps, Snapple, Kappler, Pineapple Studios and the golfer Stuart Appleby, so its 96.3% is not Apple Inc.’s rate. "apple inc" still catches “Kappler, Inc”; the matched names say which parties a figure includes.
Parties
Parties are lists, each value with its own provenance:
"respondents": [
{"role": "respondent", "seq": 0, "name": "Carolina Rodrigues", "source": "delimited"},
{"role": "respondent", "seq": 1, "name": "Fundacion Comercio Electronico", "source": "delimited"}
]
role is complainant, respondent or panelist; seq is the provider’s order, and the first complainant is the lead filer. source says how the value was obtained:
source | Meaning |
|---|---|
provider | The provider stated it — authoritative |
structural | Separated from the provider’s own listing markup |
delimited | Separated from a provider string by a known delimiter |
llm | A model separated a value whose delimiter was lost — a reconstruction |
- Respondent lists often include privacy services and registrars beside the registrant; they are kept, so the names are what to read.
- An empty
complainantslist means the complainant could not be determined, not that nothing was filed; aparties_noteis attached where this matters. Complainant coverage, measured 2026-09-05, was all but 2 of the ~73,200 WIPO decisions then held and ~91% of Forum, where it is recovered from the case caption.
Disputed domains
A lookup decision lists them in domains: each entry is an object with seq, domain and source (provenance as above). About 15 domain rows (measured 2026-09-05) hold several domains run together, e.g. nexusautopartsusa.comnexusautopartsus.com; a lookup marks them malformed: true where it can, and a domain lookup on one of the run-together names misses that case. Detection is deliberately conservative (a repeated TLD, or three or more markers), so an absent flag is not proof a value is clean: dolphins.comjets.com is a valid name. Where an entry shows a TLD in a non-final position without the flag, the case number is reliable and the domain list is not. source is provenance, not a quality claim: a fused value can genuinely come from the provider’s listing.
Panel
panel lists the panelists who decided the case, one row each with seq. They are adjudicators, not parties, which is why they have their own lookup mode. It is present on about 83% of decisions (96,826 of the 116,942 in the corpus when measured on 2026-09-05): metadata-only cases never had a panel, and about 2,100 decided cases carry a panel line that could not be separated, so an absent panel means “not recorded”. A UDRP panel has one member or three (93,575 and 2,548 decisions); the corpus also holds 647 two-member “panels” that are one name split in the wrong place, and 56 with more than three.
Counting by panelist name undercounts. There is no entity resolution: Neil Anthony Brown alone has over 200 spellings (“The Honourable Neil Anthony Brown QC”, “Hon. Neil Brown”, “the Hon Neil Brown Q.C.”). A word search finds them; a count from one exact spelling is a lower bound, though the rate over the cases found is fair.
Decision text
Text is omitted by default (the median decision is ~14,600 characters). include_text: true with a case_number returns it in full_text, 8,000 characters at a time, and the response always says what happened:
text_available: falsewith a note for a metadata-only decision.- A longer text has
text_truncated: true,text_chars(the whole length) andnext_text_offset, and the window ends with a line such as[continues at character 8000 of 14600: datapulse_disputedb_lookup(case_number="D2024-0001", include_text=true, text_offset=8000)]. That line marks a page boundary, not lost text.next_text_offsetpassed astext_offsetreads the next window; the last window has neither. Offsets count characters, not bytes. - On any other lookup the flag is ignored and the note says so.
Not found
A case number that does not exist answers found: false with total: 0, an empty decisions list and a note: no such case, not a transient failure. A domain, party or panelist with no matches returns total: 0.
datapulse_disputedb_stats
Outcome counts and the complainant win rate for the whole corpus, overall, by_source and by_year. The answer is always complete: about 11 KB, one row per year, no paging.
| Parameter | Type | Required | Description |
|---|---|---|---|
source | string | No | wipo or forum; omit for both |
from_year, to_year | integer | No | Inclusive, 1999 to next year; anything else is refused with the bounds. A year filter keeps only dated decisions |
| Field | Meaning |
|---|---|
filters | The filters applied ({} for none) |
total, counts | Decisions in the slice, and per outcome; counts sums to total |
complainant_win_rate | won, decided, rate_pct (null when nothing was decided) and excluded |
win_rate_definition | The definition below, in words |
by_source | The same figures for wipo and forum |
by_year | Ascending; the last row has year: null and holds the undated decisions, so the year rows sum to total |
undated | Decisions with no date |
first_decision_date, last_decision_date | The dated range; the latest year is partial up to last_decision_date |
note | The headline figures in a sentence |
Win rate
Complainant win rate = wins (transfer + cancelled) ÷ decided cases (transfer + cancelled + denied + split decision). terminated, withdrawn and case active are excluded: no panel ruled, and most ended in settlement. Counting them as losses would understate the rate by about 13 points (78.7% against 92.0% across the corpus). A rate is meaningful only with this definition and its won/decided figures. Lookup’s party rates use the same definition.
Measured 2026-10-19: complainants won 93,080 of 101,141 decided cases (92.0%); WIPO 92.2%, Forum 91.8%. The figures in a response are live and supersede these.
Four caveats apply to the counts:
terminatedis WIPO’s word,withdrawnForum’s, for a case that ended without a decision. Providers compare on the win rate, not on those two counts.- 17,183 decisions have no date (measured 2026-10-19), almost all terminated or withdrawn. Any year filter drops them.
- The latest year is partial;
last_decision_datesays how far it runs. - WIPO and Forum only, so these are not figures for every UDRP provider.
Errors and retries
A connection-level failure (nothing listening, or a pooled connection the server had already closed) is retried once automatically. A request that still gets no answer, or no answer within 30 seconds, is reported as “A DataPulse back-end service could not be reached or did not answer in time. This is a server-side problem, not the input; retry in a minute.”
A request the server answered is never retried. A 4xx names the mistake (an unknown source_type, a party query under 3 characters); a 5xx such as embedding failed is an upstream outage that will not clear within a call.
Generated from the live server (DataPulse MCP v0.9-341-g27163c2) on October 20, 2026. Your AI assistant reads this page by calling datapulse_help(topic="disputedb").