Cross-border intellectual property as structured, cited, versioned data, with deterministic engines on top.
$ ipatlas compare trade_mark SG US CN -a filing_system -a term -a use_requirement
| Attribute | SG | US | CN |
|---|---|---|---|
| **Filing system** | first to file | first to use | first to file |
| **Term** | 10 years from filing date, | 10 years from registration date, | 10 years from registration date, |
| | renewable for 10 years | renewable for 10 years | renewable for 10 years |
| **Use requirement**| 5 years of non-use, then | 3 years of non-use, then | 3 years of non-use, then |
| | revocation on application | cancellation on petition | cancellation on application |
## Where they differ
- Filing system
- Term
- Use requirement
## Sources
[^1]: SG — TMA 1998, s 5 and Part III generally (UNVERIFIED, checked 2026-09-12)
[^2]: US — 15 U.S.C. §§ 1051, 1057(c) (UNVERIFIED, checked 2026-09-12)
[^3]: CN — 商标法 第31条 (Trademark Law, Art 31) (UNVERIFIED, checked 2026-09-12)
...
Status: Phase 1 complete. Data model, loader, linter, comparator, office calendars, deadline engine, lifecycle (terms/renewals/copyright) and filing-route matrices are built and tested (131 tests). Packs for SG, US, CN; office packs for IPOS, USPTO, CNIPA, WIPO, EPO, EUIPO; treaty packs for Paris, PCT, Madrid, Hague. The exhaustion, takedown, portfolio and transfer engines are designed but not written — see PLAN.md for the design and phasing, CLAUDE.md for the engineering contract.
Nothing in the data is verified. Every fact carries verified: false and every output
surface says so. Do not rely on a cell without checking its citation.
Every cross-border IP question has the same shape: for right R in jurisdiction J, what is the rule on X, what is the date, and what is the source? Today those answers live in paywalled country guides, in foreign associates' heads, and in docketing systems that cost six figures. None of it is open, structured, or callable by an assistant — so when an AI is asked one of these questions it answers from memory, which is where the errors come from.
IP law is unusually tabular: the treaties (Paris, TRIPS, Berne, PCT, Madrid, Hague) harmonise the categories, so the same attributes exist everywhere and only the values differ. That is exactly the shape a comparator wants.
| Command | What it does |
|---|---|
ipatlas jurisdictions |
List loaded packs, their offices, fact counts and available rights |
ipatlas compare <right> <JUR...> |
Cited comparison table (markdown / CSV / JSON) with a "where they differ" section and gaps listed as gaps |
ipatlas brief <JUR> <right> |
One-page country note assembled from the pack, with unrecorded attributes shown as unrecorded |
ipatlas fact <JUR> <path> |
One fact with its citation, check date and notes |
ipatlas offices |
Office packs and their closure-data coverage |
ipatlas deadline <kind> <JUR> <date> |
Paris priority, PCT national phase, opposition, Madrid refusal — with a full derivation |
ipatlas term <JUR> <right> --date base=YYYY-MM-DD |
Expiry and renewal schedule with grace and restoration |
ipatlas copyright-term <JUR> <category> |
Copyright term expressions, including "whichever is earlier" limbs |
ipatlas routes <right> <JUR...> |
Filing-route matrix with the traps the packs record |
ipatlas lint |
The data quality gate; exit 1 on errors |
--as-of YYYY-MM-DD |
Answer under the law in force on that date |
$ ipatlas deadline priority CN 2026-03-13 --right trade_mark
Paris priority deadline for trade mark filing in CN
the day of first filing is excluded from the priority period (Paris Convention, Art 4C(2))
period: 6 months (pack CN:trade_mark.priority: 商标法 第25条 (Art 25)) [unverified]
trigger: 2026-03-13 (Fri)
office: CNIPA (China National Intellectual Property Administration)
month arithmetic (corresponding_date, 专利法实施细则): 2026-03-13 (Fri) + 6 months = 2026-09-13 (Sun)
2026-09-13 (Sun): Sunday. Rolled forward to the next day the office is open, 2026-09-14 (Mon)
DEADLINE: 2026-09-14 (Mon)
CNIPA closure data 2026: UNVERIFIED, source=State Council annual holiday notice
caution: INCOMPLETE. Holiday blocks are approximate and the compensating working weekends
are NOT recorded, so any computation crossing a block must be checked.
WARNING: the period was read from pack CN:trade_mark.priority, which is marked unverified
Deadlines compute against the office that receives the act, not the jurisdiction. Periods come from the packs, so the citation travels with the answer. Three properties worth noting:
- The office calendar refuses rather than assumes.
ipatlas deadline madrid-refusal CN 2026-04-01exits 3: an 18-month window lands in 2027 and the CNIPA pack has no 2027 closure data. Chinese holidays are set annually by State Council notice and include compensating working weekends, so padding the pack with guessed dates would be worse than the refusal. - Days and months are different. US trade mark opposition is 30 days from publication; Singapore's is 2 months; China's is 3 months and not extendable. The engine reads whichever the pack records and says which.
- Unverified periods produce a warning on every result, not just in the data file.
$ ipatlas term US trade_mark --date filing_date=2020-03-01
missing date: US trade_mark term runs from registration_date, which you did not supply.
Supplied: filing_date. This matters: a term running from registration expires later than
one running from filing.
That refusal is the feature. Singapore trade mark terms run from filing and US and Chinese
terms run from registration, so the same mark expires on different dates — and an engine
that quietly accepted the wrong date would hide exactly the divergence the dataset exists to
record. Renewal schedules carry the grace and restoration windows; registered_design stops
at the recorded maximum duration.
Copyright terms apply the recorded expression, including two-limb rules:
$ ipatlas copyright-term US works_made_for_hire --date publication=2000-01-01 --date creation=1990-01-01
primary: publication 2000-01-01 + 95 years = 2095-01-01
alternative: creation 1990-01-01 + 120 years = 2110-01-01
rule: whichever is earlier -> 2095-01-01
Same author's death: Singapore and the US both give 2070-12-31; China gives 2050-12-31. All three run terms to 31 December (SG s 114, US 17 U.S.C. s 305, CN Art 23), so the divergence is the period - 70 years against 50 - and a work can be in the public domain in China while still protected in the other two.
The engine warns when a jurisdiction does not declare to_end_of_calendar_year, because
silence would otherwise produce a date up to a year early. That warning is how the missing
US s 305 rule was found.
Rights covered by the three packs: trade marks, patents, utility models, registered designs, copyright, trade secrets, geographical indications, plant varieties (SG), plus transfer formalities and statutory takedown regimes.
pip install -e ".[dev]"
python -m pytest # 131 tests
ipatlas lint # 0 errors, 2 known research gapsPython 3.11+.
Every leaf is a Fact: a value plus the provenance that makes it checkable.
term: { years: 10, from: filing_date, renewable: true, renewal_years: 10,
cite: "TMA 1998, ss 18-19" }Short form expands to a full Fact; a file-level defaults: supplies checked and
verified. A fact with a value and no citation is a lint error — an uncited value is a
rumour, not a fact.
Laws change, so facts have validity periods and queries have an as_of:
$ ipatlas fact US patent.filing_system
first to file [unverified]
source: 35 U.S.C. § 102 as amended by the Leahy-Smith America Invents Act
$ ipatlas --as-of 2010-01-01 fact US patent.filing_system
first to invent [unverified]
source: 35 U.S.C. § 102(g) (pre-AIA)Superseded rules move into a history: block rather than being overwritten. The packs carry
real examples: the AIA shift to first-to-file (16 March 2013), Lexmark moving US patent
exhaustion from national to international (30 May 2017), China's design term going from 10
to 15 years (1 June 2021), and Singapore's 1987 Copyright Act giving way to the 2021 Act.
ipatlas lint enforces what CLAUDE.md promises: citations present, checked dates
present and not in the future, staleness (warn at 12 months, error at 24), closed
enumerations, unsettled values carrying an explanation, non-overlapping history, and
verified: true backed by a note recording what was checked. Negative facts ("no utility
model system here") are held to a warning, because you can rarely cite the absence of a
provision.
A genuinely un-researched attribute is written as a note with no value, so the linter reports it as an open gap:
shorter_term_rule:
notes: ["Not yet researched. Record whether SG applies Berne Art 7(8)."]Never applies: null — a null is not an answer.
The three packs are chosen to make the divergences that matter visible:
- Trade mark term runs from different events. Filing in SG, registration in US and CN — so the same mark expires on different dates in each.
- The US § 8 declaration of use has no analogue in SG or CN, and is the most commonly
missed US trade mark deadline.
maintenanceis a separate attribute fromrenewalfor exactly this reason. - Patent grace periods are asymmetric. 12 months for an inventor's own disclosure in the US and SG; 6 months on enumerated grounds only in China. A US-led programme relying on § 102(b)(1) destroys novelty in China.
- Patent annuities differ in kind. Three fees from grant in the US; annual from filing in SG and CN. Docketing on an annuity assumption mis-handles US cases.
- Trade mark assignment recordal is constitutive in China and merely protective in SG and the US — and a Chinese patent assignment to a foreign party is a technology export requiring clearance, which is a closing condition rather than a formality.
- Chinese trade mark exhaustion is recorded as
unsettled, with a note, rather than forced into a clean value. That is the honest answer and the schema permits it.
The exhaustion decision table, takedown notice validation, portfolio dockets with ICS export, and transfer-formality checklists. All designed in PLAN.md §8; phased in §9.
Known coverage gaps in what is built:
- Closure data is thin. IPOS and USPTO have 2026–2027; the rest have 2026 only. Every
year is
verified: false, and the 2027 IPOS set is projected rather than gazetted. - CNIPA compensating working weekends are not modelled. The field exists and is deliberately empty: a wrongly-designated working day shortens a deadline, so guessing is the dangerous direction.
- EPO, EUIPO and WIPO closure lists are projected, not taken from the published lists. EPO Rule 134(1) turns on all filing offices being open, which widens the effective set.
Official fees (they change quarterly and would drag the dataset into staleness), case-law retrieval, prior-art search, freedom-to-operate, and platform-specific programme rules. Statutory regimes only.
One jurisdiction, one file, one owner. data/jurisdictions/SG.yaml is the reference shape.
Read CLAUDE.md first — it binds humans and AI assistants equally. A pack PR lands with its
golden tests and passes python scripts/lint_data.py.
Code MIT. Data (data/) intended as CC BY 4.0 — see PLAN.md §12 open question 4.