Rigour you don’t have to take on trust.
This is how CommonBench actually works — how an authority is verified, what the corpus covers, how your matter is kept confidential, and where the tool’s responsibility ends and yours begins. Read it and judge for yourself.
Why the method matters.
Generative AI is fluent, and fluency is the trap. A model that returns a confident, well-formatted citation to a case that does not exist leaves the reader worse off than no tool at all — and for a practitioner, that is not an inconvenience but a misstatement to a court, made under their own name.
Everything below exists to make sure that does not happen here.
How an authority is verified.
Three things stand between a model’s fluency and what reaches your screen.
Retrieved, not generated.
Every case and statute CommonBench works from is drawn from a curated corpus of real, reported authorities — not conjured by the model. The model drafts in prose and can name an authority on its own initiative, so what matters is what happens to each citation next.
Checked against the source.
After an answer is written, every citation in it is extracted and matched — by name and citation — against the corpus, and where needed against the free public databases the courts and legislatures publish to: BAILII, HKLII, AustLII, CourtListener, and the official legislation sites. A citation is marked verified when it matches a real, reported authority. Where a stable free link exists we surface it so you can open the source yourself; for some authorities — Singapore judgments and older report-series citations among them — no stable per-case public link exists, so we point you to the right database to search.
Removed if impossible; flagged if unconfirmed.
A citation that is structurally impossible — a future year, or a court that did not exist in the year cited — is removed before the response is shown and replaced with a visible marker, as is any citation in a form the verifier cannot parse. A citation we have checked but cannot confirm is left in place and clearly flagged, rather than quietly dropped, so you can see exactly what the tool could and could not stand behind.
It confirms an authority is real and correctly cited — a strong signal, not a guarantee. It does not confirm that the case remains good law, that the proposition it is cited for is correct or complete, or that every link will resolve. A flagged citation is not necessarily wrong; it is one we could not independently confirm. That judgment stays with you, which is exactly where it should sit.
What’s actually in it.
A research tool is only as good as its scope, so here is ours, in full. CommonBench draws on 4,900+ verified authorities across five jurisdictions:
Arbitration sits across jurisdictions rather than within one, so it is held as a dedicated module — the law that governs a reference, held current for every major seat:
Commercial arbitration is private: awards are confidential and rarely published, so no tool can hold a complete record of them — and CommonBench does not pretend to. But an arbitration is decided on the substantive law that governs it — English law in a London reference, Singapore law in a Singapore one — and that is the very common-law authority the corpus above already holds in depth. So the authorities cut both ways: thin as the record of awards is, the heart of a dispute — the law the tribunal actually applies, read alongside the rules of every major seat and the statutes and conventions behind them — is firmly covered. The procedure is private; the law it turns on is not.
Where coverage of an area is thin, the tool is built to say so rather than reach for the nearest authority that happens to exist. The corpus concentrates on reported, citable authority — the apex and appellate judgments most likely to bind or persuade, together with the principal statutes — and it grows as new decisions are reported.
One engine, five systems.
The five jurisdictions share a common-law inheritance, but each has its own authorities, conventions, and procedure. Reasoning is locked to the jurisdiction you select, so the answer reflects the law that actually governs — not a blend of five. Which jurisdiction governs a cross-border matter is itself a question of law, and that one stays with you.
Your matter, kept yours.
Confidentiality is a precondition of legal work, not an add-on. In summary — and subject to the Privacy Policy and Terms, which govern:
CommonBench is headquartered in Singapore. Depending on the service, your information may be processed in Singapore, the United Kingdom, the European Economic Area, and the United States.
Your queries and work product are kept for up to twelve months by default, and you can delete an individual chat — or your entire history — at any time from your account.
Your content is not used to train foundation models without your explicit, opt-in consent. Our inference providers operate under zero- or short-retention terms and are contractually barred from training on it.
Data is encrypted in transit (TLS 1.2+) and at rest (AES-256), under role-based access control with least privilege and multi-factor authentication for staff access to production systems.
The providers we rely on — cloud hosting, AI inference, payments, email — are each bound by a written data-processing agreement and act only on our instructions.
On the Chambers tier, each seat carries its own monthly allowance; one seat’s usage does not draw on another’s.
Your relationship with CommonBench is governed by the law of Singapore, with the Singapore courts as the exclusive forum (or SIAC arbitration by election).
These are summaries. The Privacy Policy and Terms set out the full position and govern in case of any difference.
What it won’t do.
CommonBench is a research tool, not a lawyer, and it does not pretend otherwise. It will not advise on a specific matter, will not appear for you, and should never be relied on without checking the authorities yourself. When the tool is uncertain, it is built to say so rather than perform a confidence it has not earned.
Watch it work.
The homepage runs a real enquiry end to end — the engine reads the facts, flags what’s urgent, answers, cites verified authorities, and scores the matter. Everything above is what makes that sequence trustworthy.