Before you rely on a case, check what happened next

You have found a case that appears to support your advice or submissions. Before you rely on it, you need to know what later courts did with it.

That is not a trick question. It is routine legal work.

A name search gives you every mention. It does not sort those mentions into “followed”, “distinguished” and “declined to follow”. It does not start from an adverse treatment and assemble only the decisions that came afterwards. It does not show a later court using a rule through an intermediary whose name you did not know to search.

You can reconstruct all of that. The question is how many searches and judgments you must work through each time.

Three jobs, not fourteen graph operators

Our first graph benchmark contained fourteen frozen questions. That was useful for testing whether the database could perform temporal joins, treatment splits and two-step paths. It was not evidence that lawyers would ask for “a two-hop query” in practice.

The case map now presents the same evidence through three pieces of legal work.

1. Check an authority before citing it

A later CFA judgment declined to follow Li Kwok Cheung. Which later CFA judgments still cited it, and what did each say?

The frozen graph joined one extracted adverse-treatment edge to later citation edges and found three later judgments. The direct keyword arm found one of the three. More importantly, the map puts the passage behind the earlier treatment and each later citation in one place.

That is a chronology for the lawyer to inspect. A later citation is not “continued approval”, and the page does not call it that.

2. Answer the authority relied on by the other side

Which later CFA decisions followed or applied Chong Fung Yuen, which distinguished or declined to follow it, and what passage supports each classification?

Searching the case name plus “followed” produces one noisy list. Searching it again plus “distinguished” produces another. Both still leave the researcher to open and classify every passage.

The graph grouped four candidate judgments around the authority and exposed the quote used for every classification. The keyword arms missed extracted rows in both the supportive and adverse groups.

The word “candidate” matters. The extraction can misclassify a passage. The map is a faster worklist with evidence attached, not a substitute for deciding what the judgment held.

3. Trace a legal test through later cases

Where did later CFA judgments apply So Yiu Fung’s substantial-and-grave-injustice formulation, and which earlier authority should I read to trace its source?

Five later judgments in the frozen graph applied the formulation through So Yiu Fung without directly naming Chim Hon Man. Searching the earlier case name found none of them.

This path is deliberately described as a research lead. The second step is a citation. A citation alone does not prove that the same legal proposition passed through both judgments. That stronger claim needs proposition-level review.

What the graph contributes

A search index stores documents and the words or vectors that make them retrievable. The graph stores a relationship as a separate, reusable candidate fact:

This judgment treated that authority in this way, on this date. Here is the passage from which the relationship was extracted.

That stored work can be joined and reused instead of reconstructed from scratch for every question. It can also be audited: every one of the benchmark’s 119 proof legs carried a quote.

The public page shows the three workflow summaries and frozen scores without starting the graph database. Sign-in is required only to wake the live, read-only proof view and retrieve the current passages.

What the benchmark proved—and did not

The G4 benchmark was frozen before the production keyword calls. It ran over 2,745 Court of Final Appeal judgments. Direct keyword retrieval lost five of six cross-document questions and found none of the thirteen answers to the three graph-authored two-step questions. Ordinary doctrine-and-section and co-citation controls were not graph wins.

That proves that the database can perform these joins against a deep whole-court citation web. It does not prove:

• that the fourteen benchmark formulations reflect daily legal work;
• that every extracted relationship is legally correct;
• that a two-step citation path carries one proposition throughout;
• that the graph produces a better final legal answer; or
• that its benefit transfers to a shallow or more uniform corpus slice.

Those are separate questions. The next gate uses de-identified, lawyer-authored research tasks and compares complete work under equal search and reading budgets. It must measure material-authority coverage, adverse-authority coverage, treatment accuracy, proposition continuity, practical usefulness, time and cost before graph retrieval earns a place in hosted research.

What remains the lawyer’s job

Keyword search remains the right starting point for ordinary subject research. The graph may add, order and explain candidates. It may never remove a judgment from a comprehensive search.

The three public examples are quote-backed, machine-extracted research leads. They have not yet received independent legal review. Read the passages, read the judgments, confirm that the deciding court expressed the treatment, and check that the proposition you care about is the one the later court addressed.

None of this is legal advice, creates a solicitor–client relationship, or replaces reading the authorities.

The one line to take away

Search finds the documents. The case map stores some of the cross-document work that a researcher would otherwise have to reconstruct repeatedly—and shows the passage behind each proposed link.