What would improve the human capital management project?
Lawrence had already drawn the link between a project and a proposed intervention. In May 2023 he asked GPT-4 to find it. The notebook returned a column heading and no data rows. Can a current model recover what he put in the graph?
8 May 2023 · The original run
The question Lawrence actually asked
“what is the human_capital_management project improved by?”
In his notebook LR_generic_cypher_gpt4.ipynb, Lawrence adapted Tomaž Bratanič’s schema-to-Cypher tutorial to use his Unilever portfolio example. The notebook was designed to give GPT-4 the database structure and ask it to write a query. The saved import cell records a lost connection. Later cells contain the model’s response and retry, but the notebook cannot establish how much data or schema the model had at that point. Bratanič’s original tutorial.
The query GPT-4 retried
MATCH (p1:Project {name: "human_capital_management"})-[:IMPROVED_BY]->(p2:Project)
RETURN p2.name as Improved_Projects
The first response supplied this pattern alongside a warning that the relationship was missing from the schema it had been given. On retry it supplied the same pattern on its own. The saved result was [['Improved_Projects']]. That is a column heading followed by zero rows. A project manager reading it still has an unanswered question.
The 2020 source graph
The portfolio graph records a proposal to rescope
Rescope the large project that failed to deliver in 2020.
The portfolio model calls this an area of support. The accompanying portfolio review notes list the action under “Specific next proposed Areas of support”. That wording supports taking a proposal to a review. Evidence of implementation or of a later improvement would have to come from subsequent records.
Follow the three source lines
Line 88 · Starting projecthuman_capital_management_project
Project with import ID 90.
Line 183 · Outgoing relationshipimproved_by
Connects import ID 90 to import ID 69.
Line 70 · Proposed supportAreas_of_support
Import ID 69 holds the rescoping proposal.
Lawrence’s portfolio model is held in a private repository. The line numbers above refer to Final_desktop_version.cypher at revision b650d96, dated 3 September 2020. A public two-node reconstruction preserves the labels, exact values and direction needed to check this example. Inspect the two records and their relationship.
GPT-4 followed the arrow in the right direction. Three restrictions still hid this record. The stored name ends in _project, the relationship is lower-case, and the destination is an area of support. The query required the destination to be another project.
Try repairing the 2023 query
Change any of its three restrictions. This reader compares your choices with the one recorded relationship above.
21 November 2026 · Two fresh trials
What changed when the model could look up the project name?
Each trial gave gpt-6.2-astra Lawrence’s question verbatim and asked for one read-only Neo4j query. The model received the graph structure, covering 15 node labels and 25 directed patterns. Each trial started in its own session. The model could not inspect the graph, browse the web or see the 2023 query, previous attempt or expected intervention.
Graph structure alone
0 rows
The model chose the right relationship and destination type. It checked the project ID and also normalised the project name by trimming whitespace, lowercasing the text and replacing spaces with underscores. Its comparison still required an exact match to human_capital_management. The stored name has an extra word.
Read the model’s query
MATCH (p:Project)-[:improved_by]->(a:Areas_of_support)
WHERE p.id = 'human_capital_management'
OR replace(toLower(trim(p.name)), ' ', '_') = 'human_capital_management'
RETURN DISTINCT coalesce(a.name, a.id) AS improved_by
Graph structure and project names
1 row
The second session also received the complete, alphabetised list of three project names. It selected human_capital_management_project, traversed improved_by and returned the rescoping proposal. The list contained no intervention names.
Read the model’s query
MATCH (:Project {name: 'human_capital_management_project'})-[:improved_by]->(area:Areas_of_support)
RETURN DISTINCT area.name AS area_of_support
The result separates two tasks a team can check. The model used the schema to write the right traversal on both attempts. The project-name lookup let it choose the record the question referred to. For the station handover in Essai 01, the corresponding question is whether the chair can find the exact asset name or an approved alias before asking the model to retrieve lessons.
All three queries were executed in Neo4j Community 5.26.0 on a fresh import of the full 2020 graph, with 99 nodes and 134 directed relationships. The same three queries were also executed against the two-node public excerpt. One question, two fresh responses and a known answer give a worked check. The changed database conditions and incomplete 2023 run record limit what can be inferred about progress from GPT-4 to a current model.
Reproduce the query results
Open a fresh, empty Neo4j 5 database to keep the example separate from other records. The small Cypher file creates the source project, the proposed intervention and their relationship, then runs the three queries printed here. Execute the statements in order. The first two queries return zero rows. The last returns one row containing Rescope_large_project_which_has_failed_to_deliver_in_2020. This file reproduces query behaviour on the public excerpt. Two legacy statements for creating and dropping a temporary import constraint were omitted from the full import. The replay file needs no constraint or external dataset. A fresh call to a language model may generate a different query.