- A temporal knowledge graph stores linked facts, each with the date it was true and the date it was recorded.
- Delay disputes turn on what was known and when, which overwritten databases and loose document stores cannot show reliably.
- On a project, contracts, clauses, notices, activities and site records become nodes, and their links carry dates too.
- AI that answers from a temporal graph can cite its sources and give date specific answers without guessing.
The short answer
A temporal knowledge graph is a way of storing information as connected facts, where every fact carries the dates that matter to it. Each item, such as a contract, a clause, an early warning or a programme activity, is a node. The relationships between them, such as cites, affects or evidences, are links. Both nodes and links record when they became true and when they were entered on the record.
That last part is what makes the graph temporal. It lets you ask a question that ordinary systems struggle with: what was true on this date, and what did we know about it at the time?
Why ordinary records are not enough
A spreadsheet or a project database usually holds the current state. When a date changes, the old value is overwritten. A document store keeps every file, but the links between files live in people's heads. Neither can reliably reconstruct the project as it stood on a given morning.
Delay disputes depend on exactly that reconstruction. A tribunal wants to know what the contractor knew when it planned the work, when the employer was told of a problem, and whether a notice went out within the contractual period. If the record has been overwritten or the links have been lost, an expert has to rebuild them by hand, and the other side will challenge every step.
Two dates for every fact
Engineers call this structure bitemporal. Each fact holds a valid date, when it was true on site, and a recorded date, when it reached the record. Consider a diary entry for 4 March saying the piling rig stood for four hours. On 6 March the foreman finds the timesheets and corrects it to six hours.
A bitemporal record keeps both entries. Viewed as at 5 March, the record says four hours, because that is what anyone could have known that day. Viewed as at 7 March, it says six hours and shows that the earlier figure was superseded on 6 March. Nothing is deleted. The correction and its timing are themselves evidence.
What it looks like on a project
On a live project the graph links the things a commercial manager already works with:
- the contract and each clause that carries an obligation or a time bar
- early warnings and notices, with the clauses they cite and the date of delivery
- programme activities, their baseline dates and every update as issued
- site diaries, photographs and instructions, linked to the activities they affect
- compensation events or claims, linked to the evidence that supports them
Follow the links from a claim and you reach the clause it arises under, the warning that first raised it, the diary entries that record its effect and the programme in which the delay first appeared. Every step carries its dates.
Why AI needs it
Large language models are good at reading and writing, but they reason only over what they are given. Ask a model a date specific question about a project, using a folder of documents, and it may blend versions or miss the correction that changed the answer.
Give the same model a temporal knowledge graph and the question becomes a retrieval: fetch the facts valid on that date, as recorded by that date, follow the relevant links and cite each source. The answer is grounded, traceable and repeatable. For anything that may end up in front of an adjudicator or arbitrator, that traceability matters more than eloquence.
The graph is the memory. The model is the reader. Keep them separate and you can trust what the reader says.
How to start
You do not need to replace your systems. Most firms start by building the graph from records they already hold: the contract, the accepted programmes, the notice register and the diaries. From there, each new entry is added with its dates as the job runs. CIP does this natively for the projects it runs, and the same approach can be applied to a firm's wider knowledge as part of a business second brain.
Questions people ask
- Is a temporal knowledge graph the same as a database?
- No. A typical database holds the current state and overwrites old values. A temporal knowledge graph keeps every version of each fact with its valid and recorded dates, and stores the links between facts as data in their own right.
- Do we need to replace our existing systems?
- No. The graph is usually built from records you already hold, such as contracts, programmes, notice registers and diaries, and then kept up to date as new entries are made.
- How does a temporal knowledge graph help in arbitration?
- It allows the project record to be reconstructed exactly as it stood on any date, with corrections and their timing preserved, which is the basis of a credible delay analysis.