datum
Request a pilot

Benchmarks · measured 31 July 2026

Tested against what courts actually decided.

We reconstruct project records from published judgments - Wembley Stadium, The Shard, FIDIC works, offshore fabrication - and run them through the engine. The court's own findings of fact are the answer key: a chronology nobody at datum authored. Every number below comes from the engine's own scorecard against committed truth files.

11published judgments as answer keys
100%document dating accuracy, all 153 documents
86%average event recall, strict matching
26/26cited factual sentences verified on the exported claim
Per case
CaseRecordDocument datingEvent recall
Walter Lilly v MackayWalter Lilly & Co v Mackay [2012] EWHC 1773 (TCC)The landmark London EOT judgment: full extension of time to practical completion. datum also drafted and exported the claim - 26 of 26 factual sentences carried verified citations.16 docs100%11/11 · 100%
Multiplex v HoneywellMultiplex v Honeywell (No. 2) [2007] EWHC 447 (TCC)Wembley Stadium: revised programmes as acts of prevention; the time-at-large argument dismissed.12 docs100%12/12 · 100%
De Beers v AtosDe Beers UK v Atos Origin [2010] EWHC 3276 (TCC)A delayed IT programme, not a building site - the same method reads any dated project record.14 docs100%11/12 · 92%
Obrascon v GibraltarObrascon Huarte Lain v AG for Gibraltar [2014] EWHC 1028 (TCC)FIDIC Yellow Book - the Gulf's contract family. Clause 20.1 notice timing decided the claim.14 docs100%11/12 · 92%
Bluewater v MerconBluewater v Mercon [2014] EWHC 2132 (TCC)Offshore fabrication: milestones, variation orders, notice of default, termination.14 docs100%11/12 · 92%
Thomas Barnes v BlackburnThomas Barnes v Blackburn with Darwen BC [2022] EWHC 2598 (TCC)Modern concurrent-delay analysis: employer-risk steel remedials against contractor-risk roof delay.14 docs100%11/12 · 92%
Costain v HaswellCostain v Charles Haswell [2009] EWHC 3140 (TCC)A critical-path fight over piled foundations - which delay actually drove completion.13 docs100%10/12 · 83%
Cleveland Bridge v SeverfieldCleveland Bridge v Severfield-Rowen [2012] EWHC 3652 (TCC)Steelwork for The Shard: 42 days of culpable fabrication delay found against the subcontractor.14 docs100%10/12 · 83%
Adyard v SD MarineAdyard Abu Dhabi v SD Marine [2011] EWHC 848 (Comm)An Abu Dhabi shipyard: the prevention principle and concurrency under English law.Regulatory correspondence - events are positions stated inside long letters.14 docs100%8/12 · 67%
Van Oord v AllseasVan Oord v Allseas [2015] EWHC 3074 (TCC)A GBP 10m disruption claim dismissed for poor substantiation - the cautionary tale datum exists to prevent.Claim letters recounting months of history - many events buried per document.14 docs100%8/12 · 67%
Fluor v SZHIFluor v Shanghai Zhenhua [2016] EWHC 2062 (TCC)Offshore wind foundations: weld cracking, NCR chains and re-inspection campaigns.Dense NCR documents carrying several dated facts each - the hardest extraction shape.14 docs100%9/12 · 75%
How the benchmark works
  1. Reconstruct. Each corpus is built strictly from the findings of fact in a published judgment (UK National Archives open data, source XML kept for provenance). Letters, instructions, notices, minutes and certificates are re-created with the court's dates, parties and quoted words - every page marked as a reconstruction.
  2. Run blind. The engine ingests the record exactly as it would a client's: OCR, classification, indexing, event extraction. It never sees the judgment.
  3. Score strictly. A truth event counts as recalled only if the engine surfaced an event on the right date (within two days) whose linked verbatim evidence carries the operative phrase of the court's chronology entry. Extraction is gated in code: an event without an exact, verifiable quote from the record is dropped before it ever reaches a reviewer.
  4. Publish honestly. Low scores stay on the board with their reason. The hardest shapes today are dense multi-event documents - NCR chains and claim letters that recount months of history - and single-run scores on a 12-event sample move by one event (~8 points) between runs.

Also in the harness: 96 offline tests on real services in CI, a deterministic CPM engine verified against textbook schedules, and a citation gate that blocks export of any unproven sentence. These corpora are rehearsal assets: real-shaped, with adjudicated ground truth we did not write. They complement - never replace - pilot validation on a genuinely real claim record with a practitioner's review.

Want the engine graded on your own record? That is exactly what a pilot is.

Request a pilot