Five AI agents read your FIDIC contract, surface the evidence, build the dated chronology, draft the cited narrative, and quantify the claim. Every factual sentence is cited; every model call is hash-chained for audit.
One delay event in; a defensible, cited, quantified EOT claim out — through a fixed five-stage pipeline.
One LLM call per agent per click. Predictable cost, predictable latency, predictable failure modes. No chains, no streaming JSON, no tool-calling loops — each agent output is validated against a Zod schema before it touches the UI.
Surfaces every applicable Sub-Clause with a verbatim snippet and a notice / time-bar verdict. On the hero event: 5 cited clauses and a critical SC 20.2.4 84-day-deadline warning.
Ranks the 28-document corpus and returns the top matches with verbatim pull-quotes and 0–100 relevance scores. Runs on GPT-4o — proof the agent contract is provider-agnostic, both providers in one hash chain.
Builds a strictly-dated timeline from frontmatter dates parsed in TypeScript — the LLM never invents a date. It only labels and grades each of the 28 entries (critical · material · minor).
Writes the five-section EOT narrative. Every factual sentence carries a [DOC-XXX] / [CONTRACT-CL-X.X] / [PROG-XXX] citation; uncited claims are flagged [unverified] inline by the post-processor.
Quantifies prolongation cost — Hudson · Emden · Eichleay overhead + site + finance. The numbers run in TypeScript; the LLM only picks the recommended formula and writes the case-law rationale.
An agent_calls row is written for every agent invocation, with a SHA-256 row hash linking back to the previous row. The lawyer or arbitrator can replay the exact input, output, model, latency, cost, citation count and uncited-flag count of every step — and verify the chain end-to-end.
Mirzam Bridge Contractor JV · NTP 01 Mar 2025
A synthetic four-span composite rail bridge over a main highway corridor, authored from industry MENA delay typology research. The hero event is a textbook concurrent delay — an undisclosed 33 kV utility cable + a piling rig hydraulic failure — with Notice issued on Day 26 of the 28-day SC 20.2.1 window.
Red Book 2nd ed (2017) — Sub-Clauses 8.4 Advance Warning, 8.5 EOT, 8.6 Delays Caused by Authorities, 18.4 Exceptional Events, 20.2.1 Notice of Claim, 20.2.4 Fully Detailed Claim.
Delay & Disruption Protocol 2nd ed (Feb 2017) — Time Impact Analysis (MIP 3.7), As-Planned vs As-Built Windows (MIP 3.3), Collapsed As-Built, Impacted As-Planned, Core Principle 10 on concurrency.
International RP 29R-03 Forensic Schedule Analysis; RP 90R-22 Schedule Quality Index. Cross-validation of programme integrity.
Industry MENA delay root-cause typology research · UAE Civil Code (good faith, abuse of right) · DIFC arbitration seat practice.
Three capabilities the hackathon deliberately scoped out are live in the Lab — real, not mocked: AI document ingestion, a programme / XER parser, and semantic retrieval on real OpenAI embeddings + pgvector. Sign in to try them on the demo project.
Paste or upload a raw document — a real LLM pass extracts the structured fields a claim needs (type, date, parties, title, event links).
The parsed baseline activity network with critical path, plus a decoder for a pasted P6 XER export's TASK table.
Semantic search over the corpus on real OpenAI embeddings + pgvector cosine — the retrieval layer that scales past full-context.
No. Causa Claims produces a cited first draft to accelerate the Contractor's team. A qualified QS and construction counsel review and sign off before anything is submitted.
Every factual sentence must cite a [DOC-XXX], [CONTRACT-CL-X.X] or [PROG-XXX] source. A post-processor validates each citation against the real source-ID set and flags any uncited factual sentence as [unverified] — so a hallucinated claim is visually loud, not hidden.
No. Chronology dates are parsed from document frontmatter in TypeScript, and the Quantum formulae (Hudson / Emden / Eichleay) run in TypeScript. The model only writes prose and picks among typed options — it never computes a number.
The demo runs on FIDIC Red Book 2017. The methodology library and clause set extend to Yellow / Silver and the 1999 suite in Phase 1.
The hackathon demo uses a fixed synthetic project. Bring-your-own-data — document upload, P6/XER programme import and corpus-scale retrieval — is the first Phase-1 milestone.
Every tenant is isolated by Postgres row-level security. Each agent call is written to a SHA-256 hash-chained audit row a lawyer or arbitrator can replay end-to-end; tampering with any field breaks the chain.
It is FIDIC-aware, enforces a hard citation contract, computes quantum deterministically in code, and produces an arbitration-grade audit trail — none of which a general drafting tool does.
Sign in as the Contractor's PM, QS or Counsel and run all five agents on the hero event — Notice issued on Day 26 of 28.