Built for major infrastructure programmes. Every finding cited to the exact paragraph, with human in the loop.
Every assessment criterion runs through an eight-step pipeline designed for auditability — not speed. Retrieval, verification, deterministic scoring, and self-critique. The extended tier adds multi-persona debate and cross-criterion validation for the highest-stakes decisions.
Dense vector search and BM25 keyword search, run in parallel across your document set. Every passage scored for relevance — nothing buried by embedding noise alone.
RetrievalBefore reasoning begins, the system grades its own retrieval. If the top passages don't meet a confidence threshold, it re-searches with reformulated queries. No assessments built on thin evidence.
Self-correctingComplex criteria are broken into specific yes/no sub-questions. "Is the risk register versioned?" "Does it include residual risk?" Each decided independently against the evidence.
DecompositionRatings are calculated from the sub-question results using a defined rubric. The same evidence produces the same rating every time. No temperature, no improvisation.
DeterministicEvery quoted passage is re-read against the source document. Citations that don't match verbatim are rejected and the reasoning is re-run. Hallucinated citations never survive.
VerifiedThe pipeline reviews its own output against the rubric and flags weak reasoning, missing evidence, or logical gaps. Borderline ratings trigger a second pass.
ConditionalThree AI specialists — each with a distinct role — independently assess the same evidence. Disagreements are surfaced to a human reviewer. No consensus-by-default.
Extended tierThe final pass checks for contradictions across the full assessment. If risk is red but the financial case calls it manageable, the system flags it. Consistency isn't optional.
Extended tierEvery finding can be accepted, challenged with context, or force-overridden — with mandatory justification. Each action writes to a tamper-evident audit trail that flows directly into board-ready reports. Try it below: the panel on the left is live.
Every finding traces to a specific paragraph in a specific document. Scroll to watch the chain form: finding → source line → quoted text → confidence score. This is the chain that goes into your board report.
Against criterion BC-14 — financial case — the scoring step returns amber. The finding card renders on the reviewer's queue with the criterion reference, rating, and short rationale.
The system knows exactly which paragraph the finding is grounded in. Hovering or clicking the finding draws a line back to the cited passage — no hunting through the document.
The cited paragraph is brought into focus and the quoted span is surfaced in context. You see the exact phrase the system relied on — not a paraphrase, not a summary.
Alongside the rating sits a confidence score derived from retrieval quality and sub-question agreement. Low confidence is a signal — not a verdict. You decide what to do next.
When a finding needs action, the platform drafts the remediation content — section text, risk entries, mitigations — grounded in your own documents. Every generated field carries a confidence score and a citation chain. Hover any field to see where it came from.
Consultants edit in place, approve, and push to the source — no copy-paste between tools, no lost lineage.
Following review, optimism bias will be revised to 40%Green Book A.1 · p. 22 · 94% at outline stage, in line with HMT Green Book Annex A.1 for standard civil infrastructure. This will reduce to 22%Green Book A.1 · post-QRA · 91% following completion of the quantitative risk assessment scheduled for Q3 2026Programme plan · §3.2 · 87%.
Supporting analysis — including the P80 confidence intervalRisk register · §7.2 · 93% — is attached in Appendix C. All figures reconcile to the base cost estimate of £127.4MCost plan v4 · §2.1 · 96%.
Export the full assessment to PDF or DOCX with one click. Every rating carries its rationale, its citations, and the full audit trail of human decisions. Nothing decorative. Everything defensible.
SharePoint libraries are monitored for change — and answered through a chat interface that cites every sentence back to the source.
A 30-minute discovery session. We'll assess one of your documents against the framework of your choice — so you can see what the platform finds, not just how it looks.