DecisionGraphProject Evidence Desk · decision memory
16 verified casesSynthetic demo evidence
Board readiness ↗
Comparable outcomes01 / 03

Before we approve this change

What happened
last time?

Retrieve comparable project decisions. Inspect why they match. See the intervention and the measured outcome before committing again.

01
New decision

Describe the change or delivery problem

Retrieval is deterministic. Recommendation language is constrained to the evidence shown.

02Comparable history

Retrieved precedents

The closest decisions we have seen

Evidence set3 of 16 cases · 91% top match
Relationship map

Project knowledge graph

New changeCaseDecisionOutcome
10 nodes · 9 evidence links
03Human decision
Evidence-linked proposal3 COMPARABLE CASES

Conditional recommendation

Approve with an interface freeze and protected regression test.

Historical outcome range3–8 DAYS

versus 18–24 days of untreated exposure

Why this action

The strongest two precedents controlled the interface boundary and protected testing. The closest failure compressed testing and moved cutover by 17 days.

Approval conditions
  • Name one interface owner
  • Freeze the affected design for 10 days
  • Protect the full regression-test window
Evidence boundary

This is a proposal from 3 synthetic precedents, not an autonomous approval. Confirm local safety, commercial and access constraints.

Local data import

Bring your decision history

Import JSON or CSV containing previous changes, decisions, risks and outcomes. Files are parsed in your browser and never uploaded.

Requiredproject · problem · decision · outcomeRecommendedsector · phase · risks · scheduleOutcome · costOutcome · evidence
Transparent retrieval

How DecisionGraph works

  1. Normalise the question. Sector, phase, change type and problem text become an inspectable query.
  2. Retrieve comparable cases. Token overlap and exact context matches produce a deterministic similarity score.
  3. Traverse the evidence graph. Each case connects the original problem, decision, intervention, outcome and source references.
  4. Synthesise within evidence. Recommendation language uses only the retrieved cases and highlights conflicting outcomes.
  5. Require human approval. A proposal does not enter memory until a reviewer records a decision and later adds the observed result.

Production RAG could add semantic embeddings and an LLM synthesis stage. This public demonstrator intentionally keeps retrieval local, deterministic and testable.