About Services Tools Blog Contact us

Structured Case Decisions

Asks Qwen, running on our own machine, a fixed set of questions about a case and returns strictly formatted answers, checked against a published reference.

  • Classification & Tagging
  • Cross-sector
  • Tool
  • Productised
  • Local/private
  • 2026

The problem

Support and risk decisions need the same fields every time. A free-text paragraph gets read differently by each reviewer. An answer that cannot be read reliably is worse than none.

What it does

The tool loads a case file, builds a question from its details and asks Qwen through Ollama. The reply is split into fields such as urgency, department, refund eligibility and risk. A benchmark repeats a small set of cases and checks the answers against a reference. The tool refuses to run when it is called from a website.

Outcomes

  • Every case gets the same fields.
  • Each field gets a pass or fail mark against a published reference.
  • A benchmark report is produced before a model change is trusted.
  • The decision tool cannot be started from the public website.

Simplified mockup

Interactive preview with sample data. Nothing is sent to a server.

Four shop cases

Fixed questions
Question list
Urgency Score

Low or high

Department Choice

Returns, fitting, accounts, shop floor

Refund Yes or no

Eligible, not a refund, or not applicable

Risk Score

Low, medium or high

Command line
Ready on this machine
Inbox Unparsed
CaseCustomerWhat they wrote
C-1842Priya NairThe parcel arrived crushed. Please refund order 4418.
C-1847Owen BlakeUnworn, and too big. I need a smaller size, not my money back.
C-1851Helen CroweI was charged twice for the same pair. Take the second payment off.
C-1855Tom AdeyemiWhen do the autumn boots arrive in the Maidstone shop?

Published baseline

Cases stable
4 / 4
Direction
16 / 16
Parse failures
0
Wall time by case Local model, temperature 0
Field checks
CaseUrgencyDepartmentRefundRisk
C-1842PassPassPassPass
C-1847PassPassPassPass
C-1851PassPassPassPass
C-1855PassPassPassPass
Repeat
Baseline is the reference

Web server refusal

Rule

The decision script checks how it was started. A command-line run is allowed. Anything arriving through the web server is stopped before a case is read and before the model is called.

Attempt
No request sent yet

How it works

  1. Load a case file and a fixed question list.
  2. Ask the model for strictly formatted answers.
  3. Split the reply into urgency, department, refund eligibility and risk.
  4. Show a pass or fail against the reference for each field.
  5. Repeat the published cases as a benchmark before trusting a new model.
  6. Refuse to run if called from a website.

Pipeline

flowchart LR
  A[Case JSON] --> B[Questions]
  B --> C[Qwen on Ollama]
  C --> D[Parsed fields]
  D --> E[Baseline check]
  E --> F[Pass or fail]