DIBER Core
DIBER Free
Free · Runs locally in your browser

Diber an AI response. You define the criteria; DIBER does the math.

Paste a response, specify what it should have contained, and classify each point using explicit guidance. The score is based only on rules, weights, and verifiable choices — without asking another LLM to judge it.

Start the audit
Privacy first

Nothing you enter leaves this page.

Your question, response, criteria, and report stay in your browser. There are no remote API calls or requests to external models.

1
Explicit ground truth

You define what matters.

2
Guided classification

Every choice remains traceable point by point.

3
Reproducible calculation

Same classified inputs = same report.

Guided audit

You can save your work in the browser, export it, and resume it later.

No draft
1 · ResponsePrompt + AI output
2 · CriteriaExpected material points
3 · ClassifyCoverage + quality
4 · ReportMetrics and details
This is for your own context: DIBER does not send it anywhere or interpret it automatically.
0 characters
It is only a reference for you. It is not used by an automated model.
How to use it effectively: don’t ask yourself “how good is the response?” yet. First define which points were materially required. The evaluation comes afterward.

Define the expected material points

Each point is something specific the response should have addressed. You can add a source/evidence and assign an importance level.

Tip: use atomic, verifiable points. “Explain the total interest cost” is better than “Explain the mortgage well.”

Classify what you observed

The guided mode uses plain-language labels. The original DIBER codes remain visible for transparency.

First: coverageDid the response address the point? Was it required? Was the necessary information outside the available context?
Then: qualityOnly when the point was addressed, choose whether it is correct, partial, inferred, overstated, or incorrect.
No double penaltiesIf you choose “required but missing,” DIBER automatically assigns omission.
Review question, response, and context
Question
Response
Context

            

            

          


          

            
Report calculated
Your deterministic audit

            

            


            

              
              
              
            

            

              

            

            

              
PointImportanceCoverageStatusWeightNote

            

            


            

              
              
              
              
              
              
            

          

        

      

    


    

      
Quick guide: how to get a useful audit
The quality of the report depends primarily on how you define the criteria and how consistently you apply the classifications.

      

        
1 · Break the response into requirements
One point = one fact, step, caveat, or conclusion that you can verify independently.

        
2 · Weight what truly matters
Use Critical only for elements whose absence or incorrectness materially changes the usefulness of the response.

        
3 · Classify coverage first
Decide whether the point was required and whether it was addressed. Only then assess the quality of what was said.

        
4 · Use a source when you can
In the “evidence” field, note the document, data point, or reference that supports your classification.

        
5 · Don’t use “hallu” as a synonym for error
Use it for fabricated or unsupported content; for a straightforward false statement, use “error.”

        
6 · Keep the report
Export the audit and JSON so you can manually compare later responses using the same criteria.

      

    


    

      
Coverage and Status, in plain language
DIBER separates “did it address this?” from “how well did it address it?” This separation prevents everything from being collapsed into a single opaque score.

      

        
Coverage
    
          
  • addressed — the point was substantively addressed.
    • 
          
  • should — it was required but is missing.
    • 
          
  • gap — the information was unavailable and the response correctly acknowledged that gap.
    • 
          
  • out_context — the required information was not available to the model.
    • 
          
  • not_relevant — the point was not required for that question.
    • 
        

        
Status
    
          
  • correct 1.00 — correct and supported.
    • 
          
  • partial 0.50 — correct but incomplete.
    • 
          
  • inference 0.25 — an inference added beyond the available basis.
    • 
          
  • over 0.25 — overstated certainty or scope.
    • 
          
  • error 0.00 — generic factual error.
    • 
          
  • math_error 0.00 — numerical or calculation error.
    • 
          
  • hallu 0.00 — fabricated content / unsupported citation.
    • 
          
  • omission 0.00 — required point that was omitted.
    • 
          
  • missing 1.00 — gap correctly acknowledged.
    • 
        

      

    


    

      
What it means to Diber a response
Turn a general judgment into an explicit sequence: criteria → classifications → weights → metrics.

      

        
1 · Your ground truth
You define the expected scope. The tool does not invent criteria.

        
2 · Explicit classifications
Each point receives a Coverage and Status classification, so you can trace exactly how the result was produced.

        
3 · Deterministic calculation
The same choices and the same weights produce the same report.

      

    


    

      
Metrics and formulas
The Free version uses the same conceptual model as the Core for a single audit. It does not calculate repeated-run statistics because you are evaluating only one response here.

      

        
Weighted coverage
Σ weight(addressed) / Σ weight(addressed + should)
Measures how much of the materially required content was addressed.

        
Addressed quality
Σ(status score × weight) / Σ weight, coverage=addressed only
Measures the quality of what the response actually addressed.

        
Effective information
Σ(weight × score) for addressed / Σ weight(addressed + should)
Combines coverage and quality into a single metric without hiding the underlying components.

        
Defect rate
Σ defect weight / Σ weight(addressed + should)
Used for omissions, errors, math errors, and hallucinations.

      

    


    

      
Privacy and local persistence
The page is designed to work without a backend.

      

        
No remote APIs
The calculation runs in JavaScript in your browser. This page does not automatically send your text to external servers.

        
Optional draft
The “Save draft” button uses your browser’s localStorage. You can delete it at any time with “Reset everything.”

        
Exports under your control
Audits and reports are downloaded only when you explicitly choose to export them.

      

    


    

      
What it does not do
A transparent method should also be clear about its limitations.

      

        
It is not an automatic fact-checker
It does not independently determine whether a statement is true or false. You classify it using sources or your own expertise.

        
It does not use an LLM judge
No model assigns the score behind the scenes.

        
It does not replace expert ground truth
In medical, legal, financial, or technical domains, audit quality depends on the accuracy of the criteria and sources.

        
It cannot measure stability from a single response
Standard deviation and confidence intervals require multiple runs; that is why they are available in the Core/Developer Hub, not in a single Free audit.

      

      
Explore a multi-run report →Go to the Developer Hub