The proof is in the Accuracy Score.
Anyone can claim their AI is accurate. We publish the rubric, ask a real multi-document council question, and score every answer the same way — including our own.
One question. Twelve months of packets.
A question no single document answers — it has to be assembled from a year of minutes, then counted correctly.
Across all the council meetings in these minutes, how many total residential housing units were approved through public hearing votes?
- Context
- 12 months of packets/minutes
- Correct answer
- 928 units across 7 hearings
- Copilot8%
- ChatGPT54%
- Claude80%
- Madison Ai95%
Scores on this question, graded against the five dimensions below.
The 5 dimensions of an Accuracy Score.
Every answer is graded on the same five dimensions, weighted to what a government actually needs to defend a decision.
- 30%CorrectnessIs the answer correct?
- 30%CompletenessAll key facts included?
- 20%FaithfulnessBacked by docs?
- 10%RetrievalFound the right docs
- 10%AuthorityUsed official records?
How the score works
- Accuracy Score
- A weighted sum of all five dimensions.
- Our target
- 90%+ on every question.
- Automatic fail
- Any factual error caps the score at 70% — no matter how well the answer scores everywhere else.
No magic wand. Just hard work.
The same question, run two ways — and the difference is architecture, not prompting.
Drop 24 files into an LLM
The workaround most teams try first.
Still time-intensive. Staff re-upload, re-prompt, and re-verify every single time.
Inconsistent. Same question, different answer. Capped at ~20 files at a time.
Not trustworthy. No grounding, no citations. Nothing you can take to the dais or present to a citizen.
Madison Memory + Intelligence
Two layers, built for the whole task.
- Connect
- Convert
- Organize
- Ground
PDFs · Tabular · GIS · Permit data · ERP · Decades of files/decisions
- Analyze
- Synthesize
- Output
- Cite Work
Live in 65+ local governments























Run the same question against your own record.
We’ll load Madison with a sample of your files and score the answers in front of you.