Progress monitor
Overview
Updated
Can a local AI model read real quotes and invoices well enough to pre-fill a Conduit request? We are measuring it in the open.
Loading the latest status
Next:
Extraction
0 of 0 documents
Overall completion
0 of 0 phases
Extraction
The extraction process is not running. It resumes where it left off when restarted; no finished document is lost.
0 of 0 documents
Documents run smallest first, so the last few are the longest. Text volume is the better guide to how much work is left.
Complete 0Partial 0Failed 0
Now reading
Recent pace
Measured ETA
Latest scored round
When will this be done?
The extraction time is measured. Everything after it is an estimate. The early date assumes round 2 passes; the late date includes a third round.
Who does what
| Role | Who | What it does |
|---|---|---|
| Reader | Qwen3.8-27B, running locally in LM Studio on Albert's Mac | Reads every document and returns facts and line items in a fixed JSON format. It never decides where a value goes in Conduit. |
| Reviewer | Claude Opus, with vision | Labels a sample of documents from their page images, sorts every error into a failure bucket, and proposes changes for the next round. |
| Scorer | Python scripts | Compare the reader's output with the labels and compute every score. No model grades its own work. |
| Approver and critic | Claude Fable | Approved this plan, will approve the results, and will try to break the build plan before anyone builds it. |
| Planner | A BMAD product manager agent (Claude) | Turns the results into a build plan for the AI assist feature. |
This site is kept current by a Claude Sonnet agent that edits the narrative only. Every number comes from a script that reads the evaluation files directly.