Automation
Rules that decide where documents belong, and the learning loop that improves those decisions.
Learning loop
Classifier versions are separate from routing policy. The rules a decision was evaluated against live on the Rules Engine tab.
Current version
45 approved examples · illustrative 87% accuracy
Illustrative — not a measured result.
Feedback queue
Read-only — Lead Reviewer approves
| Document | Original type | Corrected type | Reviewer | Submitted | Validation | Actions |
|---|---|---|---|---|---|---|
| unknown-document.pdf | None | Invoice | Sarah Palmer | 4h ago | Pending | |
| letter-acme-vendor.pdf | Correspondence | Contract | Sarah Palmer | 9h ago | Approved by Emily Zhang | |
| resume-johnson.pdf Misread — this is an application, not an employment contract | Job Application | Contract | Michael Torres | 20h ago | Rejected by Emily Zhang | |
| invoice-missing-total.pdf | Invoice | Contract | Sarah Palmer | 6h ago | Pending Conflicting labels | |
| invoice-missing-total.pdf | Invoice | Correspondence | Michael Torres | 5h ago | Pending Conflicting labels |
Only a submitted classification creates feedback. A routing override never creates a classification label, and a correction alone never changes the current version.
Approved examples — build of new version v1.1
Not yet in use
45 base examples + 1 approved this session.
- letter-acme-vendor.pdfContractSarah Palmer · Next version
Rejected corrections are never included.
Evaluation
10 held-out labelled documents.
Every document behind an approved example is excluded in code. "Held out" here means a fixture list and nothing more.
Show the 10 documents
invoice-2024-0892.pdf, nda-draft-techstart.pdf, invoice-high-value.pdf, contract-renewal-autorenew.pdf, invoice-to-james-wilson.pdf, memo-to-david-kim.pdf, application-martinez.docx, invoice-attn-torres.pdf, notice-attn-maya-chen-legal.pdf, memo-from-sarah-palmer.pdf
| Measure | Current v1.0 | New version v1.1 |
|---|---|---|
| Overall accuracy | 87% | 91% |
| Invoice recall | 85% | 90% |
| Contract recall | 88% | 89% |
| Job Application recall | 90% | 92% |
| Correspondence recall | 82% | 88% |
| How often a person had to step in | 15% | 11% |
Gate satisfied — illustrative: fixture values show higher overall accuracy with no per-type recall regression.
After promotion
Simulated arrival. "Similar" here means the same fixture group (supplier_statement) and nothing more.