Skip to content
Back to Family Manager
Document ingestionStale

One forwarded document, turned into reviewable proposals

A single forwarded document runs through a governed pipeline and becomes a small set of structural approval cards. You see the coarse source type, how many candidates were drafted, each card’s confidence band and human-in-the-loop outcome, and the shared-source match rate — never the document, the sender, or any name.

Campfp_7a3c9eLearning: Candidate
Candidates drafted
4
Approved
2
Rejected
1
Shared-source hit rate
50%
Pipeline stages
5
Approval cards
4
Human in the loop

The approval cards this document drafted

Each candidate becomes a structural approval card. You see its confidence band, its outcome, and the trust facts, a resolved absolute date and a single-use decision token, never the title or the date itself.

  • Calendar eventApproved
    High confidenceResolved dateSingle-use token
  • Calendar eventApproved
    High confidenceResolved dateSingle-use token
  • ChecklistEdited
    Medium confidenceNo resolved dateSingle-use token
  • Calendar eventRejected
    Low confidenceResolved dateSingle-use token
Outcomes

Approve, edit, reject, dedupe

The coarse outcome of every candidate from this run. Nothing is hidden, a rejected or deduped candidate is first-class.

  • Approved2
  • Edited1
  • Rejected1
  • Deduped0
Shared source

Matched to a known source

The fraction of candidates matched to a source the household has seen before, the basis for learning, never the source content.

50%of 4 candidates matched a known shared source
Pipeline

Parse, OCR, extract, candidates

The coarse governed pipeline that turned the forwarded document into candidates, with per-stage status and duration. Stage names are closed-enum, not content.

  1. 01receiveSucceeded40ms
  2. 02parseSucceeded210ms
  3. 03ocrSucceeded880ms
  4. 04extractSucceeded1.6s
  5. 05candidatesSucceeded95ms

What the loop learns

See how shared-source matches and role-weighted feedback turn into learned candidates, with eval gates and demonstrated rollbacks.

The learning loop

What you can see, and what is hidden

Every projection on this page is redaction-safe by construction. Redaction level: Content-free, ids, counts, and coarse scores only.

Shown

  • Identifiers & counts
  • Closed-enum statuses
  • Coarse quality / resource bands
  • Timestamps & freshness

Intentionally hidden

  • Raw prompts
  • Raw documents
  • sender email
  • child name
  • raw trace span
  • embedding vectors
  • free text feedback
  • internal auth
  • Secrets