Agents, with a checkpoint

Data

Every decision traces back to what the agent saw.

Data moves one way, in clear stages, and each agent run reads a frozen copy.

  1. RawAs it arrived
  2. CleanChecked, joined
  3. FeaturesPer customer or case
  4. SnapshotFrozen, with a hash
  5. AgentsRead only

Raw records are never edited. Cleaning and joining happen in a separate step, so a bad input can be traced. Before a run, the inputs are frozen into a snapshot with a fingerprint (a content hash).

That means a month later anyone can ask "why was this customer flagged?" and replay the exact inputs. Agents read; they don't write back into the source data.

Demo data
Synthetic, with an answer key
Customer details
Only the fields a step needs
Write-back
None

↑ ↓ to move · ↵ to open · esc to close