Experiment

Schema Validation Reliability for AI-Drafted Content and Automation Recovery

Test how reliably the blueprint's automated checks catch mistakes in AI-drafted content, and how reliably its scheduled safety net recovers automation steps that a missed event leaves undone.

DESIGNED · NOT RUN YET

Hypothesis

Deterministic checks (required fields, diagram rules, and every number traced back to the author's own notes) will catch nearly all seeded factual and structural errors before a human sees the draft, and a scheduled sweep will recover every missed automation within one cycle.

Problem

The blueprint lets AI draft content, but only a person publishes. That rule only holds up if the checks in front of the reviewer actually catch errors, and if automations that depend on events don't silently skip work when an event is missed.

Setup

Validation test: take a set of correct drafts and seed known errors (missing required fields, invalid diagram placements, invented numbers not in the source notes, contradictory figures). Run the checks and record which errors they catch. Recovery test: deliberately drop a share of publish events so their follow-up steps (refresh, asset generation) don't run, then let the scheduled sweep run and record how many are recovered and how long it takes. Metrics: error catch rate by error type, false alarms on clean drafts, recovery rate and time to recovery.

PROCESS FLOW

The seeded-error test

  1. 1. Input

    Clean drafts

  2. 2. Action

    Seed known errors

    Missing fields, invented numbers

  3. 3. Decision

    Run content checks

  4. 4. Output

    Caught vs missed, by type

Known errors are planted in clean drafts, the checks run, and every planted error is marked caught or missed.

Input

Seeded drafts with a known error list; a controlled run of publish events with a known set of dropped ones.

Output

Catch rate by error type and a recovery timeline.

Decision

Status: designed, not run. The checks and the sweep are in use; their reliability has not been measured.

Next Step

Prepare the seeded drafts and the event-drop harness, run both tests, and publish the results here.

Use Cases

  • ▸Human-in-the-loop AI content pipelines
  • ▸Webhook-driven automations that need a safety net