Skip to content
Research Insights Made Simple
Software factories · series 2/6Approach review

Dark factories: independent acceptance

Who checks the code nobody reads

/ Research Insights Made Simple · Software factories 2/6 · Independent acceptance

Slide contents

  1. 1. Dark factories: independent acceptance

    Who checks the code nobody reads

  2. 2. A verdict needs independent evidence

    Contract, evaluator, test environment and future change

  3. 3. People define acceptable behavior

    In the strict model, routine implementation review disappears

  4. 4. Self-checks can share a blind spot

    One wrong assumption can enter both implementation and test

  5. 5. Workers cannot rewrite their final exam

    Final scenarios live outside the worker’s write permissions

  6. 6. Hide probes; keep requirements public

    Expected behavior is available before the run

  7. 7. Check exact outcomes exactly

    Deterministic checks, model judges and experts answer different questions

  8. 8. Satisfaction concerns observed trajectories

    Likely satisfactory outcomes depend on the scenario set

  9. 9. A twin makes failures reproducible

    A behavioral API model provides controllable states

  10. 10. The simulator also creates blind spots

    A useful external-service model has testable boundaries

  11. 11. The factory depends on its evaluation system

    Specification, agent, scenarios and environment form a loop

  12. 12. Attractor publishes implementation specifications

    NLSpecs: orchestrator, agent loop and unified client

  13. 13. The validator returns defects to workers

    Factory Missions separates execution, validation and orchestration

  14. 14. Validation consumes substantial work

    Published Slack-like project: 16.5 hours in total

  15. 15. A repeated request exposes an extra effect

    T1: one identifier, repeated delivery, one promised effect

  16. 16. Reports return evidence, not a ready fix

    Observation, expectation, revision and reproduction guide repair

  17. 17. Current correctness does not guarantee evolvability

    Read the lights-off account with its date and context

  18. 18. The next task tests inherited code

    T4 starts from that system’s accepted T2 result

  19. 19. A coverage map beats one green status

    Automatic checks, expert judgment and untested properties

  20. 20. Trust grows around verification boundaries

    Contract, independent verdict, controlled environment and the next change