FIELD NOTE / X
A demo can pass with no state model.
The short film, the complete written thought, and the evidence behind it.
The X edition will be linked here after its public post is verified.
A demo can pass with no state model.
Video caption
A demo can pass with no state model. A job resumes from a checkpoint after a partial write. Require an external probe for the consequential branch. #EricFieldNotes
Full written post / accessible read
A working demo is useful, but it can hide that nobody specified the legal state transitions. Two models agreeing on the UI is still one evidence stream if neither has the production policy, failure history or independent state readback.
Take an illustrative backfill. The visible result is correct on the happy path. A worker writes half a batch, loses its checkpoint acknowledgment, and replays. Whether that is safe depends on the write key and recovery state, which the demo never exercised.
Before asking another model to approve the architecture, write the transition table: start, partial write, checkpoint, retry and reconciliation. Mark which effects can repeat and which component owns a stable operation key. Then execute a failure at the seam.
Use the model loop to find missing transitions, then make a test harness kill each dangerous assumption. Record the final database state, not just the agent's explanation. Do this because agreement narrows hypotheses; it does not observe the side effect.
#EricFieldNotes
X thread draft
Video belongs on the first post; the rest gives the full argument. Check the native composer before sending.
A demo can pass with no state model. A job resumes from a checkpoint after a partial write. Require an external probe for the consequential branch. #EricFieldNotes
A working demo is useful, but it can hide that nobody specified the legal state transitions. Two models agreeing on the UI is still one evidence stream if neither has the production policy, failure history or independent state readback.
Take an illustrative backfill. The visible result is correct on the happy path. A worker writes half a batch, loses its checkpoint acknowledgment, and replays. Whether that is safe depends on the write key and recovery state, which the demo never exercised.
Before asking another model to approve the architecture, write the transition table: start, partial write, checkpoint, retry and reconciliation. Mark which effects can repeat and which component owns a stable operation key. Then execute a failure at the seam.
Use the model loop to find missing transitions, then make a test harness kill each dangerous assumption. Record the final database state, not just the agent's explanation. Do this because agreement narrows hypotheses; it does not observe the side effect.
Evidence and boundary
On-screen label: ILLUSTRATIVE DESIGN REVIEW. Illustrative cases are not measured incidents. Research papers and vendor documents support the stated mechanism only within their studied or documented scope.