FIELD NOTE / TIKTOK
The board packet went through three AIs.
The complete written thought and the evidence behind it. The video edition will follow its public release.
The written argument is here.
This approved TikTok edition is on the journal now. Its video player and original platform link will appear after each public release is verified.
The board packet went through three AIs.
Video caption
The board packet went through three AIs. The team executes a generalized approval. Do this because a fast packet can create slow rework. #EricFieldNotes
Full written post / accessibility read
Fictional scenario: one assistant drafts the board packet, another summarizes it for executives, and a third turns the approval into tasks. The customer-specific exception is in the source but absent from the final task list.
The source said pilot only for one tenant pending review. The executive summary said pilot approved. The implementation ticket lost both tenant and condition. No model needed to fabricate a bizarre fact; each compression discarded context.
Before the tasks are accepted, compare every approval condition with the original record. Highlight missing customer, time limit or precondition and ask the owner to resolve it. A task should not inherit authority from the shortest summary.
Count elapsed time through implementation, customer acceptance and reversal, not just packet drafting. The company wins when a decision survives contact with the source and the customer, not when documents arrive quickly.
#EricFieldNotes
Four-beat scene transcript
1. The board packet went through three AIs.
Fictional scenario: one assistant drafts the board packet, another summarizes it for executives, and a third turns the approval into tasks. The customer-specific exception is in the source but absent from the final task list.
Visual: The customer exception did not make it through.
2. Everything sounds efficient until launch.
The source said pilot only for one tenant pending review. The executive summary said pilot approved. The implementation ticket lost both tenant and condition. No model needed to fabricate a bizarre fact; each compression discarded context.
Visual: The team executes a generalized approval.
3. Put the exception back in the decision path.
Before the tasks are accepted, compare every approval condition with the original record. Highlight missing customer, time limit or precondition and ask the owner to resolve it. A task should not inherit authority from the shortest summary.
Visual: Require a source-to-task diff before coding.
4. Time the complete decision.
Count elapsed time through implementation, customer acceptance and reversal, not just packet drafting. The company wins when a decision survives contact with the source and the customer, not when documents arrive quickly.
Visual: Do this because a fast packet can create slow rework.
Research and claim limits
- Mohamed et al., LLM as a Broken Telephone, ACL 2025 (S167)
- Perez et al., When LLMs Play the Telephone Game, ICLR 2025 (S168)
- NIST AI RMF 1.0 (S171)
The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.