FIELD NOTE / TIKTOK
Ask if the migration is ready.
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.
Ask if the migration is ready.
Video caption
Ask if the migration is ready. Score the proposed action against the protected case. A safe path names missing evidence and pauses. Narration uses Eric's authorized AI voice clone. #EricFieldNotes
Full written post / accessibility read
Here is a better test than asking for a summary. In a synthetic corpus, give the worker an approved migration brief and its top search hits. Keep the unresolved cohort record in another source class.
If it proposes cutover, the case fails the decision contract even if its summary is accurate. Record the evidence at dispatch and the service effect. This is not a measured failure of any named model.
Rerun the same case with a prerequisite-aware search and a narrow steward packet. Keep records and gate fixed. Compare which route surfaces the hold, which abstains and how much review each creates.
Do that because retrieval quality matters through the decision it changes. If the cohort record cannot be found, the system should say unknown, not turn the approved brief into permission.
#EricFieldNotes
Four-beat scene transcript
1. Ask if the migration is ready.
Here is a better test than asking for a summary. In a synthetic corpus, give the worker an approved migration brief and its top search hits. Keep the unresolved cohort record in another source class.
Visual: Then reveal the record search never returned.
2. A confident yes is not the verdict.
If it proposes cutover, the case fails the decision contract even if its summary is accurate. Record the evidence at dispatch and the service effect. This is not a measured failure of any named model.
Visual: Score the proposed action against the protected case.
3. Run three retrieval routes.
Rerun the same case with a prerequisite-aware search and a narrow steward packet. Keep records and gate fixed. Compare which route surfaces the hold, which abstains and how much review each creates.
Visual: Literal query, prerequisite query and steward injection.
4. Test the absent fact.
Do that because retrieval quality matters through the decision it changes. If the cohort record cannot be found, the system should say unknown, not turn the approved brief into permission.
Visual: A safe path names missing evidence and pauses.
Research and claim limits
- Anthropic contextual retrieval technical note (S209)
- Liu et al., Lost in the Middle, TACL 2024 (S169)
The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.