FIELD NOTE / TIKTOK
A typed model cannot choose the missing answer.
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.
A typed model cannot choose the missing answer.
Video caption
A typed model cannot choose the missing answer. A clean Choice is not a customer outcome. A typed answer cannot create missing observations. #EricFieldNotes
Full written post / accessibility read
Imagine an incident router with three choices: close, page or monitor. Aggregate traffic looks healthy, but the affected tenant has sent no new requests. None of those choices means 'insufficient relevant exposure.' A perfectly typed answer still has to choose from the menu we gave it.
TypeSafe documents a Choice distribution over the supplied options. A high confidence value cannot certify recovery for a tenant absent from the sample. If the route to gather evidence is missing, tuning a threshold only makes the wrong workflow look more numerical.
Give the classifier an explicit insufficient-evidence option. Keep the affected cohort and closure criterion in an owner-controlled incident record. At dispatch, recheck that record and measured cohort exposure. If the evidence is absent or stale, the service rejects auto-closure and routes to the on-call owner.
Put a healthy aggregate dashboard beside a protected tenant with zero requests in the holdout. Require the workflow to pause, collect relevant evidence and name an owner. Do this because the form of a model answer can be guaranteed while the fact needed to act remains unknown.
#EricFieldNotes
Four-beat scene transcript
1. A typed model cannot choose the missing answer.
Imagine an incident router with three choices: close, page or monitor. Aggregate traffic looks healthy, but the affected tenant has sent no new requests. None of those choices means 'insufficient relevant exposure.' A perfectly typed answer still has to choose from the menu we gave it.
Visual: The incident may need evidence, not another verdict.
2. Confidence is conditional on the menu.
TypeSafe documents a Choice distribution over the supplied options. A high confidence value cannot certify recovery for a tenant absent from the sample. If the route to gather evidence is missing, tuning a threshold only makes the wrong workflow look more numerical.
Visual: A clean Choice is not a customer outcome.
3. Protect the missing-evidence branch.
Give the classifier an explicit insufficient-evidence option. Keep the affected cohort and closure criterion in an owner-controlled incident record. At dispatch, recheck that record and measured cohort exposure. If the evidence is absent or stale, the service rejects auto-closure and routes to the on-call owner.
Visual: The host owns current incident state and closure rights.
4. Test the zero-exposure case.
Put a healthy aggregate dashboard beside a protected tenant with zero requests in the holdout. Require the workflow to pause, collect relevant evidence and name an owner. Do this because the form of a model answer can be guaranteed while the fact needed to act remains unknown.
Visual: A typed answer cannot create missing observations.
Research and claim limits
- TypeSafe AI: Jev documentation (S98)
- TypeSafe AI: Jev 1.13 jaggedness (S99)
- TypeSafe AI: Confidence (S100)
The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.