FIELD NOTE / LINKEDIN
A steward is a hypothesis.
The complete written thought and the evidence behind it. The video edition will follow its public release.
The written argument is here.
This approved LinkedIn edition is on the journal now. Its video player and original platform link will appear after each public release is verified.
A steward is a hypothesis.
Video caption
A steward is a hypothesis.
Misses, false holds and people time belong together.
Move facts, revoke holds, suppress the steward.
My rule: Complexity needs evidence of net value.
#EricFieldNotes
Full written post / accessibility read
I would compare four complete workflows on the same owner-labeled migration cases: baseline search, long context, source-aware retrieval and a domain steward. Every route gets the same service rule and target-state oracle.
Measure critical-fact recall at dispatch, wrong cutovers, unnecessary holds, reviewer minutes, latency, maintenance burden and recovery cost. Keep expensive exceptions visible rather than letting a routine-case average hide them.
A protected set should include approval-only, open and resolved cohorts, contradictory sources and absent records. Rerun after a policy update and with the supervisor offline. Require service and target readback.
Do that because a second agent creates a new update stream and review queue. If source-aware retrieval plus a deterministic gate passes the contract, use it; add a steward only for a measured gap.
#EricFieldNotes
Four-beat scene transcript
1. A steward is a hypothesis.
I would compare four complete workflows on the same owner-labeled migration cases: baseline search, long context, source-aware retrieval and a domain steward. Every route gets the same service rule and target-state oracle.
Visual: It earns its place against simpler retrieval.
2. Score accepted work.
Measure critical-fact recall at dispatch, wrong cutovers, unnecessary holds, reviewer minutes, latency, maintenance burden and recovery cost. Keep expensive exceptions visible rather than letting a routine-case average hide them.
Visual: Misses, false holds and people time belong together.
3. Test perturbations, not demos.
A protected set should include approval-only, open and resolved cohorts, contradictory sources and absent records. Rerun after a policy update and with the supervisor offline. Require service and target readback.
Visual: Move facts, revoke holds, suppress the steward.
4. Choose the smallest passing system.
Do that because a second agent creates a new update stream and review queue. If source-aware retrieval plus a deterministic gate passes the contract, use it; add a steward only for a measured gap.
Visual: Complexity needs evidence of net value.
Research and claim limits
- Liu et al., Lost in the Middle, TACL 2024 (S169)
- LongMemEval, ICLR 2025 (S206)
- Anthropic contextual retrieval technical note (S209)
- Microsoft Research GraphRAG paper (S210)
The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.