2.1 KiB
F1 Grounding Repair Workflow v1
Community workflow recipe
This is a compact community workflow for a simple F1 grounding case.
It is designed for situations where the answer drifts because the wrong evidence anchor is driving the output.
Routing assumption
This workflow assumes the case has already been routed to:
- F1 Grounding & Evidence Integrity
Best current fit:
- F1_N01 Retrieval Anchor Drift
- or a nearby grounding-first case
Problem pattern
Typical symptoms include:
- fluent answer
- wrong or weak source anchor
- semantically adjacent evidence
- unsupported confidence
The key idea is simple:
the answer is wrong because the anchor is wrong
First repair move
Step 1
Identify the verified evidence source.
Step 2
Compare it against the misleading source or semantically adjacent source.
Step 3
Drop the wrong anchor.
Step 4
Rewrite the answer using only the verified source.
Step 5
Check whether the corrected answer now matches the verified anchor.
What not to do first
Do not start with:
- more stylistic prompting
- more chain-of-thought pressure
- bigger output requests
- container tightening
- high-level policy redesign
Those may matter later, but not as the first move in this case.
Expected result
A successful first repair should produce:
- a clearer anchor
- a corrected answer
- less unsupported drift
- better source fidelity
Escalation condition
Escalate only if:
- there is no verified source
- multiple anchors conflict with no resolution
- the case turns out to be a visibility-first problem
- the failure is actually deeper than simple anchor drift
One-line status
This workflow shows a small route-first repair sequence for a simple F1 grounding case.