In the case of tx-explain hallucinating decimals and/or having other inaccuracies in the response in spite of having all the data available in asset_changes json object, the currently accepted method of solving this issue is by switching from Haiku to a stronger model.
During testing, this was done manually. The challenge is to do this automatically. This would require an automated evaluation of the summary provided by tx-explain.
Suggested solution:
We could create a heuristic that would take the amounts present in the summary and compare them with the amounts in asset_changes json object. The amounts should be 1) identical and 2) all should be present.
In the case of tx-explain hallucinating decimals and/or having other inaccuracies in the response in spite of having all the data available in asset_changes json object, the currently accepted method of solving this issue is by switching from Haiku to a stronger model.
During testing, this was done manually. The challenge is to do this automatically. This would require an automated evaluation of the summary provided by tx-explain.
Suggested solution:
We could create a heuristic that would take the amounts present in the summary and compare them with the amounts in asset_changes json object. The amounts should be 1) identical and 2) all should be present.