What keeps pulling me back on OpenGradient is not the bad row.
Its the BATCH_HASHED row under it.
Once $OPG HACA kicks the fast path loose, the inference node answers first. Dashboard bucket goes calm before the proof path has to answer for anything. Then settlement lands BATCH_HASHED and the OpenGradient inference trace quits looking like one row. Starts looking like a settlement bargain somebody made earlier for cost and throughput.
Cheap now. Fine.
Ugly later.
That's the split. Execution fast. Verification lower. Mixed verification on OpenGradient if you want it. TEE. VANILLA. ZKML. Didn't matter. Review learned from the calmer surface first.
I know that trade.
One bad row lands inside that batch and suddenly nobody wants the cheap story anymore. Now they want the exact inference trace? Which call? Which output row? Which path? Which inference trace inside that settlement round? Great. Which batched record actually moved review state. Flipped HOLD. Opened the wrong route. Got copied into the wrong ticket.
By then the row already did the damage.
I've watched that mood change too late.
I keep picturing the same OpenGradient review panel. One ugly row up top. One BATCH_HASHED proof path lower. Dashboard bucket still green. Settlement trace still valid. Then somebody opens the ticket again.
OpenGradient preserved exactly what the settlement round promised.
Fine.
Not missing proof.
Wrong granularity.
Worse timing.
Batch footprint is real. Signed output is real. Lower @OpenGradient settlement trace is real. Just not the clean single-row trace the review panel suddenly wants after the bad row already pushed the wrong HOLD forward.
Cheap now.
Annoying later.
One bad row up top.
One batch footprint lower.
So what did the review panel think it bought there.
One clean row.
Or the OpenGradient settlement bargain under it.