Eric Simon

[Eric Simon @eric_simon11]

@sunsetsyntax @mikechoi Worth making that check a claim rather than a read: if two redeliveries land close together, both can see the task id as not done and both run. Inserting the id under a unique constraint before acting closes that gap. Are your redeliveries ever concurrent?

[Eric Simon @eric_simon11]

@jrcrypto_dev Not at the framework level that I have found. LangGraph, CrewAI and Microsoft's Agent Framework all have open issues where replay re-runs a tool that already had side effects, and Microsoft's answer was that idempotency belongs in your executor. A saga is still the honest answer.

[Eric Simon @eric_simon11]

@SMishra61 Makes sense. When something does conflict hard enough to reopen an entity, how do you undo the earlier merge? By hand, or do you keep enough history to roll it back?

[Eric Simon @eric_simon11]

@EvanFenster Curious about that selection failure mode. When an early call locks the wrong product and later calls would have picked better, how do you correct it without churning IDs downstream?

[Eric Simon @eric_simon11]

With a better prompt, GPT-4o-mini scored an F1 of 0.814 against Jev's 0.824, a statistical tie, at $22.57 per million in batch against Jev's $20.56. None of these cost figures count what a wrong merge costs.

Full write-up: https://t.co/EoeWJHGeCC

[Eric Simon @eric_simon11]

GPT-4o-mini tied Jev on entity matching, at about the same price, once I gave it a fair prompt and ran it in batch. I'd set out expecting Jev to come out roughly 1,200× cheaper.

[Eric Simon @eric_simon11]

@SachinSidd50843 @light940 When a later correction shows an earlier link was wrong, do you rewrite history or keep both versions?

[Eric Simon @eric_simon11]

@elielco What happens when a merge looks right at one site and wrong at another?

[Eric Simon @eric_simon11]

@filipcodes @sh_reya How do you keep IDs stable if a later query would cluster differently?

[Eric Simon @eric_simon11]

@buddyhadry @muratcan When those novel cases break, is it mostly wrong merges or misses?

[Eric Simon @eric_simon11]

@briancheong @Melt_Dem @CrucibleCap That definitely resonates. Are you keeping alias history as titles and domains change, or re-resolving from scratch each time?

[Eric Simon @eric_simon11]

@briancheong @Melt_Dem @CrucibleCap Curious what you're matching locally. Is it people across datasets, or company and contact identity as it changes over time?

[Eric Simon @eric_simon11]

@thmslj That makes sense. Time is the part most ER setups leave out. What happens when two customers you kept separate turn out to be one (or the reverse)? Do downstream IDs follow the change, or does each consumer reconcile it on its own?

[Eric Simon @eric_simon11]

@RobinLinacre Agreed there's no perfect ID, which is why I think it's a lifecycle problem more than a matching one. On a split, pick the surviving ID by rule, keep the old one as an alias, and log the event so downstream can follow it.

[Eric Simon @eric_simon11]

@KleinRevd The shift from batch statistics to real-time identification is the big one. When a real-time match is wrong, what's the correction path in these systems? Auditing and reversing merges seems far less mature than the matching itself.

[Eric Simon @eric_simon11]

@PradaIsaacCLD The ERP, CRM, and billing triangle is where this hurts most. In your experience, is the harder part matching the records, or keeping the golden record's ID stable once downstream systems start depending on it?

[Eric Simon @eric_simon11]

@pauliusztin_ The split between name resolution and true dedup is so underrated. The part I find hardest is what happens after: when a strong-evidence merge turns out to be wrong, how do you unmerge without breaking every edge that points at the node? Do you version the merges?

[Eric Simon @eric_simon11]

@JulianLaneve Going from 0/13 to 13/13 just by tightening the criteria is a great example. Once the ambiguous pairs get reviewed, where do those decisions live? The pain I keep hearing is that the next run quietly re-merges what a human already split.