No evidence, no answer.

I stored tags but left them out of the comparison

In the snapshot collector that became Stratum, tags were stored in a separate column. The content hash and change comparison did not include them. Changing only a resource's environment tag could therefore look like no change at all.

An assistant reading that history would not have the missing change to retrieve. Better answer wording could not fix the underlying record.

One comparison input for the hash and the diff

I introduced one normalized representation containing both the resource attributes and tags. Both the content hash and the diff now use it.

The regression test ingests a resource tagged env=dev, then the same resource tagged env=prod. It expects one UPDATED change with the old and new tag values. Another test checks that a tag change affects the hash.

Same resource attributes
Before: env=dev
After:  env=prod
Expected: one UPDATED change, including the tag

What a source link does and does not prove

Stratum's read-only assistant retrieves records and returns source links. Those links let a reader check the underlying record; they do not guarantee that every model claim is correct.

The portfolio lab uses fixed answers, including a question the sample workspace cannot answer. It does not call a model or run Stratum's retrieval tools.

The beta is still limited

The live beta requires an invited account and has shared AI quotas. Its sample workspaces are not customer tenants. A private workspace without collected data stays empty instead of falling back to sample records.

The optional private Azure pilot has not been verified against a real subscription. I am not presenting the beta as evidence of customer adoption or a completed production rollout.