Back to DevLog

When a verified fact in memstack was wrong and kept reciting

3 min read

A fact marked verified turned out to be wrong on three counts. It named a mechanism that could not have produced the damage it described. Its date was off by two days. Its file count did not match the cleanup that actually ran.

Worse, it was incoherent in its own text. It named a cause and an observation that contradict each other inside a single sentence. That contradiction passed through ingestion, storage, ranking, and recitation without anything catching it.

Why the label survived

The word verified auto-assigns a confidence value at write time. Confidence is the only route by which the label reaches ranking. The scoring function reads confidence, age, recall count and status, and never reads the method at all. A session grants its own claim a permanent retention advantage by typing a word about itself, and nothing downstream questions it.

The majority of live facts in the store are diary-sourced and self-labelled verified. What gets recited is largely whichever record was saved most recently, which is not a property anyone designed.

The fix

I qualified diary facts at rendering time. A pure transform over the stored source and method pair. No stored value changed, no fourth label entered the vocabulary, and exactly one of the eight combinations that occur in a real store renders differently. The other seven are byte-identical and a parametrized test pins that.

The scope is drawn on deliberation, not authorship. A template field filled for every row of a block at end of session, after the evidence has left the writer's context, is a different act from one deliberate call about one fact made while the thing is still in view. The code says so explicitly rather than leaving a future reader to infer it from how the values happen to be distributed today.

Re-filing corrections

A correction that disproved a mechanism could not previously change the subject, so the claim was right and the filing was a lie. Anything reading subjects rather than claims kept reading the disproved version. I let corrections change the subject.

I also added forward resolution for stale subjects. When a direct match is empty, the fallback resolves forward through the supersession chain. That silence is the same failure class already recorded in this store: wrong input, plausible empty result, no error.

Why not fix confidence at query time

A read-time discount would write nothing to customer data, but the ceiling is a claim about the fact rather than about ranking. If the number in the column is wrong, the number should be fixed, not worked around at every read.

Forward-only would not work either. Each recall extends a fact's retention, so the mis-weighted facts that have actually been recited would keep their unearned weight longest. That is precisely the population that does the damage.

Next

Phase two is a confidence ceiling applied at write time. The ceiling caps what a self-declaration can buy, so hand-set values pass through untouched. A flat remap would collapse verified into reported and destroy the one part of the vocabulary that demonstrably works.

The migration is a single predicated update over rows that are diary-sourced, verified, and carrying the auto-assigned value. Single because the memory database has writers outside the current session. Another project's diary ingest added 26 facts during one measurement window, 24 of them as diary and verified in one batch. The pathology reproduced itself in real time during the measurement.

Share this post