"Dwell regret" scoring gate removed; Under the Hood shows legal withholding
- Removed scorer
value_model_gate(537 lines). It was a per-viewer linear gate: it scored features from the viewer's recent action sequence and compared the score to a threshold to decide whetherranking_scorershould use the alternative "dwell regret" scoring path. Inside the hysteresis band, a hash of the user ID made the call. With it go 18 params: everyDwellRegret*alpha/negative/gate/temperature param andValueModelMode.ranking_scorer.rsloses 465 lines and gains 12. - New param
PhoenixColdStartMaxResults(default0, i.e. off).PhoenixRetrievalMOEInferenceClusterIdchanges from"Experiment2Memy04"to"Experiment3Memy04". - Under the Hood: the README's "September 18th, 2026" note says reports now show whether an account's or post's visibility was limited for legal compliance (e.g. withheld in a specific country). That matches the changes to
UthPctdAccountTakedownEventsJob.scalaandunderTheHoodLabels.strato. - Visibility filtering gets a new
tes_compositehydration path and a batchevaluate_tweetsAPI.
Full analysis → Summary: hand-written notes, checked against the diff