Bernstein 3.x recap: deterministic replay, evidence bundles, tournament runs

Bernstein's 3.x line takes the capabilities added in 2.x — trackers, the web UI, a cost economy, multi-agent runs — and gives each one a verifiable record. 3.0 collapsed lineage into a single Ed25519-signed chain and turned on deterministic replay by default: two runs from the same seed and state are now compared by hash, so drift shows up as an exact-position divergence rather than a flaky retest.

3.1's verification evidence bundles turned "done" into a recomputable artefact, with test, coverage and lint outputs stored in a content-addressed store and sealed into the audit chain. The durable work ledger and detached run service introduced across 3.2-3.4 let an interrupted run resume without redoing finished work, while tournament runs and cost-aware scheduling turned scheduler decisions into model-free, signed receipts that can be recomputed offline.

For engineers, the practical shift is that operational trust no longer rests on the orchestrator's say-so but on evidence anyone can recheck. The web dashboard caught up too, gaining scoped access tokens and a deterministic run-review board built directly from the journal.