Bernstein 2.16: The Output-Economy Suite and Safer Agent Summaries
Bernstein 2.16 adds per-role cost attribution, a cost-quality A/B harness, safer proactive context compaction, and schema-enforced completions.
Bernstein's 2.16 release targets a problem that only shows up once teams run dozens of agents rather than a handful: spend that arrives as one opaque number, summaries that can't be trusted, and completion payloads that are hard to parse reliably at scale.
The new 'output-economy' suite adds four pieces: per-role response-style profiles so verbose and terse roles no longer pay the same token cost, cost-ledger attribution by profile so spend breaks down by verbosity tier, a three-arm A/B harness that runs the same task across cheap, balanced, and thorough model configurations and reports cost against a quality gate, and an operator-gated compression tool that only rewrites role templates after a human approves the diff.
On the safety side, context compaction is now proactive rather than reactive, validated against task invariants, and logged with a signed receipt. A new gate scans any summary sent to a model for credential-shaped content and redacts matches before they leave. Worker completion payloads are now schema-enforced, with refusals returned as a typed, unambiguous outcome instead of free-form prose.