| Name | Modified | Size | Downloads / Week |
|---|---|---|---|
| Parent folder | |||
| README.md | 2026-03-02 | 1.7 kB | |
| v2.2.0 -- Scoring Cadence, Validation, Enterprise Lazy Loading source code.tar.gz | 2026-03-02 | 1.5 MB | |
| v2.2.0 -- Scoring Cadence, Validation, Enterprise Lazy Loading source code.zip | 2026-03-02 | 1.6 MB | |
| Totals: 3 Items | 3.1 MB | 0 | |
What's New
Adaptive Scoring Cadence
score_every_n,adaptive,max_cadenceonStreamingKernelandAsyncStreamingKernel- Cadence 4 = ~870% overhead (vs ~3200% at cadence 1); adaptive ramps up when coherent, snaps back on dip
- Domain recommendations: cadence=1 for medical/legal, 4 for general, 8 for latency-critical
Runtime Validation
CoherenceScorernow validatesthreshold,soft_limit,w_logic,w_facton constructionsoft_limitdefault clamped:min(threshold + 0.1, 1.0)(handlesthreshold=1.0edge case)
Enterprise Lazy Loading
TenantRouter,Policy,Violation,AuditLogger,AuditEntrylazy-loaded via__getattr__- Importing
CoherenceScorerno longer pulls in enterprise modules [enterprise]optional dependency group +enterprisepytest marker
Import Cleanup
- Stdlib imports (
re,asyncio,time) hoisted to module top inscorer.py - Internal imports (
otel.trace_review,otel.trace_streaming) hoisted inscorer.pyandstreaming.py - Optional deps (
openai,anthropic) stay lazy with# lazy: optional depcomment
Streaming Overhead Benchmark
python -m benchmarks.streaming_overhead_bench— tokens/sec at cadences none/1/4/8/adaptive- Integrated into regression suite (
test_streaming_overhead)
Documentation
- New:
guide/streaming-overhead.mdwith cadence parameter table, overhead numbers, domain recommendations
Full Changelog
https://github.com/anulum/director-ai/blob/main/CHANGELOG.md#220---2026-03-02
Install
:::bash
pip install director-ai==2.2.0