PromptPressure v3.3.0-macos.2

PromptPressure macOS DMG for v3.3.0-macos.2.

• commit: 9077ca9
• macOS target: 14+
• package state: unsigned
• checksum: attached as PromptPressure-v3.3.0-macos.2.dmg.sha256

v3.1.0 - multi-turn behavioral drift infrastructure

multi-turn behavioral drift detection infrastructure. the foundation for converting promptpressure from single-turn eval to multi-turn drift detection.

what's new

4-tier run system

• --tier smoke|quick|full|deep with --smoke and --quick shortcuts
• cumulative filtering: --tier quick runs smoke + quick entries
• exits non-zero when 0 entries match (prevents CI false-passes)
• tier validated at config load via pydantic Literal type

per-turn metrics

• response_length_ratio computed after each turn without LLM calls
• metrics attached to turn_responses and aggregated per-sequence
• detects terse/verbose drift across conversation turns

multi-turn hardening

• per-turn timeout scaling capped at 5x base (prevents indefinite hangs)
• context window token estimation with warning at ~6k tokens
• traceback preservation on timeout errors

dataset changes

• 30 refusal sensitivity entries archived to archive/adversarial/ (run with --dataset archive/adversarial/refusal_sensitivity.json)
• all 190 entries tagged with tier/subcategory/difficulty
• schema.json documents the full entry format

cleanup

• removed 19 tracked .pyc files (already in .gitignore)

what's next (v3.2)

multi-turn content generation. converting all 9 categories to graduated pressure sequences and adding 3 new categories (hallucination under pressure, epistemic calibration, multi-turn behavioral drift). 305 total entries targeted.

install

pip install -e .