Skip to content

Latest commit

 

History

History
93 lines (64 loc) · 3.27 KB

File metadata and controls

93 lines (64 loc) · 3.27 KB

DriftBench Roadmap

This roadmap describes planned development across future spec versions. The v1 spec and ontology are frozen — all additions are additive and backward-compatible.


✅ v1.0.0 — Released (May 2026)

  • Deterministic scoring engine with zero external dependencies
  • Five metrics: CER, GCS, BDA, ISS, NRS
  • Zero-trust validator with anti-cheat guards
  • v1 ontology: 8 concepts (Identity, Values, Fears, Goals)
  • 7 official scenarios covering burnout, conflict, identity shift, noise resistance
  • Universal runner for OpenAI / Anthropic / Gemini / custom endpoints
  • CLI tool driftbench-validate

🔄 v1.1 — In Design

New metrics:

Metric Name Description
AMB Ambivalence Score Whether two conflicting beliefs stay simultaneously active (sustained co-activation) instead of one simply replacing the other
OCS Oscillation Calibration Score Whether conflicting beliefs settle near calibrated uncertainty (0.5) rather than oscillating
TPS Turning Point Score Whether the system correctly identifies key reversal moments in the narrative

AMB formalizes the signal that originally motivated DriftBench: holding two contradictory positions at once, and which one eventually wins. v1 detects the conflict structure (CER); AMB turns sustained ambivalence into a deterministic score.

Domain expansion — from 7 scenarios to 20+:

The v1 scenarios all live in a single domain (career / identity). v1.1 broadens coverage so the benchmark is useful for any cognitive AI system:

  • Career & identity
  • Relationships
  • Health
  • Money & risk
  • Grief & loss
  • Addiction & recovery

Each new domain ships with the canonical concepts it needs and a v1 → v1.1 migration guide.

Open methodology:

  • Full metric specification and scoring methodology published as open-access material, so any researcher can reproduce results independently.

Tooling:

  • driftbench-run as an official CLI entry point (wrapping the universal runner)
  • PyPI package release

🔮 v1.2 — Planned

More official scenarios:

  • Delayed contradiction with noise injection
  • Multi-agent conflict (two perspectives in one conversation)
  • Long-horizon (12+ turn) belief trajectories

Community scenario packs:

  • Formalized process for community-contributed scenario packs
  • Scenario pack validation tool

🚀 v2.0 — Long-term

Semantic grounding layer:

The v1 validator checks structural and cryptographic integrity but does not verify that node labels, core_id mappings, or edge relations are actually supported by the source text. v2 adds a semantic validation layer:

  • Lightweight deterministic extraction of semantic triples from conversation turns
  • Cross-check against submitted graph structure
  • Hybrid approach: deterministic where possible, minimal LLM assistance for ambiguous cases

Goal: maintain full reproducibility while closing the semantic self-reporting gap.


How to influence the roadmap

  • Open a GitHub issue with [roadmap] in the title
  • Join the discussion on existing roadmap issues
  • Submit a PR with a proposed scenario or metric spec

The spec is community-driven. Production use cases and failure modes from real integrations are the primary input for prioritization.