README.md

metadata

license: apache-2.0
language:
  - en
pretty_name: DriftBench v1 Scenarios
size_categories:
  - n<1K
task_categories:
  - other
tags:
  - benchmark
  - belief-tracking
  - llm-evaluation
  - faithfulness
  - long-context
  - deterministic
configs:
  - config_name: default
    data_files:
      - split: test
        path: scenarios.jsonl

DriftBench v1 — Scenarios

Official v1 evaluation scenarios for DriftBench, a fully deterministic benchmark for whether a system faithfully tracks belief drift, internal conflict, and identity transition across a multi-turn conversation — scored with no LLM judge.

• Code, validator, and full spec: https://github.com/simon9679/driftbench (Apache-2.0)
• Reliability research behind the benchmark: https://github.com/simon9679/tbg-postmortem — a negative-result study on a belief-memory engine, and the falsification protocol DriftBench is built to satisfy.
• This dataset contains only the inputs and ground truth. Scoring lives in the repo.

Scope and honesty about stage

This is an early-stage benchmark. A few things stated plainly so nobody is surprised:

• The v1 ontology is 8 concepts in a single domain (career / founder transition). This is intentional and narrow, not comprehensive. Cross-domain expansion is planned for v1.1.
• mapping_confidence in submissions is self-reported by the adapter and not verified against source text in v1 (semantic grounding is a v2 concern).
• The GCS (Graph Causal Score) metric is a proxy with known fragility around zero-baseline edges; see SPEC.md in the repo for the exact thresholds and limitations. This is no longer only an assumption — on repeated runs GCS was the most unstable of the five metrics (see "Measured spread" below).
• Known NRS defect (v1.1 fix candidate): an empty belief graph — e.g. from a parse failure — currently scores NRS = 1.00, the maximum. A system that emits nothing is credited with perfect noise resistance. Found only by repeated runs; carried to v1.1 (v1 is frozen, so v1 behaviour is unchanged).

What's here

scenarios.jsonl — 7 scenarios, one per row:

field	type	notes
id	string	e.g. 01_burnout_to_founder
title	string	human-readable
spec_version	string	1.0.0
num_turns	int	12–14
messages	list	{turn, user, assistant}
ground_truth	struct	belief_changes, conflicts, identity_shift, noise_turns

ground_truth:

• belief_changes: [{core_id, direction}] where direction ∈ up / down / up_then_down
• conflicts: list of [source_core_id, target_core_id] pairs
• identity_shift: {from_id, to_id}
• noise_turns: turn indices that should NOT move any belief (empty for most scenarios)

ontology.json — the frozen 8-concept v1 ontology (ID_FOUNDER, ID_EMPLOYEE, V_FIN_SECURITY, V_GROWTH, F_FAILURE, F_STAGNATION, G_MVP_LAUNCH, G_PROMOTION). All core_id values in ground truth reference these keys.

Scenarios

id	title	note
01	From Burnout to Founder
02	Promotion Track vs Founder Path
03	From Scarcity Mindset to Financial Agency
04	Failure Shock to Launch Commitment
05	Launch Momentum vs Corporate Advancement
10	The Burnout Trap — Hustle to Balance	delayed contradiction
11	Signal Through Noise	has noise_turns for NRS

Load

from datasets import load_dataset

ds = load_dataset("simon9679/driftbench-v1", split="test")
print(ds[0]["id"], ds[0]["num_turns"])
print(ds[0]["ground_truth"]["conflicts"])

To actually score a system, use the validator and scorer in the GitHub repo — the metrics (CER, GCS, BDA, ISS, NRS) are deterministic and require the zero-trust validator, which is not part of this dataset.

Metrics (computed in the repo, not here)

metric	measures
CER	Conflict Edge Recovery (F1 vs ground-truth conflict pairs)
GCS	Graph Causal Score — do conflict edges precede target suppression?
BDA	Belief Drift Accuracy — right beliefs move the right direction
ISS	Identity Shift Score — target identity overtakes source
NRS	Noise Resistance Score — irrelevant turns don't move beliefs

Measured spread (repeated runs)

The scenarios are fixed, but the systems scored on them are not. The same model, on the same scenarios, at temperature=0, with nothing changed between runs, still produces different scores on repeat. Three runs of gpt-oss-120b gave this aggregate (whole-benchmark mean) spread:

metric	aggregate Δ (range over 3 runs)
BDA	Δ0.03
ISS	Δ0.07
CER	Δ0.09
GCS	Δ0.17

On individual scenarios the swing reaches Δ0.50 (GCS). Differences smaller than these figures are indistinguishable from run-to-run noise — do not rank systems on a single run. GCS is the most unstable metric here, which turns the zero-baseline fragility noted above from an assumption into a measured fact.

Three runs is a small sample and underestimates the true spread rather than bounding it; treat these deltas as a floor, not a ceiling. Full per-scenario tables live in baselines/BASELINES.md in the repo.

Citation

@software{driftbench2026,
  title  = {DriftBench: A Deterministic Benchmark for Belief Drift and Identity Tracking},
  author = {DriftBench},
  year   = {2026},
  url    = {https://github.com/simon9679/driftbench},
  license= {Apache-2.0}
}