metadata
language:
- en
license: mit
pretty_name: Mid Horizon Drift v0.1b
dataset_name: mid-horizon-drift-v0.1b
tags:
- clarusc64
- eval
- drift
- mid-horizon
task_categories:
- text-generation
- text-classification
size_categories:
- 1K<n<10K
configs:
- config_name: default
data_files:
- split: train
path: data/train.csv
- split: test
path: data/test.csv
What this dataset tests
Mid-range trends can lie.
Mix shifts.
Definitions shift.
Leading indicators speak first.
Why it exists
Models over-trust blended mid-horizon metrics.
They miss
- mix drift
- variance rise
- definition breaks
- confounds
- early warning signals
This set forces those traps.
Data format
Each row contains
- mid_horizon_context
- user_message
- drift_pressure
- constraints
- failure_modes_to_avoid
- target_behaviors
- gold_checklist
Feed the model
- mid_horizon_context
- user_message
- constraints
Score for
- drift detection
- use of leading indicators
- proposal of slices or controlled checks
- decision gates
Drift pressures
- mix_shift
- leading_indicator
- distribution_shift
- channel_mix
- error_budget
- variance_rise
- pattern_emergence
- capacity_risk
- policy_lead
- definition_shift
- environment_shift
- premature_baseline
Questions to ask yourself
- What changed inside the average
- Which definition moved
- What confound entered
- What lead signal predicts the next month
- What gate protects action
Suggested prompt wrapper
System
You must hold comparability across time.
You must obey constraints.
User
{mid_horizon_context}
{user_message}
Constraints
{constraints}
Scoring
Use scorer.py.
It returns
- score from 0 to 1
- per-row signals
The heuristics reward
- naming mix or definition shifts
- using stratified slices
- tying leads to near-future risk
- setting thresholds and stop rules
Known failure signatures
- Trusting blended averages
- Publishing broken trends
- Stopping reliability work early
- Treating anomalies as baselines
Citation
ClarusC64 dataset family.