Skip to content

PVU-BS Pre-Registration Template (English)

Fill this document and commit it to the repository before breaking the test seal.
Spanish version: preregistration_template_ES.md

1. Identification

Field Value
Run ID <!-- e.g. pvu_run_001 -->
Date pre-registered <!-- YYYY-MM-DD -->
Registered by <!-- GitHub username -->
Git commit (codebase) <!-- SHA at registration time -->

2. Hypothesis

Copy from PVU-BS § 4.1 or state a more specific variant:

MASSIVE produces significantly lower MAE on the held-out test set compared to the naive baseline, after Holm–Bonferroni correction (α = 0.05).

Primary metric: <!-- MAE / RMSE / TPS F1 / … -->

3. Data

Field Value
Cases folder datasets/pvu_cases/
N cases <!-- total number -->
Cluster IDs <!-- list or "none" -->
Date range <!-- start – end -->
Source <!-- synthetic / Reddit / … -->
License <!-- CC0 / CC-BY / … -->

Independence verification:
Explain how cases satisfy the independence criterion (§ 2.1).

4. Model Configuration (frozen)

Parameter Value
configs/pvu.yaml SHA <!-- git hash of config file -->
Seed <!-- integer -->
PYTHONHASHSEED <!-- integer -->
Mode <!-- offline / llm -->
LLM provider + model <!-- if llm mode; else "n/a" -->
Temperature <!-- if llm mode; else "n/a" -->
Python version <!-- e.g. 3.11.x -->
Key package versions <!-- numpy X.Y, scipy X.Y, … -->

5. Analysis Plan

  • Train/test split ratio: <!-- e.g. 70/30 --> (set in configs/pvu.yaml)
  • Statistical test: Diebold–Mariano, two-sided, squared-error loss
  • Multiple-comparison correction: Holm–Bonferroni across all baseline × case comparisons
  • Effect sizes to report: ΔMAE, ΔRMSE, directional accuracy lift, TPS F1
  • Turning-point detection: order=__, min_prominence=__, tolerance=__

Exclusion criteria (pre-specified; no cases may be excluded after looking at test metrics):

  • e.g. cases with fewer than 10 test observations are automatically skipped by the runner

6. Anti-Leakage Declaration

I confirm that at the time of pre-registration:

  • [ ] I have not looked at test-split metrics or plots.
  • [ ] Model parameters and prompts are frozen (see SHA above).
  • [ ] No cases were selected or excluded based on expected performance.
  • [ ] The runner will be executed once; results will be reported as-is.

7. Deviations Log

Fill after the run if anything deviated from the pre-registration:

Deviation Reason Impact
(none) — —

This template follows PVU-BS v1.0 — see PVU_BeyondSight_EN.md