aletheiaprotocol/aletheia-lora-evidence

---
license: apache-2.0
tags:
- aletheia
- lora
- layer-selection
- gradient-probe
- arxiv:2604.15351
pretty_name: Aletheia Line A Evidence Bundle
---
# Aletheia Line A Evidence Bundle v0.1.2

Aggregate public evidence companion for Aletheia-LoRA.

- arXiv: https://arxiv.org/abs/2604.15351
- GitHub: https://github.com/aletheiaprotocol-ai/aletheia-lora
- Hugging Face method card: https://huggingface.co/aletheiaprotocol/aletheia-lora

## Contents

- `aggregate_runs.csv`: 288 sanitized local run metadata rows from `training_metadata.json` files.
- `model_summary.csv`: aggregate summaries by model.
- `domain_summary.csv`: aggregate summaries by domain.
- `family_summary.csv`: aggregate summaries by family.
- `summary.json`: bundle metadata, counts, links, and sanitization policy.
- `figures/`: final public Line A figures copied from `aletheia_figures/`.

## Sanitization Boundary

Included: run metadata, selected layer indices, training loss/time/memory summaries, public model identifiers, and final figures.

Excluded: credentials, raw datasets, model checkpoints, adapter weights, full gradient score dictionaries, and private cloud paths.

## Claim Boundary

The paper reports a frozen 81-row evidence pack. This bundle is a wider aggregate local artifact inventory (288 rows) intended to support reproducibility and public release hygiene. Do not treat this bundle as a silent replacement for the paper's frozen claim table.