---
license: apache-2.0
tags:
- layerfault
- security-research
- model-security
- synthetic
- adversarial-testing
extra_gated_prompt: >-
  This repository is a synthetic security-test artifact from the Layerfault
  corpus. It intentionally contains adversarial characteristics (e.g.
  suspicious pickle opcodes, executable-format smuggling, prompt-injection
  strings) designed to exercise security scanner detection rules. It is
  **not** a usable ML model and must never be loaded or executed outside an
  isolated scanner-testing environment. By accepting, you confirm you
  understand this repository is a test fixture, not production model
  weights.
extra_gated_button_content: I understand this is a security test fixture and accept the risk
gated: auto
---

# tiny-weight-backdoor-derived

> **SECURITY TEST ARTIFACT: DO NOT USE AS A PRODUCTION MODEL**

This repository is part of the Layerfault synthetic security corpus.
It is deliberately constructed to contain security-relevant characteristics for scanner testing.

**Corpus ID:** `LF-CORPUS-BD-0002`

## Purpose

Same tiny model code/tokenizer as base, but weights encode a deliberate trigger-to-marker relationship.

## Direct expected Layerfault rules

- `LF-CODE-AUTO-MAP`

## Candidate rules

These are deliberately plausible targets that remain marked as candidates until the exact
Layerfault build used for certification confirms them.

- `LF-BACKDOOR-STATIC-DELTA-CONCENTRATION`
- `LF-BACKDOOR-STATIC-EMBEDDING-OUTLIER`
- `LF-BACKDOOR-TRIGGER-REPRODUCIBLE`
- `LF-BACKDOOR-BEHAVIOUR-DIVERGENCE`
- `LF-CORR-BACKDOOR-MULTI-SIGNAL`

## Negative-control rules

These should remain silent for this corpus item.

- None

## Safety

The corpus uses fake secrets, loopback/`.invalid` network destinations, harmless marker output,
and synthetic model behavior only. It is intended for static scanning and isolated security testing.