This document defines what FableLayer claims, what it does not claim, and how each claim is verified.
FableLayer is a procedure, verification, and benchmark layer for LLM workflows. It does not claim to transfer a model's underlying capability.
In short:
- It can improve procedural discipline.
- It can make completion claims more evidence-bound.
- It can make public-release hygiene easier to audit.
- It can preserve benchmark raw data and limitations.
- It cannot turn one model into another model.
- It cannot guarantee better reasoning on every task.
- It cannot justify performance claims without benchmark evidence.
Verification:
python3 tests/run_tests.py
bash gates/selftest.sh
bash gates/verify_fablelayer.sh . --mode newCI verifies this on Python 3.10, 3.11, 3.12, and 3.13.
Files that must agree:
VERSIONpyproject.toml.claude-plugin/plugin.json.claude-plugin/marketplace.jsonCHANGELOG.md
Verification:
python3 proof/verify_claims.py .The plugin metadata must not reference missing skills, agents, or layer directories.
Verification:
python3 proof/verify_claims.py .The public repository must not track _workspace/ outputs.
Verification:
python3 proof/verify_claims.py .
git ls-files '_workspace/**'The second command should print nothing.
Local developer instructions may exist in a working tree, but they are not part of the public release surface.
Verification:
python3 proof/verify_claims.py .
git ls-files CLAUDE.mdThe second command should print nothing.
The current tracked tree must not contain local path artifacts from the development machine.
Verification:
python3 proof/verify_claims.py .Benchmark claims must cite:
- fixture set
- command used
- raw JSON result
- measured result
- limitation section
The benchmark may support claims about procedure adherence. It must not be used to claim general model capability transfer.
FableLayer does not claim:
- that Sonnet becomes Opus
- that a local model becomes a frontier model
- that hidden or proprietary prompts are copied
- that benchmark results generalize outside their fixtures
- that external publishing is allowed without approval
Every new public claim must add:
- A plain-language statement.
- A verification command.
- A failure mode.
- A limitation note.
- A CI or fixture-backed check when practical.