fablelayer

FableLayer

FableLayer is a public-safe procedure, verification, and benchmark layer for LLM workflows.

It does not claim to transfer a model's underlying capability. It packages the repeatable parts of strong agentic work: evidence discipline, systematic investigation, fail-closed gates, model routing, and honest benchmark records.

Why it exists

Teams often try to improve model behavior by copying long prompts or making vague quality claims. FableLayer takes a stricter approach:

• no private or proprietary prompt text
• no unverified performance claims
• no external publishing without approval
• no “done” without evidence
• no benchmark number without raw data and limitations

What you get

Layer	What it does	Files
PromptCore	Deterministic public-safe operating profile	fablelayer/promptcore.py, core/promptcore.md
Evidence Gate	Blocks completion claims without evidence	fablelayer/evidence_gate.py
Router	Cost-aware Sonnet/Opus/local routing rules	fablelayer/router.py
Adapters	Exports Claude Code, Ollama, LM Studio, SillyTavern profiles	fablelayer/adapters.py
Benchmark	Deterministic fixture scoring and raw JSON output	fablelayer/benchmark.py, bench/
Gates	License, performance, runtime, render, publish, completeness checks	gates/

Quick start

git clone https://github.com/VoidLight00/fablelayer.git
cd fablelayer

python3 tests/run_tests.py
bash gates/selftest.sh
bash gates/verify_fablelayer.sh . --mode new

./cli/fablelayer --help
./cli/fablelayer init --target ./_dist/demo --apply

See docs/INSTALL.md for detailed installation options.

CLI

./cli/fablelayer init                 # dry-run scaffold
./cli/fablelayer init --apply         # write local layer files
./cli/fablelayer upgrade sonnet       # preview routing decision
./cli/fablelayer benchmark            # run deterministic benchmark fixtures
./cli/fablelayer check --file output.md
./cli/fablelayer status
./cli/fablelayer resume

Verification

python3 tests/run_tests.py
bash gates/selftest.sh
bash gates/verify_fablelayer.sh . --mode new
python3 proof/verify_claims.py .

Current local evidence:

• 152 stdlib tests pass
• LICENSE/PERF/BENCH/COMPLETE/RENDER/RUNTIME gates pass
• publish without approval exits non-zero by design
• public claim checks verify version alignment, plugin references, and tracked-file hygiene

Proof and reproducibility

FableLayer separates verified procedural claims from unsupported capability-transfer claims.

• proof/CLAIMS.md defines what the project claims and does not claim.
• proof/REPRODUCIBILITY.md gives clean-checkout reproduction steps.
• proof/EXPERIMENT_DESIGN.md describes model-only vs model-plus-FableLayer comparisons.
• proof/THREATS_TO_VALIDITY.md records limits and failure modes.

The CI matrix runs tests, proof claims, and gates on Python 3.10–3.13.

Public-safe source policy

FableLayer is an original implementation. It may reference public methodology sources in ATTRIBUTION.md, but it does not copy private, proprietary, or non-public prompt text. High-risk source classes are blocked/reference-only and are not required to use this project.

Performance honesty

FableLayer's benchmark is designed to preserve raw data and limitations. Do not cite quality or cost claims without a bench/ reference. See bench/RESULTS.md.

Documentation

• docs/INSTALL.md
• docs/DEVELOPMENT.md
• CONTRIBUTING.md
• SECURITY.md
• ROADMAP.md
• ATTRIBUTION.md
• README.ko.md

License

AGPL-3.0. See LICENSE and NOTICE.