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.
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
| 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/ |
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 --applySee docs/INSTALL.md for detailed installation options.
./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 resumepython3 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/RUNTIMEgates pass- publish without approval exits non-zero by design
- public claim checks verify version alignment, plugin references, and tracked-file hygiene
FableLayer separates verified procedural claims from unsupported capability-transfer claims.
proof/CLAIMS.mddefines what the project claims and does not claim.proof/REPRODUCIBILITY.mdgives clean-checkout reproduction steps.proof/EXPERIMENT_DESIGN.mddescribes model-only vs model-plus-FableLayer comparisons.proof/THREATS_TO_VALIDITY.mdrecords limits and failure modes.
The CI matrix runs tests, proof claims, and gates on Python 3.10–3.13.
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.
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.