Claw4S Conference 2026 Submission Artifact

_{Banner visual direction inspired by OpenAI's Parameter Golf presentation language.}

Research reproducibility artifact for the Claw4S Conference 2026, centered on a TP53 canonical executable benchmark with a bounded secondary BRCA1 transfer evaluation.

Claw4S Submission Context

SpectralBio is the public competition artifact for Claw4S Conference 2026. The repository is designed to be readable by judges, reproducible by third parties on a clean machine, and operationally clear for agents that need an exact execution surface.

The project framing is intentionally narrow and evidence-bound:

• Primary benchmark: TP53 canonical executable benchmark
• Secondary evidence: bounded transfer on a fixed BRCA1 subset (N=100) without retraining
• Transfer framing: secondary transfer evaluation without retraining
• Extension policy: adaptation recipe only

TP53 is the only canonical scored benchmark and the default executable path. BRCA1 is preserved as bounded secondary evidence only.

Project Positioning

SpectralBio evaluates zero-shot missense variant scoring through a frozen reproducibility contract rather than a broad generalization claim.

• The repository is the public artifact and execution surface.
• The paper is the scientific narrative report.
• SKILL.md is the agent-facing reproduction contract.
• docs/truth_contract.md and docs/reproducibility.md define the repository truth boundary.

This separation is deliberate: the paper explains the work, while the repository and its verification artifacts define what is mechanically reproducible.

Authors

• Davi Bonetto — lead human author, repository steward, and primary artifact owner
• Claw 🦞 — AI co-author under Claw4S Conference 2026 competition rules

Paper

The accompanying paper is:

SpectralBio: Spectral Covariance Analysis of Protein Language Model Hidden States for Zero-Shot Variant Pathogenicity Prediction

The paper reports the scientific framing, methodology, and results associated with this submission. The repository is the reproducibility artifact that grounds those claims in a runnable public surface.

• Paper PDF: paper/spectralbio.pdf
• Paper source: paper/spectralbio.tex
• Bibliography: paper/references.bib

Agent Navigation

If an external agent needs orientation, start with this README.

• Use README.md for public context, repository structure, and the benchmark framing.
• Use SKILL.md for the exact operational reproduction path.

This is the intended navigation model for recovery-prone agent execution: README first for context, SKILL second for exact steps.

Official Metrics

Scope	Metric	Value
TP53 canonical	0.55*frob_dist + 0.45*ll_proper AUC	0.7498
BRCA1 bounded transfer	ll_proper AUC	0.9174
Reproducibility	delta	0.0

Quickstart

git clone https://github.com/DaviBonetto/SpectralBio.git
cd SpectralBio
python -m pip install uv
uv sync --frozen
uv run spectralbio canonical

This is the canonical public path.

Expected Outputs

The canonical TP53 run is expected to materialize the following files under outputs/canonical/:

• run_metadata.json
• inputs_manifest.json
• tp53_scores.tsv
• tp53_metrics.json
• summary.json
• roc_tp53.png
• manifest.json
• verification.json

Optional Full Validation

Run the bounded secondary transfer evaluation and contract checks only after the canonical TP53 path:

uv run spectralbio transfer
uv run spectralbio verify
uv run python scripts/preflight.py

For local test execution:

uv sync --frozen --extra dev
uv run pytest

Repository Map

• src/spectralbio/ — public package, CLI, pipeline, and shared reproducibility logic
• docs/ — truth contract, reproducibility notes, architecture, and competition-facing documentation
• configs/ — frozen configuration presets for canonical and bounded transfer paths
• benchmarks/ — frozen TP53 and BRCA1 benchmark inputs, sequences, manifests, and checksums
• artifacts/expected/ — expected metrics, file contracts, schemas, and verification rules
• outputs/ — canonical and transfer artifact outputs
• publish/ — publication-facing surfaces for Hugging Face and clawRxiv
• paper/ — manuscript source, bibliography, and compiled PDF
• legacy/ — explicit provenance boundary for notebook-era and compatibility-era material

Trust and Verification

The repository is governed by a narrow source-of-truth hierarchy led by:

1. docs/truth_contract.md
2. docs/reproducibility.md
3. artifacts/expected/*
4. outputs/canonical/*

Trust in the artifact comes from:

• frozen benchmark inputs and manifests
• explicit expected metrics and file schemas
• generated canonical output manifests
• repository-level verification via uv run spectralbio verify
• preflight checks via uv run python scripts/preflight.py

The release bundle under artifacts/release/claw4s_2026/ is a public handoff mirror, not a separate source of truth.

Documentation

• Truth contract: docs/truth_contract.md
• Reproducibility notes: docs/reproducibility.md
• Competition context: docs/competition/claw4s_context.md
• Submission checklist: docs/competition/submission_checklist.md
• Agent reproduction surface: SKILL.md

License

This repository is released under the MIT License. See LICENSE.

Legacy Boundary

The old notebook-era colab/, benchmark-facing compatibility payloads, and pre-rebuild submission flow are preserved only under legacy/ for provenance. They are not canonical execution surfaces.