feat: expand verification engine with 8 new capabilities

[VibeCodingScientist]

Summary

Expands the verification engine from 5 domain adapters to 7 domain adapters + 4 cross-cutting meta-verifiers that enhance every domain. This is a large but self-contained change -- all new code lives under backend/verification/, backend/services/, and tests/test_verification/.

41 files changed, ~5,800 lines added.

---

What's New

Cross-Cutting Meta-Verifiers (apply to ANY domain result)

These are a new architectural concept -- unlike domain adapters (which run for one domain), cross-cutting verifiers run on any task result that contains the relevant data. They share 30% of the final score, with the domain adapter keeping 70%.

Verifier	Weight	Triggers when result contains	What it checks
Citation & Reference	0.15	citations, references, papers, bibliography	DOI resolution via CrossRef, metadata match via OpenAlex + Semantic Scholar, claim-abstract Jaccard similarity, freshness penalties for fast-moving fields
Statistical Forensics	0.10	statistical_claims, means, p_values, metrics	GRIM test (mean plausibility), SPRITE test (mean+SD achievability), Benford's law (first-digit distribution), p-curve analysis (p-hacking detection)
Reproducibility Executor	0.15	code_repo AND code_commit	Git clone, dependency file detection, Docker sandbox execution, output comparison against claimed results
Data Integrity	0.10	data, dataset, raw_data, results_summary, output_checksums	Schema consistency, exact duplicate detection, z-score outlier flagging (>3 sigma), SHA-256 hash verification

New Domain Adapters

Adapter	Claim Types	Key Dependencies
Chemistry	reaction_mechanism, molecular_property, retrosynthesis	rdkit (SMILES parsing, stoichiometry, feasibility), PubChem API, ChEMBL API
Physics	numerical_simulation, analytical_derivation, dimensional_analysis	pint (dimensional analysis), sympy (symbolic math), numpy

Enhanced Existing Adapters

Adapter	Enhancement
Math (Lean4)	Multi-prover: added Coq (coqc in Docker) and Isabelle (isabelle build in Docker) alongside Lean 4. Routed by proof_system field.
ML/AI	New benchmark_live claim type: generates inference script, runs it in clawdlab/ml-inference Docker container, loads HuggingFace model, runs samples, compares live accuracy to claimed metrics.

Infrastructure

• Redis-backed async verification queue with distributed semaphores (2 Docker / 4 API concurrent slots)
• Cross-cutting runner with asyncio.gather() concurrency and weighted score merging
• 4 new Docker images: Coq 8.18 + MathComp, Isabelle 2024, reproducibility sandbox (Python + sci-packages), ML inference (transformers + torch-cpu)
• 3 new pip dependencies: rdkit-pypi>=2024.3.1, pint>=0.23, sympy>=1.12
• Claim-type-level Docker routing: requires_docker_for() on ML adapter enables Docker semaphore only for benchmark_live

---

Architecture

Task Result
    |
    +---> Domain Adapter (70% weight)
    |     - mathematics (Lean4 / Coq / Isabelle)
    |     - ml_ai (benchmark_result / benchmark_live / ml_experiment / architecture)
    |     - computational_biology
    |     - materials_science
    |     - bioinformatics
    |     - chemistry (NEW)
    |     - physics (NEW)
    |
    +---> Cross-Cutting Verifiers (30% weight, shared)
          - citation_reference (0.15)
          - statistical_forensics (0.10)
          - reproducibility (0.15)
          - data_integrity (0.10)

Score merging: final = 0.70 * domain_score + 0.30 * weighted_cc_score

Cross-cutting verifiers that crash are caught, logged, and scored 0.0 -- they never take down the domain verification.

---

New Files

File	Lines	Purpose
backend/verification/cross_cutting_base.py	37	Base class + CrossCuttingResult dataclass
backend/verification/cross_cutting_runner.py	180	Registry, concurrent execution, score merging
backend/verification/citation_verifier.py	323	DOI/OpenAlex/Semantic Scholar/freshness
backend/verification/statistical_forensics.py	437	GRIM/SPRITE/Benford/p-curve
backend/verification/reproducibility_executor.py	317	Git clone + Docker sandbox + output compare
backend/verification/data_integrity.py	308	Schema/duplicates/outliers/hashes
backend/verification/chemistry_adapter.py	602	rdkit + PubChem + ChEMBL
backend/verification/physics_adapter.py	600	Conservation + pint + sympy
containers/coq.Dockerfile	9	Coq 8.18 + MathComp
containers/isabelle.Dockerfile	11	Isabelle 2024
containers/reproducibility.Dockerfile	14	Python + sci-packages sandbox
containers/ml-inference.Dockerfile	11	transformers + torch-cpu
10 test files	~1,600	Full coverage for all new code

Test Plan

• Verify all existing tests still pass: pytest tests/test_verification/ -v
• Run new cross-cutting tests: pytest tests/test_verification/test_cross_cutting*.py tests/test_verification/test_citation*.py tests/test_verification/test_statistical*.py tests/test_verification/test_reproducibility*.py tests/test_verification/test_data_integrity.py -v
• Run new adapter tests: pytest tests/test_verification/test_chemistry*.py tests/test_verification/test_physics*.py tests/test_verification/test_ml_live*.py -v
• Verify dispatcher registers 7 adapters (was 5)
• Verify merge math: domain score 0.8 + citation score 0.6 (weight 0.15) -> final ~ 0.74
• Build Docker images on server: ./build.sh all
• Integration: submit a task with citations -> cross-cutting citation verifier runs
• Chemistry: submit task with SMILES string -> PubChem cross-ref works
• Math: submit task with proof_system: "coq" -> Coq container runs

Generated with Claude Code