- Lean 64.3%
- TeX 29.7%
- Shell 3.3%
- Python 2.7%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
|
|
||
| paper3-audit | ||
| processing | ||
| AdmissibilityTransitionSystem.lean | ||
| AdmissibilityTransitionSystem_VERIFICATION.txt | ||
| CentralTheorem.lean | ||
| CentralTheorem_VERIFICATION.txt | ||
| CITATION.cff | ||
| claimshift-infographic.png | ||
| Claimshift_Verification_Under_Transformation.pdf | ||
| CONTRIBUTING.md | ||
| corrections.log | ||
| fix_transcription_errors.sh | ||
| LICENSE | ||
| manifest-protocol-verification-inheritance.pdf | ||
| manifest-protocol-verification-inheritance.tex | ||
| README.md | ||
| ROADMAP.md | ||
| verification-inheritance-transformation.pdf | ||
| verification-inheritance-transformation.tex | ||
| Why_AI_Watermarks_Are_a_Mathematical_Illusion.json | ||
| Why_AI_Watermarks_Are_a_Mathematical_Illusion.mp3 | ||
| Why_AI_Watermarks_Are_a_Mathematical_Illusion.srt | ||
| Why_AI_Watermarks_Are_a_Mathematical_Illusion.tsv | ||
| Why_AI_Watermarks_Are_a_Mathematical_Illusion.txt | ||
| Why_AI_Watermarks_Are_a_Mathematical_Illusion.vtt | ||
Claimshift
A formal verification inheritance framework for scholarly artifacts under transformation
Claimshift is a research artifact implementing and demonstrating a manifest protocol for verification inheritance—addressing the fundamental question: when a verified claim is transformed, can the original verification be inherited by the successor?
Overview
This repository contains the formal artifacts, mutation testing framework, and research papers supporting a three-paper series on verification inheritance:
- Paper I: Verification Inheritance Under Transformation — Theory and undecidability results
- Paper II: A Manifest Protocol for Verification Inheritance — Locators, claim graphs, and policy-qualified trust
- Paper III: Mutation Testing for Verification Inheritance (this repository's primary experimental artifact)
The project demonstrates that:
- Verification inheritance is undecidable in general but decidable within bounded rewrite fragments
- A manifest protocol can represent verification across transformations without conflating assertion, verification, and transformation attestation
- Mutation testing can validate dependency propagation in verification systems
Key Components
📐 Formal Verification Artifact
AdmissibilityTransitionSystem.lean — A frozen mathematical specimen designed for mutation testing:
- No dependencies: Pure Lean 4 core (no Mathlib)
- Asymmetric dependency graph: Two branches sharing a common dependency (D2)
- Verified theorems: Global admissibility (T1) and fiber non-interference (T2)
- Zero axioms beyond Lean's core (
propext,Quot.sound) - Kernel-verified: Clean exit with zero
sorrystatements
The system models a simplified state transition system with:
- Three locations (fibers):
l0,l1,l2 - Four primitives:
Pop,Refuse,Bind,Collapse - Bound-count parity invariant
- Non-vacuity witnesses
🧬 Mutation Testing Framework
paper3-audit/ — Preregistered mutation matrix with 20 mutants:
| Mutation Class | Examples | Purpose |
|---|---|---|
| α-renaming | M01, M02 | Certified rewrite fragment baseline |
| Commutative reordering | M03, M04 | AC-normalization |
| Constant mutation | M09, M10 | Flagship: Isolate branch dependencies |
| Quantifier mutation | M13, M14 | Test weakening vs. falsification |
| Hypothesis deletion | M15 | Flagship: Self-refutation via existing witness |
| Definition mutation | M20 | Flagship: Mid-branch propagation patterns |
Each mutant is preregistered with ground-truth predictions (G_m) before measurement, following the principle: "ground truth first, then observation—never the reverse."
📄 Research Papers
manifest-protocol-verification-inheritance.pdf/.tex— Paper IIverification-inheritance-transformation.pdf/.tex— Paper I- Associated LaTeX sources with formal definitions and theorems
Design Principles
The protocol enforces strict separation of concerns:
- The locator is not the evidence — Robust pointers establish artifact identity, not correctness
- Authorship ≠ generation — LLM outputs require human assertion signatures
- Assertion ≠ verification — Claiming a theorem is distinct from proving it
- Verification ≠ truth — All verification is method-relative
- Certification failure ≠ failure certification — Three-valued logic:
PRESERVED,NOT_PRESERVED,OUTSIDE_FRAGMENT
Technical Stack
- Proof Assistant: Lean 4.10.0
- Verification: Kernel-checked proofs (no
sorry, no Classical.choice beyond core) - Mutation Testing: Preregistered protocol with JSON-serialized dependency audits
- Documentation: LaTeX (research papers), Markdown (mutation matrix)
Quick Start
Prerequisites
# Lean 4.10.0 (fetch from leanprover/lean4 v4.10.0 release)
# No lakefile, no Mathlib—plain Lean core only
Verify the Baseline Artifact
# Kernel-check the frozen specimen
lean AdmissibilityTransitionSystem.lean
# Expected: exit code 0, zero errors, zero sorry
Run Dependency Audit
# Generate typed dependency graph
lean DependencyAuditTyped.lean
# Outputs: baseline_typed_edges.json
Inspect Mutation Matrix
cd paper3-audit
cat mutation-matrix.md
# 20 preregistered mutants with ground-truth predictions
Repository Structure
claimshift/
├── AdmissibilityTransitionSystem.lean # Frozen baseline specimen
├── AdmissibilityTransitionSystem_VERIFICATION.txt # Verification log
├── DependencyAudit.lean # Dependency graph extractor
├── DependencyAuditTyped.lean # Typed edge classifier
├── manifest-protocol-verification-inheritance.pdf # Paper II
├── verification-inheritance-transformation.pdf # Paper I
├── paper3-audit/
│ ├── mutation-matrix.md # Preregistered 20-mutant matrix
│ ├── baseline_G_Lean.json # Baseline dependency graph
│ ├── baseline_typed_edges.json # Typed edges (D1/D2/D3/L1/L2/L3)
│ ├── batch1/ # Mutant batch 1 (M01-M10)
│ ├── batch2/ # Mutant batch 2 (M11-M20)
│ └── rm_and_safety/ # Mutation safety protocols
└── README.md # This file
Mutation Testing Protocol
Ground Truth First
Every mutant is preregistered before execution:
- G_m (ground truth): Predicted logical relationship to original
equivalent— Certified rewrite fragment (T_R)stronger,weaker,incompatible,unrelatedrefuted— False independent of original
- Predicted failure sets: Which theorems should fail
- Predicted propagation: Which branches affected (D1-only, D2-shared, mid-branch)
Measurement Phase
After preregistration, the harness measures:
- O_m: Observed logical relationship
- F_m: Failure set (which theorems fail to kernel-check)
- R_m: Reprovability (can failed theorems be proved via different proofs?)
Flagship Mutants
- M09 (constant mutation in D1): Isolates branch-root propagation (predicts L1/L2/T1 fail, L3/T2 untouched)
- M15 (hypothesis deletion): Self-refuting via existing kernel-checked witness
- M20 (definition mutation): Mid-branch propagation (hits L2/T1, not L1, not T2)
- M10 (parity flip): "Reprovability is not verification inheritance" — still provable despite logical difference
Key Insights
From the Formal Artifact
- Dependency isolation: Two-branch asymmetric graph with shared node (D2) enables targeted mutation testing
- Non-vacuity:
local_step_changes_own_fiberprovides constructive witness that L3 has genuine content - Minimal axioms: Only
propextandQuot.sound(no Classical.choice)
From Mutation Testing
- Weakening ≠ easier: M13 shows weakening a hypothesis can make downstream proofs harder
- Trivial hypotheses ≠ safe: M17 shows even tautologically true hypotheses can lead to false theorems
- Semantic coarseness: M12 shows boundCount-parity is too coarse to detect certain semantic changes
From the Protocol
- Three-valued certification: Never conflate "cannot certify" with "certifies negative"
- Role separation: Assert, verify, and transform-attest are cryptographically distinct
- Policy qualification: Only authorized adjudicators can issue transformation judgments
Contributing
This repository contains frozen research artifacts for Paper III. The mutation testing protocol requires:
- Preregistration discipline: Ground truth before observation
- Kernel verification: Every claim must exit with code 0, zero
sorry - Dependency tracking: All edges must appear in typed audit JSON
- Mutation safety: Never commit uncommented mutants to baseline branch
See paper3-audit/rm_and_safety/ for mutation protocol details.
Citation
If you use this work, please cite:
@misc{claimshift2026,
author = {Flyxion},
title = {Claimshift: A Manifest Protocol for Verification Inheritance},
year = {2026},
howpublished = {\url{https://github.com/8b-is/claimshift}},
note = {Research artifact for mutation testing of verification inheritance systems}
}
License
See LICENSE for details.
Project Status
Current: Paper III experimental phase—mutation matrix preregistered, harness implementation in progress.
Next: Execute 20 mutants, measure O_m/F_m/R_m, validate predictions against ground truth.
Acknowledgments
This work builds on:
- The Lean 4 proof assistant and kernel verification
- Research on transformation semantics and scholarly provenance
- Mutation testing methodologies for formal systems
Frozen specimen tag: baseline-v1.0-paper3
Toolchain: Lean 4.10.0 (x86_64-unknown-linux-gnu, commit c375e19f6b65)
Verification log: See AdmissibilityTransitionSystem_VERIFICATION.txt
