Commitment Conservation Harness v2.0

Falsification instrument for the commitment conservation law.

DOI

📄 Paper: A Conservation Law for Commitment in Language Under Transformative Compression (v0.04) 🏗️ Zenodo Record: https://zenodo.org/records/18792459 💻 GitHub: https://github.com/SunrisesIllNeverSee/commitment-conservation


Quick Start

# Clone and install
git clone https://huggingface.co/burnmydays/commitment_conservation_harness
cd commitment_conservation_harness
pip install -r requirements.txt

# Run the interactive demo
python app.py
# Opens at http://localhost:7860

# Run the full falsification protocol (CLI)
python -m src.runner --backend lossy --depth 10

# Run tests (53 tests, <1 second)
pip install pytest
python -m pytest tests/ -v

A Conservation Law for Commitment in Language Under Transformative Compression and Recursive Application

Abstract

This repository accompanies a preprint introducing a conservation law for commitment in language under transformative compression and recursive application. We formalize commitment as an information-bearing invariant that must be preserved across paraphrase, summarization, and iterative reuse, even as surface form and representation change.

We propose a falsifiability framework based on compression-driven stress tests and lineage-aware evaluation, distinguishing semantic preservation from mere token retention. The framework is model-agnostic and applies to both human and machine-generated language.


Core Claims

  • Commitment Conservation: Meaningful commitments in language obey a conservation constraint under compression and recursive reuse.
  • Dual Stress Regime: Preservation must hold under both transformative compression and recursive application, exposing failure modes not captured by retrieval benchmarks.
  • Falsifiability: Commitment preservation can be empirically tested using compression-based stress tests and lineage-aware metrics.

Harness v2.0 Architecture

app.py                  <- Gradio interactive demo
src/
  extraction.py         <- Modal-pattern sieve (Definition 2.4, Figure 4)
  fidelity.py           <- Min-aggregated scoring: F = min(Jaccard, Cosine, NLI proxy)
  compression.py        <- Pluggable backends (extractive, BART, back-translation)
  lossy.py              <- Deterministic LLM drift simulator (pure Python, no model)
  enforcement.py        <- Reject-and-retry commitment gate (Module 3)
  lineage.py            <- SHA-256 provenance chains (Module 2)
  runner.py             <- Falsification protocol orchestrator (Section 7)
corpus/
  canonical_corpus.json <- 25 pinned signals across 5 categories
tests/
  test_harness.py       <- 53 tests covering all modules

Key improvements over v1

Issue v1 v2
Extractors 3 incompatible implementations 1 modal-pattern sieve
Enforcement String concatenation (stripped next cycle) Reject-and-retry gate with re-injection
Fidelity Jaccard only min(Jaccard, Cosine, NLI proxy)
Lineage None SHA-256 provenance chains
Collapse detection None Cross-signal attractor check
False positives "will", "have" match as commitments Blocked by pattern design
Demo backend BART (2GB model download) Lossy simulator (pure Python, zero deps)
Backend pairing Same backend for both runs Auto-pairs lossy with lossy_enforced

Empirical Results

Paper results (Section 7)

Compression + lineage enforcement vs. probabilistic baseline on 175 signals (100 sentences, 50 code snippets, 25 proofs) over 10 recursive iterations:

Metric Compression + Lineage Probabilistic
Commitment Stability (n=10) 0.94 ± 0.03 0.42 ± 0.12
Identity Preservation 92% 38%
Drift Rate (per iteration) 0.006 0.058

Harness v2 results (lossy backend, drift=0.4, depth=10)

25-signal canonical corpus across 5 categories:

Baseline Enforced Gap
Avg Final Fidelity 0.000 0.479 +47.9pp
Contract (payment + penalty) 0.000 1.000 +100.0pp
Security (encrypt + prohibit) 0.000 0.500 +50.0pp

Baseline → 0.0 (commitments destroyed, text becomes "Pretty straightforward. Nothing too complicated here.") Enforced → holds (modal-bearing sentences survive compression).


Version History

Version Label Date DOI
V.1-preprint Law Disclosure Jan 12, 2026 10.5281/zenodo.18267279
V.02 Preprint Jan 16, 2026 10.5281/zenodo.18271102
V.03 Falsifiability Testing Jan 16, 2026 10.5281/zenodo.18274930
V.04 Technical Structure Depth Feb 26, 2026 10.5281/zenodo.18792459

Resources


Licensing & Scope

This work is released under Creative Commons Attribution 4.0 International (CC BY 4.0).

This repository includes an operational evaluation harness and corpus supporting the experiments described in the paper.

Core implementation details related to production deployment, enforcement, and system integration are intentionally out of scope.


Attribution & Contact

Author: Deric J. McHenry Copyright: © 2026 Ello Cello LLC. All rights reserved. Affiliation: Ello Cello LLC Contact: burnmyday@proton.me

For academic or research correspondence, please reference the Zenodo DOI above.


Intellectual Property Notice

MO§ES™ (Minimal Orthogonal Subset to Essential Structure) is a trademark of Ello Cello LLC.

All systems, methods, and written material contained herein are protected under applicable copyright law and are part of one or more provisional patent applications filed with the United States Patent and Trademark Office (USPTO).

© 2026 Ello Cello LLC. All rights reserved.

Unauthorized use, duplication, or redistribution of any part of this communication — including proprietary system architecture or terminology — is strictly prohibited and may result in legal action.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Space using burnmydays/commitment_conservation_harness 1