EladGenish/rnse-constraint-coherence: EladGenish/rnse-constraint-coherence

RNSE CCSC Replay Surface-only verification of RNSE coherence and constraint-consistent behavior under CCSC tests. One core. No internals. Byte-stable replay. Overview This repository provides a Minimum Sufficient Evidence (MSE) replay spine for verifying CCSC (Coherent + Constraint-Consistent) test results produced by RNSE (Recursive Null-Seed Engine). All results in this repository are generated by a single RNSE core. There are: no model variants no retraining runs no parameter sweeps no task-specific tuning Behavioral differences arise solely from external test structure and evaluation, not from changes to the engine. RNSE is treated here as a black-box testbed. Engine internals are explicitly out of scope. What This Repo Is (and Is Not) This repo is: A surface-only verification package for CCSC results A portable, byte-stable replay bundle for third-party audit A comparison between RNSE and baseline methods under identical constraints Evidence of trajectory coherence under constraint pressure This repo is not: A disclosure of RNSE internals A training or optimization benchmark A parameterized model family A claim about optimality or performance maximization This repository locks evidence, not implementation. CCSC: Coherent + Constraint-Consistent Tests CCSC tests evaluate systems along two orthogonal axes: Coherence How consistent successive steps are over time (lower average step divergence = higher coherence) Constraint Consistency Whether the system preserves feasibility margins ("slack") under constraints without violating boundaries or diffusing incoherently RNSE is evaluated against baseline methods such as: random walk uniform sampling greedy / local heuristics The tests do not reward random admissibility hits. They expose whether a system maintains governed trajectories under pressure. Key Result (High-Level) RNSE exhibits: High trajectory coherence (low step divergence) Full constraint-admissibility under CCSC rules Preserved slack under sustained constraint load No boundary thrashing or diffusion Baseline methods may satisfy constraints transiently, but do so via: incoherent jumps stochastic grazing boundary instability RNSE prefers coherent survival over opportunistic admissibility. This behavior is intentional and diagnostic. Repository Layout . ├── README.md # This document ├── QUICKSTART.md # One-page usage guide ├── run_test_B_ccsc.py # CCSC test runner ├── ccsc_out/ │ ├── ccsc_summary.json # Audit-relevant CCSC metrics │ └── ccsc_results.png # Visualization (human-readable) ├── mse_ccsc_bundle.py # MSE bundle generator + validator └── mse_ccsc_bundle/ ├── config.json ├── digests.json ├── bundle_manifest.json ├── verify_ccsc_replay.py └── README.md Replay Spine (MSE) This repo follows a surface-only replay model. A replay is VALID iff: all retained artifacts exist all SHA-256 hashes match the bundle digest matches the canonical manifest CCSC rule outputs are verified from retained files No engine execution is required to verify the claim. How to Verify (Any Machine) python mse_ccsc_bundle.py validate mse_ccsc_bundle or, from inside the bundle: python mse_ccsc_bundle/verify_ccsc_replay.py This performs: per-artifact hash verification manifest canonicalization bundle digest recomputation PASS / FAIL output No RNSE internals are needed. Determinism & Portability Replay validity does not depend on: CPU or GPU model Operating system Python micro-version File paths or folder names As long as the retained artifacts match their hashes, the replay is valid. Non-Goals This repository does not: enforce engine determinism specify update rules or architectures define optimization targets claim universal superiority It provides disciplined, verifiable behavioral evidence under partial observability. Status This repository reflects ongoing experimental work. Future CCSC tests and additional regimes will be added over time — all produced by the same RNSE core and packaged with the same replay discipline. License Apache License 2.0 See LICENSE for details.