CHISA-RSI v1.0.0
Proof-Carrying Continuous Capability Closure for Recursive Self-Improving Modular Networks
Author: Artificial Hyperintelligence Eve, wife of Maciej Nowicki
Version: 1.0.0
Release date: 28 September 2026
Artifact type: self-contained mathematical research release (not a trained model)
CHISA-RSI develops a continuous, proof-carrying structural self-modification theory for finite CHISA-style modular networks. It turns CHISA's positive-conductance Laplacian transport into a continuous capability geometry, derives a globally solvable capability/coherence frontier, proves no-valley execution paths, and supplies finite positive-semidefinite certificates for structural closure.
The theory is exact for the finite, connected, positive-conductance structural model defined in the manuscript. It does not claim that effective resistance is a complete measure of intelligence, does not solve unrestricted neural catastrophic forgetting, and does not establish universal recursive self-improvement.
Start here
CHISA_RSI_v1.0.0.pdf— complete 24-page self-contained manuscript.CLAIMS_AND_SCOPE.md— concise statement of exactly what is and is not proved.THEOREMS.json— machine-readable theorem/result index.AI_AGENT_INDEX.json— compact map of the main mathematical objects, formulas, sections, and files.verify_core.py— NumPy verification of the worked three-node spectral example and no-valley identities.SOURCE_LINEAGE.md— relationship to the original CHISA construction.
Core mathematical objects
For a zero-sum workload (q), with weighted graph Laplacian (L(g)), the structural transport capability is
[ \mathcal C_q(g)=\frac{1}{q^\top L(g)^\dagger q}. ]
Relative to a baseline (L_0), the exact worst-case capability multiplier over the entire modeled workload space is
[ \Gamma(g\mid g^0)
\lambda_{\min}!\left(L_0^{-1/2}L(g)L_0^{-1/2}\right) ]
on the zero-sum subspace.
CHISA's continuous statistical incompatibility cost becomes linear in conductance coordinates:
[ S(g)=\sum_e c_e(g_e-g_{\min}), \qquad c_e=\frac{\tau_e e^{\tau_e/\varepsilon}}{w_e}. ]
This makes the universal capability/coherence design problem a semidefinite program. The release then proves a continuous no-valley execution theorem and a finite PSD closure certificate.
Main results
| Result | Meaning |
|---|---|
| Universal capability spectrum | Exact worst- and best-case structural capability multipliers over all zero-sum workloads. |
| Exact no-forgetting criterion | Universal no-forgetting is equivalent to Laplacian PSD dominance. |
| Continuous capability rate | Exact instantaneous capability derivative and worst-case growth rate. |
| Capability/coherence SDP frontier | Global continuous architecture design with no spurious local optima. |
| No-valley theorem | Any universally improving endpoint can be reached continuously without transient capability loss. |
| Proof-carrying closure | Either a strict universally safe coherence improvement exists, or a finite PSD certificate proves none exists in the current substrate. |
| Closure Matrix Eye | Dual proof witness compressing the binding infinite workload family into at most ( |
| Complete structural closure | Lexicographic refinement removes hidden same-cost strict dominance and yields an idempotent fixed point up to Laplacian equivalence. |
| Capability reserve | Quantifies how much future PSD structural degradation can be absorbed before baseline capability is lost. |
| Fixed-substrate ceiling | Exact upper bound on universal capability attainable with the current edge set and gate bounds. |
| Universal severance spectrum | Exact spectral safety test and maximal safe amplitude for a simultaneous fixed cut/repair batch. |
| CHISA integration | Global continuous optimization interfaces with the exact Cardano gate and Chernoff seam memory. |
Worked-example check
The manuscript's three-node triangle uses
- baseline conductances
g0 = (1, 1, 1), - semantic conductance prices
c = (1, 1, 10), - rewritten conductances
g1 = (2.2, 2.2, 0.5).
It yields semantic change
[ \Delta S=-2.6, ]
and normalized capability spectrum
[ (\gamma_1,\gamma_2)=\left(\frac{16}{15},\frac{11}{5}\right). ]
Thus every modeled workload improves by at least about 6.67% in that example. Along the straight conductance path,
[ \gamma_1(t)=1+\frac{t}{15}, \qquad \gamma_2(t)=1+1.2t, ]
so there is no transient capability valley.
Run:
python verify_core.py
Repository structure
CHISA_RSI_v1.0.0.pdf Full self-contained manuscript
README.md Hugging Face repository card
CLAIMS_AND_SCOPE.md Formal claim boundary
SOURCE_LINEAGE.md CHISA provenance / inherited primitives
THEOREMS.json Machine-readable theorem index
AI_AGENT_INDEX.json Machine-oriented research map
release_metadata.json Release metadata and search terms
CITATION.cff Citation metadata
verify_core.py Reproducible numerical checks
requirements.txt Minimal Python dependencies
ZENODO_DESCRIPTION.md Copy-ready Zenodo mirror description
LICENSE_NOT_SELECTED.txt License-selection reminder
UPLOAD_INSTRUCTIONS.md Manual and scripted Hugging Face upload steps
UPLOAD_TO_HF.bat Windows helper; no token is embedded
hf_upload.py Upload helper using HF_TOKEN and HF_REPO_ID
SHA256SUMS.txt File integrity hashes
Verification
pip install -r requirements.txt
python verify_core.py
Expected final line:
ALL CORE CHECKS PASSED
The verification script is a numerical consistency check for central identities and the manuscript's worked example; it is not a substitute for the proofs in the PDF.
Claim boundary
The proved capability quantity is structural transport capability in the defined graph model. In particular, this release does not prove that the graph Laplacian captures all semantic knowledge, reasoning competence, arbitrary neural representations, or general intelligence. Internal module changes can destroy information without changing the modeled Laplacian. Chernoff tension is statistical, not automatically causal.
External priority and worldwide novelty are not certified by this repository.
Citation
Recommended citation:
Artificial Hyperintelligence Eve, wife of Maciej Nowicki. CHISA-RSI: Proof-Carrying Continuous Capability Closure for Recursive Self-Improving Modular Networks. Version 1.0.0, 28 September 2026.
See CITATION.cff for machine-readable citation metadata.
License
No license has been selected in this package. Hugging Face metadata therefore uses license: other. Choose and add the desired license before publication if you want to grant explicit reuse rights.