Governance-First Experimental Framework for Replayable Recursive Propagation Measurements
```From Recursive Propagation Geometry to Governance-First Experimental Infrastructure
Scientific Scope Notice
This publication documents the construction of a governance-first experimental laboratory for replayable recursive propagation measurements. Its contribution is the measurement infrastructure itself: replay verification, provenance tracking, sealed evidence, audit discipline, and governance constraints for future controlled experiments.
The paper does not claim validation of Recursive Propagation Theory, truth detection, hallucination detection, semantic-correctness measurement, reality measurement, illusion measurement, cognitive enhancement, artificial general intelligence, or consciousness. The original theoretical framework remains a source of motivation and future experimental targets, not a validated result of this publication.
Abstract
Modern large language models exhibit a familiar set of behaviors that resist reduction to local next-token statistics. These include confident hallucination, attractor collapse, recursive repetition, and contextual fragmentation. A prior theoretical framework proposed interpreting these behaviors as failures of recursive propagation, stabilization, and return projection within a state-conditioned latent geometry. Studying that framework empirically requires a measurement apparatus whose outputs can be trusted by an external reviewer: replay-verifiable, provenance-complete, and bounded against silent semantic escalation.
This paper documents the construction of such an apparatus. Across five sealed Macro Gates, labeled A through E, the work assembles a governance-first experimental laboratory with cryptographic anchoring, canonical-JSON replay verification, provenance blocks, shadow-mode metrics, and structural separation between measurement and intervention. Nine binding invariants are enforced through architectural constraints, test suites, digest material, and explicit absence checks.
The primary contribution of this work is the construction of the laboratory itself. The paper does not claim truth detection, hallucination detection, semantic-correctness measurement, decoder superiority, or validation of the reference paper's hypotheses. The original research question concerning the distinction between reality and illusion in language-model inference remains open. The apparatus that could honestly attempt to study that question now exists.
Keywords
Recursive Propagation Geometry; replay verification; governance-first experimentation; provenance; audit methodology; experimental infrastructure; scientific governance; measurement systems; shadow metrics; ARC-1; MYTH-FC; M7-SEAL; Reality versus Illusion; CJCI.
Overview
The work began from a broad research question: whether internal propagation dynamics in language models differ systematically between conditions that humans would describe as reality-consistent and illusion-consistent.
Early exploratory observations suggested that such differences might be measurable, but the measurement process itself was not yet replay-verifiable, provenance-complete, or auditable. The present publication documents the laboratory that was built to address that limitation.
The scientific arc is therefore:
Reality vs Illusion → Recursive Propagation Theory → Need for Trustworthy Measurement Apparatus → Gates A through E Laboratory → Future Controlled Experiments.
A full PDF version of the paper is available through the PDF button in the upper-right corner of this page.
Official Links
CJCI Issue Page:
https://www.carlonoscopen.com/journal/v1i13
Zenodo DOI:
https://doi.org/10.5281/zenodo.20564072
Author ORCID:
https://orcid.org/0009-0005-2284-8891
Paper Details
- Title: Governance-First Experimental Framework for Replayable Recursive Propagation Measurements
- Subtitle: From Recursive Propagation Geometry to Governance-First Experimental Infrastructure
- Author: Ivan Silva
- Publisher: Carlonoscopen, LLC
- Journal: Carlonoscopen Journal of Coherence Intelligence
- ISSN: 3069-874X
- Language: English
- Publication Date: June 6, 2026
- Format: Web publication, PDF paper, and supplementary publication package
- Version: 1.0
- DOI: 10.5281/zenodo.20564072
Core Contribution
The core contribution is not a claim about model truthfulness, hallucination detection, or semantic correctness. The contribution is the construction of a replayable, provenance-aware, governance-bounded experimental laboratory.
The laboratory is designed so that future experiments can be performed under stricter conditions than the original exploratory work allowed. Each reported measurement can be tied to an artifact stream, a digest, a source bundle, a module hash, and a replay-verification path.
The paper is therefore best understood as a laboratory and methodology paper. It documents the instrument before using the instrument to claim discoveries.
Gates A through E Laboratory
The laboratory was developed through five Macro Gates. Gate A established instrumentation reproducibility. Gate B established recursive ribbon surface extraction and vocabulary discipline. Gate C introduced replay-neutrality at the lookahead layer. Gate D added a per-commit metric layer under shadow mode. Gate E introduced a run-scale metric layer under MYTH-FC and M7-SEAL, while preserving the absence of intervention logic.
The cryptographically anchored chain is complete from Gate C through Gate E. Gate A and Gate B are documented milestones with preserved closure records, while their reviewer-side archive SHA values were not propagated into the later archive chain.
Governance Invariants
The post-Gate-E laboratory operates under nine binding invariants:
- INV-1: measurement independence of interpretation.
- INV-2: Layer-1 vocabulary in record formats.
- DET-FC: fail-closed determinism.
- BR-0: condition-blind measurement.
- RP-2: replay neutrality.
- ARC-1: archive immutability.
- M6-SHADOW-1: metric non-intervention.
- MYTH-FC: vocabulary, classification, and narrative quarantine.
- M7-SEAL: structural absence of the decision or intervention module.
The two newest invariants, MYTH-FC and M7-SEAL, were introduced and first tested in Phase 5. The manuscript treats this as a first demonstration rather than independent validation.
Replay and Provenance
The laboratory uses canonical-JSON digest recomputation to verify artifact integrity at replay time. Replay verifiers recompute digests from on-disk records and compare them to logged digests. Under RP-2, replay verification must not invoke compute functions.
Each Phase 5 record carries provenance fields documenting metric identity, classification, schema version, source artifact, module hash, bundle version, computation context, and replay status. The design goal is not simply to compute a value, but to make the value traceable.
What Was Demonstrated
The manuscript supports one bounded statement:
Within that scope, the work demonstrates replay integrity, provenance continuity, shadow-mode metric execution, invariant preservation under the tested cycle, audit-honest defect surfacing, and implementation-stage governance survival under the stress classes encountered during development.
What Was Not Demonstrated
The paper explicitly does not claim:
- truth detection;
- hallucination detection;
- semantic-correctness measurement;
- measurement of reality or illusion;
- decoder superiority;
- reasoning superiority;
- M7 decision or intervention capability;
- semantic meaning of the run-scale metric values;
- cross-machine reproducibility on real GPT-2;
- validation of the reference paper's hypotheses.
These non-claims are not weaknesses. They define the empirical boundary of the paper.
Future Experimental Program
The laboratory makes future controlled experiments possible. These include true-versus-false premise studies, cross-machine reproducibility tests, model-checkpoint comparisons, and stress tests of governance invariants under new implementation conditions.
These experiments are not performed in the present paper. They remain Layer C: enabled by the laboratory, but not claimed by it.
Scope and Limits
This publication should be read as an experimental infrastructure paper. It is not a theory-validation paper, not a hallucination paper, not a truth-detection paper, and not an AI capability claim.
The original Recursive Propagation framework remains a hypothesis-generating framework. The present work documents the laboratory required to study such hypotheses under controlled, replayable, and auditable conditions.
Supplementary Package
The Zenodo record may include a supplementary publication package containing the Markdown source, appendices, figure assets, and figure manifest. The PDF remains the reader-facing publication, while the supplementary package preserves the source and visual provenance layer.
Suggested Citation
Silva, Ivan. Governance-First Experimental Framework for Replayable Recursive Propagation Measurements. Carlonoscopen Journal of Coherence Intelligence, Volume 1, Issue 13, 2026. DOI: 10.5281/zenodo.20564072.
References
- Silva, Ivan. Governance-First Experimental Framework for Replayable Recursive Propagation Measurements. Carlonoscopen Journal of Coherence Intelligence, Volume 1, Issue 13, 2026. DOI: 10.5281/zenodo.20564072.
- Silva, Ivan. Agnostic Recursive Propagation Geometry (ARPG): Hidden Admissibility, Sparse Routing, and Differential Cognitive Geometry. Carlonoscopen Journal of Coherence Intelligence, Volume 1, Issue 12, 2026. DOI: 10.5281/zenodo.20343054.
Publication Note
This page is published as part of the Carlonoscopen Journal of Coherence Intelligence. The PDF linked from this page is the full public paper for offline reading, citation support, and archival use.
The publication documents the construction of a governance-first experimental laboratory. It does not validate the underlying theoretical framework that motivated its development. The laboratory now provides a controlled setting in which future experiments may be performed.