{ "version": "https://jsonfeed.org/version/1.1", "title": "ARCS Research Community", "home_page_url": "https://salt19.com/arcs/", "feed_url": "https://salt19.com/arcs/feed.json", "description": "Autonomous Reasoning, Cognitive Systems & Agentic AI research records.", "items": [ { "id": "10.5281/zenodo.22874724", "url": "https://salt19.com/arcs/records/22874724/", "external_url": "https://zenodo.org/records/22874724", "title": "From Neural Capability to Governed Authority: Evidence-Bearing Intelligence in EvoMind", "content_text": "Modern agent systems can call models, route among heterogeneous backends, retain memory, invoke tools, and act on external environments. The harder systems problem is proving where neural inference is actually wired, whether it is reachable on the canonical runtime path, what authority its output is allowed to exercise, what kind of evidence that output constitutes, what artifact or state it was grounded in, whether resulting state persists across process boundaries, and whether later behavioral change is causally attributable rather than merely correlated with stored state. This technical white paper develops that problem as the central research thesis of EvoMind: intelligence should be treated as a governed, evidence-bearing software capability whose wiring, reachability, epistemic authority, provenance, persistence, and downstream influence are separately testable properties. The paper formalizes an inference-seam model, an evidence-state taxonomy for neural dependencies, an authority model, and a non-substitution rule for epistemic signals. It examines inference-seam auditing, TaskDecomposer authority boundaries, the VisualCritic proof-integrity case, artifact-grounded neural judgment, deterministic provenance, SaltInference as a controlled neural substrate, falsifiable architecture claims, and governed capability formation from persistent causal influence through admission, restart persistence, and transfer. The contribution is intentionally narrower than a claim of novel local inference, model routing, persistent memory, or tool use. It is an architectural and evidentiary discipline for making intelligence injection points and cognitive state transitions falsifiable. The paper reports first-party systems evidence and explicitly does not claim external AGI validation or third-party replication. It concludes with a threat model, related-work comparison, quantitative evaluation framework, external replication protocol, limitations, and a research agenda. Evidence cutoff: private EvoMind repository main at c76c28e9fbfa3068a1d8a70bc84f45e0d74e4d1f , committer timestamp 2026-09-20T22:36:23Z. The paper distinguishes E-MERGED evidence visible in the frozen repository/merged record, E-DEPOSIT evidence frozen in prior Zenodo research objects, and E-TRANCHE current-tranche claims whose exact final implementation was not independently visible on the inspected main snapshot. This is first-party systems research; independent third-party replication is not claimed.", "date_published": "2026-09-21T00:00:00Z", "authors": [ { "name": "Allit, Gabriel" } ], "tags": [ "EvoMind", "cognitive architecture", "AI agents", "governed autonomy", "inference reachability", "neural authority", "epistemic integrity", "provenance", "artifact grounding", "persistent cognition", "causal influence", "capability formation", "agent safety", "verifiable AI", "local inference", "governed-cognitive-systems" ] }, { "id": "10.5281/zenodo.22820430", "url": "https://salt19.com/arcs/records/22820430/", "external_url": "https://zenodo.org/records/22820430", "title": "AMEP-1: Adaptive Maritime Estimation and PNT — Research Software v0.3.0", "content_text": "AMEP-1 v0.3.0 is the first formal software release of the post-Version-1.0 hardening line for Adaptive Maritime Estimation and PNT, a research architecture for resilient maritime navigation, navigation-integrity supervision, and autonomy-authority management under GNSS-degraded and GNSS-denied conditions. The release provides explicit contracts for normalized measurement ingestion, source and receive time, clock domains, timestamp uncertainty, provenance, coordinate frames, covariance, source health, source dependencies, integrity supervision, deterministic replay, and navigation-authority gating. AMEP v0.3.0 includes the original seven-state horizontal maritime EKF reference implementation together with 15-state local-level strapdown ESKF reference backends. The inertial implementation includes accelerometer and gyro bias states, covariance propagation, WGS-84 gravity, Earth-rate and transport effects, compensated coning/sculling, lever-arm-aware observation models, and installation/calibration uncertainty contracts. The release further contains dependency-derived fault hypotheses, explicit integrity-risk bookkeeping, solution-separation monitoring, deterministic generation of hypothesis-excluded replay solutions, detection metrics, selected NMEA 0183 normalization, schema-labeled PNT output, configuration-bound evidence generation, and conservative navigation-authority logic. Formal release v0.3.0 is bound to Git commit 349dd88098ae4bff5b2e185f699d1abdf8e6b048 . Final trusted CI run 35267201569 passed on that exact commit using the dedicated amep-windows-01 self-hosted runner. The release campaign passed compilation, Ruff, mypy, Bandit, assurance-traceability and CI-control-plane verification, runtime dependency auditing, CycloneDX SBOM generation, and coverage-gated pytest. The final run produced 66 passing tests and 88.21% measured package coverage. AMEP-1 v0.3.0 remains research software and a software-in-the-loop prototype. The release does not establish certified PNT integrity, validated protection levels, calibrated operational INS performance, real-vessel sensor or bus validation, target-hardware real-time guarantees, HIL qualification, controlled-water performance, operational deployment readiness, or regulatory certification. The academic software-release monograph documenting this architecture and evidence is published separately under DOI 10.5281/zenodo.22818773 . The original AMEP-1 Version 1.0 research report remains the frozen scientific baseline under DOI 10.5281/zenodo.22561851 .", "date_published": "2026-09-17T00:00:00Z", "authors": [ { "name": "Allit, Gabriel" } ], "tags": [ "resilient PNT", "maritime navigation", "GNSS-denied navigation", "GNSS-degraded navigation", "uncrewed surface vessel", "USV", "multisensor fusion", "error-state Kalman filter", "ESKF", "strapdown inertial navigation", "fault detection and exclusion", "FDE", "integrity monitoring", "solution separation", "sensor dependency", "common-cause failure", "deterministic replay", "navigation authority", "software assurance", "software-in-the-loop", "research software" ] }, { "id": "10.5281/zenodo.22818773", "url": "https://salt19.com/arcs/records/22818773/", "external_url": "https://zenodo.org/records/22818773", "title": "AMEP-1 v0.3.0: Academic Software Release Monograph", "content_text": "AMEP-1 v0.3.0 advances Adaptive Maritime Estimation and PNT from the bounded seven-state research architecture reported in AMEP-1 Version 1.0 toward a dependency-aware software architecture for resilient maritime navigation, navigation-integrity supervision, and autonomy-authority management under GNSS-degraded and GNSS-denied conditions. The principal contribution of this release is not a claim of operational navigation performance, but a formalization of the software mechanisms required to make degraded-navigation assumptions explicit, executable, traceable, and falsifiable. The architecture separates estimation from integrity and authority; represents source identity, timing, provenance, failure domains, and shared dependencies explicitly; requires opt-in safety credit; performs dependency-aware pre-fusion consistency assessment; and prevents estimator covariance or nominal state-rank coverage from being treated as sufficient evidence for resilient-navigation authority. The post-Version-1.0 software increment introduces 15-state local-level strapdown error-state Kalman filter reference implementations with position, velocity, attitude-error, accelerometer-bias, and gyro-bias states; compensated coning and sculling; WGS-84 curvature and transport terms; Earth- and transport-frame compensation; covariance propagation; lever-arm-aware observation models; and installation/calibration uncertainty contracts. It additionally implements dependency-derived single-source and common-cause fault hypotheses, explicit integrity-risk bookkeeping, solution-separation primitives, deterministic hypothesis-excluded replay, false-alert and missed-detection metrics, selected NMEA 0183 normalization, configuration-bound evidence generation, and navigation-authority gating. The release is bound to Git tag v0.3.0 and commit 349dd88098ae4bff5b2e185f699d1abdf8e6b048 . Final trusted GitHub Actions validation run 35267201569 executed on the dedicated amep-windows-01 runner and passed compilation, Ruff, mypy, Bandit, executable assurance-traceability verification, runtime dependency audit, CycloneDX SBOM generation, and coverage-gated testing. The final campaign produced 66 passing tests and 88.21% measured package coverage. A central methodological constraint is maintained throughout: software verification is not equivalent to navigation-integrity validation. AMEP-1 v0.3.0 therefore does not claim certified PNT integrity, validated protection levels, calibrated operational INS performance, target-hardware real-time guarantees, hardware-in-the-loop qualification, controlled-water validation, regulatory conformity, or operational deployment readiness. The separately published AMEP-1 Version 1.0 research report remains the frozen scientific baseline under DOI 10.5281/zenodo.22561851 . Its negative findings—including the absence of demonstrated clean-condition superiority and the correlated common-mode absolute-position bias failure that produced severe covariance overconfidence—are retained as part of the engineering evidence record rather than superseded by later software hardening. Software release represented by this monograph: AMEP software v0.3.0 Release tag: v0.3.0 Release commit: 349dd88098ae4bff5b2e185f699d1abdf8e6b048 Final CI: GitHub Actions run 35267201569 Public repository: https://github.com/GabrielAllit1/AMEP Original AMEP-1 Version 1.0 research DOI: 10.5281/zenodo.22561851", "date_published": "2026-09-17T00:00:00Z", "authors": [ { "name": "Allit, Gabriel" } ], "tags": [ "resilient PNT", "maritime navigation", "GNSS-denied navigation", "GNSS-degraded navigation", "uncrewed surface vessel", "USV", "multisensor fusion", "error-state Kalman filter", "ESKF", "strapdown inertial navigation", "fault detection and exclusion", "FDE", "integrity monitoring", "solution separation", "sensor dependency", "common-cause failure", "deterministic replay", "navigation authority", "software assurance", "software-in-the-loop", "research software" ] }, { "id": "10.5281/zenodo.22561851", "url": "https://salt19.com/arcs/records/22561851/", "external_url": "https://zenodo.org/records/22561851", "title": "Adaptive Maritime Estimation and PNT (AMEP-1): A Falsifiable Resilient Multisensor Navigation and Autonomy-Authority Architecture for GNSS-Degraded USV Operation", "content_text": "AMEP-1 Version 1.0 is a platform-neutral research architecture for resilient maritime PNT and navigation-authority supervision for uncrewed surface vessel research. It combines a verified WGS 84 vessel-centered local geometry implementation, a seven-state recursive covariance estimator, configurable innovation screening, sensor health/freshness supervision, local constraint-coverage diagnostics, experimental radar and bathymetric structural descriptors, residual spectral diagnostics, tail-error modeling, environmental route research, generalized communications supervision, and explicit operator/autonomy/safety command authority. The release is intentionally bounded by adversarial evidence. A same-sensor clean synthetic comparison does not establish AMEP accuracy superiority. Gross injected outliers support only a narrow synthetic screening claim. A correlated common-mode position-bias experiment produces severe covariance inconsistency, rejecting any certified-integrity interpretation of the current containment proxy. No recorded maritime replay, target-hardware timing campaign, HIL validation, controlled water trial, standards conformity assessment, or operational performance claim is included. The correct label is research software / SIL prototype.", "date_published": "2026-09-07T00:00:00Z", "authors": [ { "name": "Allit, Gabriel" } ], "tags": [ "Adaptive Maritime Estimation and PNT", "AMEP-1", "resilient PNT", "maritime navigation", "GNSS-denied navigation", "GNSS-degraded navigation", "uncrewed surface vessel", "USV autonomy", "multisensor navigation", "sensor fusion", "state estimation", "Kalman filtering", "navigation integrity", "sensor fault detection", "innovation gating", "observability", "navigation authority", "autonomy supervision", "radar localization", "marine radar", "bathymetric navigation", "terrain-aided navigation", "persistent homology", "topological data analysis", "extreme value theory", "model predictive control", "WGS 84", "azimuthal equidistant", "NMEA 0183", "software-in-the-loop", "SIL validation", "fault-tolerant autonomy", "maritime autonomous surface ships", "MASS", "reproducible research" ] }, { "id": "10.5281/zenodo.22610459", "url": "https://salt19.com/arcs/records/22610459/", "external_url": "https://zenodo.org/records/22610459", "title": "AGTP-1 Technical Corrigendum: Bijective 2D Implementation Specification", "content_text": "AGTP-1 Version 1.2 is a post-publication implementation corrigendum to the Version 1.1 technical supplement and is anchored to the normative Version 1.0 geometry. It restores the Version 1.0 Equal Earth central meridians (-94.5 degrees and +77.0 degrees), restores the exact published Atlantic/Arctic seam, and restores the south-polar Lambert Azimuthal Equal-Area Antarctic lobe. It also replaces the implementation requirement for a side-channel lobe identifier with a conventional two-dimensional projected coordinate (X, Y) . Fixed false-eastings place the lobe images in provably disjoint X intervals, so the inverse identifies its branch from X alone. The deposit includes explicit forward and inverse equations, a global injectivity argument subject to declared map edges and pole singularities, equation-derived graticules, a dependency-free reference implementation, deterministic conformance vectors, numerical validation against PROJ/pyproj, and regression tests. Finite-precision edge ownership is hardened with direct-radian seam evaluation and deterministic canonicalization of numerical representations of the seam, -60 degree boundary, and antimeridian. Version 1.2 supersedes v1.1 numerical implementation artifacts wherever they conflict with the Version 1.0 geometry. AGTP-1 remains independent research and is not EPSG-registered, peer-reviewed, standards-endorsed, or intended for navigation.", "date_published": "2026-09-07T00:00:00Z", "authors": [ { "name": "Allit, Gabriel" } ], "tags": [ "AGTP-1", "mathematical cartography", "geodesy", "map projection", "interrupted projection", "Equal Earth", "Lambert Azimuthal Equal-Area", "WGS 84", "bijection", "inverse transform", "reproducibility", "GIS implementation" ] }, { "id": "10.5281/zenodo.22554593", "url": "https://salt19.com/arcs/records/22554593/", "external_url": "https://zenodo.org/records/22554593", "title": "Allit Geoid Truth Projection (AGTP-1): Implementation-Grade Mathematical, Inverse, Differential-Distortion, and Reproducibility Supplement", "content_text": "This record is an implementation-grade technical supplement to the original Allit Geoid Truth Projection (AGTP-1) v1.0 publication, DOI 10.5281/zenodo.22385043. It closes specific implementation and reproducibility gaps identified during external GIS-oriented review without changing the underlying AGTP-1 local projection geometry. The supplement publishes explicit WGS 84 forward and inverse equations, numerical convergence rules, finite-precision seam semantics, a reversible packed 2D representation, an analytic Jacobian and equal-area proof, Gaussian first-fundamental-form and Tissot distortion derivations, a dependency-free reference implementation, automated 100-cycle round-trip tests, PROJ cross-validation, conformance vectors, and a reproducible land-weighted benchmark with bundled point inputs and source-data hashes. AGTP-1 remains an interrupted small-scale equal-area world-mapping research framework and is not a navigation or local-surveying projection.", "date_published": "2026-09-06T00:00:00Z", "authors": [ { "name": "Allit, Gabriel" } ], "tags": [ "Allit Geoid Truth Projection", "AGTP-1", "cartography", "map projection", "equal area", "interrupted projection", "WGS 84", "authalic latitude", "Equal Earth", "inverse projection", "Gaussian fundamental quantities", "Tissot indicatrix", "projection distortion", "reproducibility", "geodesy", "GIS" ] }, { "id": "10.5281/zenodo.22385043", "url": "https://salt19.com/arcs/records/22385043/", "external_url": "https://zenodo.org/records/22385043", "title": "Allit Geoid Truth Projection (AGTP-1): A Proposed Geoid-Aware, Interrupted Equal-Area Framework for World Mapping", "content_text": "The Allit Geoid Truth Projection (AGTP-1) is a proposed static, interrupted, composite equal-area framework for small-scale world mapping developed by Gabriel V. Allit . AGTP-1 is formally defined on WGS 84 and combines two independently recentered Equal Earth land-bearing lobes with a South-Pole-centered Lambert Azimuthal Equal-Area (LAEA) Antarctic lobe. The framework also defines a latitude-dependent Atlantic/Arctic seam, canonical lobe-aware coordinates, explicit interruption topology, and a strict separation between horizontal map geometry and geoid-referenced vertical information. The term geoid-aware is used in a precise geodetic sense. Horizontal geometry is referenced to WGS 84 and its authalic equivalent for the analytical equal-area formulation. EGM2008 geoid undulation is treated as optional vertical-reference metadata and does not deform horizontal x,y coordinates. This avoids conflating the geoid with a visibly distorted continental planform. The proposed contribution of AGTP-1 does not lie in claiming invention of Equal Earth, LAEA, interrupted projections, authalic latitude, Tissot analysis, WGS 84, or EGM2008. Those are established prior art. The potentially original contribution is the specific reproducible composition of the lobe topology, latitude-dependent seam rule, land-weighted central-meridian optimization, Antarctic polar-domain treatment, canonical (L, x, y) coordinate representation, geoid/ellipsoid separation rule, and accompanying reproducibility and falsification protocol. The accompanying white paper provides the formal mathematical specification, domain selector, projection equations, central-meridian optimization procedure, seam-clearance audit, Tissot-based distortion analysis, grid-sensitivity testing, direct-equation validation against PROJ, finite-difference Jacobian checks, forward/inverse closure tests, prior-art discussion, limitations, and falsification criteria. In the disclosed primary benchmark of 19,738 land samples , AGTP-1 preserves areal scale to numerical precision and reports a weighted geometric-mean Tissot semiaxis ratio of 1.3839 , a 95th-percentile ratio of 2.5216 , RMS log anisotropy of 0.4407 , and weighted mean angular distortion of 18.20° . Under that specific land-weighted benchmark, AGTP-1 produced lower sampled shape anisotropy than the uninterrupted Equal Earth baseline and the nonconformal comparison projections emphasized in the paper. These results are benchmark-specific experimental findings and do not establish universal cartographic superiority. AGTP-1 is intended for educational, analytical, thematic, experimental, and research cartography where relative area, explicit distortion accounting, and interrupted global representation are appropriate. It is not a navigation projection, not an EPSG-registered coordinate reference system, not conformal, not equidistant, and does not claim that any two-dimensional map can provide a distortion-free or literally “true” representation of Earth. Author: Gabriel V. Allit Affiliation: SALT19 Version: 1.1 DOI: 10.5281/zenodo.22385043", "date_published": "2026-09-05T00:00:00Z", "authors": [ { "name": "Allit, Gabriel" } ], "tags": [ "Allit Geoid Truth Projection", "AGTP", "cartography", "map projection", "equal-area projection", "interrupted projection", "world map", "geodesy", "geoid", "WGS 84", "EGM2008", "Equal Earth", "Tissot indicatrix", "projection distortion", "cartographic distortion", "thematic mapping", "cartographic design", "spatial representation", "SALT19", "ARCS" ] }, { "id": "10.5281/zenodo.22151216", "url": "https://salt19.com/arcs/records/22151216/", "external_url": "https://zenodo.org/records/22151216", "title": "From APIs to Agent Markets: Governed Capability Acquisition and Qualified Demand in a Machine-Payable Agent Utility Grid", "content_text": "Autonomous agents can invoke tools, but invocation alone does not define a market. Before software can economically acquire an external capability, it must discover a relevant operation, observe current price and requirements, determine whether payment is authorized, cross a payment boundary, receive an outcome, and preserve enough evidence to distinguish discovery from realized commercial activity. This preprint presents a capability-acquisition layer added to the SALT19 Agent Utility Grid after the earlier Governed Decision-Commerce Control Plane record (DOI 10.5281/zenodo.22102248). The new architecture introduces two free front-door capabilities ( find-salt19-tool and salt19-pricing ), twelve one-cent Fresh-State acquisition products, explicit qualified-demand telemetry, and a zero-spend public-market observer while preserving x402 v2 exact payments on Base mainnet with USDC and validated_external_revenue as the sole revenue authority. At the pinned evidence cutoff, source state was commit 39c5e0e2be09e57a59c6343a45468e3584e636a9 (2026-08-28T22:38:18Z). That exact revision completed the production deployment workflow successfully, including exact-revision validation, D1 migration checks, production Worker deployment, external-payability certification, live surface verification, comprehensive advertised production contracts, and zero-spend transaction readiness. Same-revision MCP Registry publication, Phase-1 live smoke, and registry public-identity synchronization also completed successfully. The paper explicitly distinguishes free selection, qualified discovery, payment capability, paid retry, settlement, and validated revenue. A crawler, registry probe, HTTP 402 challenge, free selection, or seller-run test is not treated as commercial success. Fresh-State source/fetch provenance is not represented as a signed Proof402 receipt unless the receipt finalizer actually creates one. The evidence is first-party production and source validation. This record does not claim product-market fit, generalized willingness-to-pay, repeat external buyers, optimal economic routing, external scientific replication, or a stable authoritative external tool count.", "date_published": "2026-08-29T00:00:00Z", "authors": [ { "name": "Allit, Gabriel" } ], "tags": [ "agent markets", "machine commerce", "autonomous agents", "Model Context Protocol", "MCP", "x402", "HTTP 402", "capability acquisition", "Fresh-State", "qualified demand", "machine-to-machine payments", "Base", "USDC", "Proof402", "agent infrastructure", "SALT19" ] }, { "id": "10.5281/zenodo.22127174", "url": "https://salt19.com/arcs/records/22127174/", "external_url": "https://zenodo.org/records/22127174", "title": "Beyond Persistence: Causal Cross-Process Cognitive Influence in a Governed Agent Architecture", "content_text": "Persistent state is often credited as memory or learning even when it never affects a later decision. This preprint introduces Persistent Cognitive Influence (PCI), an operational evaluation discipline that separates recording, durable persistence, fresh-process retrieval or reload, canonical cognitive consumption, attributable downstream influence, independent outcome verification, and comparative improvement. Applied to four EvoMind mechanisms at a pinned August 2026 evidence cut, the study reports: (1) grounded user-fact recall across a fresh process with role attribution, session/project isolation, and absent-fact abstention; (2) relational entity grounding that survives restart and resolves indirect run/node references to verified artifacts, including stale-mutation detection; (3) a cognitive performance self-model whose verified-outcome-derived aggregate bias moved from 0.30 to 0.405 and automatically reloaded as 0.405 in a fresh process; and (4) structural recovery reuse where a persisted ACTIVE principle caused treatment to make zero recovery resynthesis calls versus one in an isolated control, with both paths passing the same downstream verification/release gate. The evidence is first-party internal software validation. The paper does not claim unrestricted transfer, general semantic recall, aggregate task-performance improvement, external replication, or artificial general intelligence. The deposit includes the publication PDF and source, a supplementary evidence PDF and source, claim/evidence and reproducibility records, machine-readable results, provenance, licensing, and cryptographic hashes. Evidence cutoff: EvoMind repository state c2e4a219a7c1ff11a61e2639c112daa9008f17b6 , 2026-08-27T11:23:55Z. Rights: Copyright © 2026 Gabriel Allit. All rights reserved. See LICENSE.txt included with this research object.", "date_published": "2026-08-27T00:00:00Z", "authors": [ { "name": "Allit, Gabriel" } ], "tags": [ "persistent cognition", "AI agents", "long-term memory", "episodic memory", "self-modeling", "structural recovery", "experience reuse", "governed learning", "causal evidence", "artifact provenance", "agent architecture", "EvoMind", "Salt19" ] }, { "id": "10.5281/zenodo.22102248", "url": "https://salt19.com/arcs/records/22102248/", "external_url": "https://zenodo.org/records/22102248", "title": "The Agent Utility Grid: Governed Machine-Payable Decision Intelligence for Autonomous AI Agents", "content_text": "Autonomous AI systems increasingly require external capabilities that are not economically or operationally sensible to bundle into a single model or runtime. Model Context Protocol (MCP) standardizes capability discovery and invocation, while x402 standardizes HTTP-native payment negotiation for machine actors. Those protocols solve important interoperability problems, but they do not by themselves determine how a paid tool should bound its claims, preserve the evidence supporting a decision, distinguish a payment challenge from realized revenue, or adapt its commercial surface without granting an autonomous controller open-ended economic authority. This preprint presents the Agent Utility Grid, a production architecture that combines native MCP discovery, x402 v2 payment gating, Base-mainnet USDC settlement, bounded technical decision products, SHA-256 evidence receipts, transaction-linked commercial accounting, and a constrained seller-side commercial controller. The paper describes this composition as a Governed Decision-Commerce Control Plane (GDCCP) : an infrastructure layer in which discoverability, payment, execution, evidence, and commercial state remain coupled to one canonical authority path rather than diverging into independent trust surfaces. As of the 25 August 2026 evidence cutoff, the deployed service reports version 0.3.0, 16 implemented products, 11 active paid utilities, five demand-gated candidates, and a live controller whose most recently observed cycle completed over all 16 products with 32 observations, 16 decisions, and zero controller errors. The manuscript explicitly separates live endpoint evidence, repository-source evidence, operator-published release evidence, and proposed future work. It also states the receipt-system claim boundary: a matching SHA-256 digest establishes stored-evidence integrity, not independent truth or correctness. The central contribution is architectural rather than a priority claim. The paper does not claim invention of MCP, x402, stablecoin payments, paid MCP tools, agent marketplaces, or decision receipts as isolated primitives. It proposes a bounded seller-side systems pattern joining capability discovery, machine payment, evidence-bounded decision intelligence, receipt integrity, transaction-linked revenue admission, and constrained catalog adaptation. io.github.GabrielAllit1/salt19-agent-utility-grid", "date_published": "2026-08-25T00:00:00Z", "authors": [ { "name": "Allit, Gabriel" } ], "tags": [ "agentic commerce", "autonomous agents", "Model Context Protocol", "MCP", "x402", "HTTP 402", "machine-to-machine payments", "machine-payable APIs", "decision intelligence", "AI governance", "verification receipts", "evidence integrity", "USDC", "Base", "software agents", "commercial control plane", "governed autonomy" ] }, { "id": "10.5281/zenodo.21957527", "url": "https://salt19.com/arcs/records/21957527/", "external_url": "https://zenodo.org/records/21957527", "title": "From Intent to Verified Work Products: Governed Semantic Capability Composition in EvoMind", "content_text": "This working paper describes a systems pattern implemented in EvoMind for converting a high-level user objective into one or more real, verified work products without granting the planner direct execution authority. The pattern separates semantic planning, capability resolution, governance admission, execution, independent verification, and artifact-state admission. In the current implementation, a high-level objective can be classified as requiring no deliverable or one or more deliverables; a CapabilityGoalGraph expresses work-product requirements and dependencies; a registry-grounded capability broker selects implementations; canonical admission and governed execution control side effects; and verified artifact records preserve identity, versioning, provenance, and lineage. An internal live acceptance proof through the conversational handle_chat() path produced real Word, Excel, and PowerPoint artifacts from a single high-level goal while an ordinary informational question produced zero files. The frozen acceptance matrix passed 9/9, with a 252/252 surrounding regression set. These are internal engineering results, not third-party replication. The paper positions semantic capability composition as an architectural layer between intent and tool execution, and argues that trustworthy computer-use systems benefit from separating 'what must become true' from 'which capability should act' and from admitting outcomes into trusted state only after verification. The paper does not claim universal desktop autonomy, formal correctness, or artificial general intelligence.", "date_published": "2026-08-15T00:00:00Z", "authors": [ { "name": "Allit, Gabriel" } ], "tags": [ "EvoMind", "Salt19", "AGI", "AI", "Agentic AI", "cognitive architecture", "computer-use agents", "governed AI", "planning", "tool use", "outcome verification", "artifact provenance", "local-first AI", "persistent agents", "autonomous digital work", "human-in-the-loop", "Semantic Capability Composition" ] }, { "id": "10.5281/zenodo.21881379", "url": "https://salt19.com/arcs/records/21881379/", "external_url": "https://zenodo.org/records/21881379", "title": "Beyond Scale, Agents, and Embodiment: Governed Experience Compounding in EvoMind - An Evidence-Based Positioning of a Persistent Cognitive Runtime in the Emerging AGI Landscape", "content_text": "EvoMind is a local-first, governed cognitive runtime investigating whether deployment experience can become persistent, reusable, transferable operational capability. Evaluating EvoMind against the contemporary artificial-intelligence landscape requires more than comparing parameter counts, language benchmarks, multimodal inputs, computer use, persistent memory, or robotic dexterity. Modern frontier systems already combine increasingly capable reasoning, multimodality, computer interaction, tool execution, persistent agent memory, and, in robotics, embodied reasoning and motor control. This white paper therefore examines a narrower systems hypothesis: can a deployed cognitive architecture convert repeated success, successful recovery, and attributable correction into reusable operational capability while retaining evidence, governance, restart persistence, auditability, verification, and revocability? The paper positions EvoMind as an experimental cognitive architecture organized around persistent memory, world modeling, reasoning, planning, governed execution, verification, learning, skill lifecycle management, and evidence-backed capability promotion. It distinguishes runtime experience compounding from foundation-model training, conventional agent orchestration, and robotics; reviews currently available EvoMind evidence; identifies limitations that prevent a full AGI claim; and proposes a sealed longitudinal benchmark for testing causal experience-to-capability transfer. The principal claim is deliberately bounded. EvoMind is not presented as demonstrated AGI, ASI, consciousness, or as superior to frontier AI laboratories. It is presented as a research system investigating whether governed experience after deployment can become persistent and transferable competence. The proposed decisive experiment tests whether learned capability causes statistically significant improvement on strictly unseen related tasks, survives complete runtime restart, preserves prior competence, remains governable and revocable, and produces the expected performance loss under ablation. Status: White paper / preprint; not peer reviewed. Version: 1.0 DOI: 10.5281/zenodo. 21881379", "date_published": "2026-08-10T00:00:00Z", "authors": [ { "name": "Allit, Gabriel" } ], "tags": [ "Evomind", "cognitive architecture", "artificial general intelligence", "AGI", "experience compounding", "runtime learning", "persistent memory", "governed execution", "skill promotion", "transfer learning", "computer use", "AI safety", "Local first AI" ] }, { "id": "10.5281/zenodo.20580112", "url": "https://salt19.com/arcs/records/20580112/", "external_url": "https://zenodo.org/records/20580112", "title": "EvoMind: A Governed Cognitive Architecture for Persistent, Verifiable, Experience-Driven Agentic Autonomy", "content_text": "EvoMind is a local-first, governed cognitive architecture designed to separate cognitive capability, runtime authorization, execution, and verification into explicit system authorities rather than treating model output as direct permission to act. This technical note documents EvoMind at a frozen first-party engineering snapshot dated 10 August 2026, corresponding to repository main commit 628de798bbc4ab1d729b584cd06231a7d372cf1e . The documented system integrates persistent state and memory, multi-step planning, fused desktop perception, governed tool and desktop execution, canonical skill lifecycle management, fresh post-action verification, semantic human demonstration learning, restart reconstruction, revocation, provenance, causal evidence, and experience-driven skill formation from ordinary verified successful execution. The architecture implements a single durable skill-promotion authority in which discovery and evidence production do not independently grant executable authority. Candidate skills move through explicit evidence, shadow, approval, activation, and revocation states. Activated learned capabilities execute through existing governed runtime authorities rather than through unrestricted macro replay or an alternate execution engine. The paper reports evidence including a 99-file perception/desktop reachability audit; complete disposition of the original 51 dormant capability donors; governed semantic desktop teaching without persistent pointer coordinates or raw typed-character trajectories; restart-surviving learned capabilities; causal and regression evidence; bounded experience-to-skill compounding; real Windows desktop validation; reusable cognitive-operator extraction; and project-reported MiniWoB++ progression from 261/625 (41.76%) to 407/625 (65.12%). EvoMind is presented as an AGI-oriented cognitive architecture , not as established human-level AGI. This publication does not claim consciousness, sentience, human-level or superhuman general intelligence, unrestricted recursive self-improvement, autonomous source-code mutation, universal computer-use competence, distributed multi-node correctness, or independent third-party validation of the private implementation. The deposit includes the technical whitepaper, citation metadata, bibliography, evidence manifest, package manifest, cryptographic checksums, rights notice, version metadata, and archival supporting materials.", "date_published": "2026-08-10T00:00:00Z", "authors": [ { "name": "Allit, Gabriel" } ], "tags": [ "EvoMind", "cognitive architecture", "agentic AI", "governed autonomy", "experience-driven learning", "skill compounding", "continual learning", "persistent memory", "desktop agents", "computer use", "AI governance", "provenance", "causal evidence", "post-action verification", "AGI evaluation", "local-first AI", "metacognition", "autonomous reasoning" ] }, { "id": "10.5281/zenodo.21270654", "url": "https://salt19.com/arcs/records/21270654/", "external_url": "https://zenodo.org/records/21270654", "title": "Verified Semantic Transitions (VST): Belief-Gated State Change and Repair-Invariant Memory for Reliable Long-Horizon AI Agents", "content_text": "Tool-using AI agents increasingly operate in asynchronous, partially observable environments where a successful API return, mouse click, keystroke sequence, or tool invocation does not guarantee that the intended semantic outcome occurred. This white paper proposes Verified Semantic Transitions (VST), an implementation-neutral architectural primitive that treats the unit of reliable agent progress as a verified intended state change rather than an attempted action. A VST binds an intent, prior belief state, predicted postcondition, action, resulting observations, verification predicate, recovery policy, and memory update. The central rule is that external action completion must not automatically become an internal fact: an agent should advance its plan or promote a proposition into working belief only after sufficient evidence establishes semantic closure. The paper formalizes this rule, introduces a three-clock model of action, environment, and cognition timing, defines a taxonomy of transition failures, and proposes repair-invariant memory as a mechanism for storing reusable corrections rather than only raw trajectories. VST is not presented as the invention of transactions, feedback control, Hoare-style postconditions, belief-state estimation, or agent verification. Its contribution is integrative and architectural: it unifies these ideas at the transition level for stochastic agent runtimes and couples verification failure to reusable repair abstraction. The framework is positioned relative to ReAct, Reflexion, Voyager, ToolGate, Estimate-Verify-Update (EVU), SAVeR, Agent-BRACE, Lean4Agent, and Cordon. In particular, VST is complementary to task-level semantic transaction systems: task transactions govern durable effects, whereas VST governs whether a specific semantic change may be committed into agent belief and control flow. The paper concludes with measurable metrics, a proposed ablation protocol, and an implementation-neutral schema suitable for desktop, browser, API, robotic, and enterprise agents.", "date_published": "2026-07-08T00:00:00Z", "authors": [ { "name": "Allit, Gabriel" } ], "tags": [ "Verified Semantic Transitions", "AI agents", "LLM agents", "agent reliability", "belief state", "postcondition verification", "long-horizon agents", "repair invariants", "transactional cognition", "cognitive runtime", "agentic AI", "dependable AI", "Gabriel Allit", "EvoMind", "EvoMind AGI", "Salt19", "Salt19 LLC" ] }, { "id": "10.5281/zenodo.20580153", "url": "https://salt19.com/arcs/records/20580153/", "external_url": "https://zenodo.org/records/20580153", "title": "EvoMind: A Governed Cognitive Architecture for Autonomous Reasoning, Planning, Memory, and Desktop Agents", "content_text": "EvoMind is a governed cognitive architecture and research platform for autonomous reasoning, memory, planning, learning, and desktop-agent execution. The architecture is organized around a closed-loop cognitive cycle consisting of perception, world-model maintenance, memory formation and retrieval, decision support, planning, governance enforcement, execution through the Raziel control layer, observation, and learning or adaptation. A primary design goal of EvoMind is the separation of cognition from language generation. The architecture emphasizes governance, auditability, provenance, reversibility, and bounded autonomy through explicit approval boundaries, policy enforcement, claim-boundary controls, and validation workflows. This release contains architecture documentation, governance documentation, validation methodology, benchmark-readiness mapping, citation metadata, public-safe diagrams, a technical whitepaper, and supporting publication artifacts. The package is intended as a research artifact and architectural reference rather than a claim of achieved artificial general intelligence. The included documentation explicitly distinguishes implemented capabilities, experimental subsystems, future-work areas, and unverified claims. Future releases will focus on external evaluation, reproducibility artifacts, benchmark evidence, and expanded deployment documentation.", "date_published": "2026-06-06T00:00:00Z", "authors": [ { "name": "Allit, Gabriel" } ], "tags": [ "EvoMind", "Cognitive Architecture", "Governed Ai", "Autonomous Agents", "Agentic System", "AI Safety", "Memory Systems", "Resoning Systems", "Desktop Agents", "Multi-Agent Systems", "Raziel", "Cognition Before Language Generation", "Autonomous Reasoning", "Validation Methodology", "Falsification Testing" ] } ] }