ASF v0.1.0 — First public deposit: integration substrate and epistemic architecture

ASF v0.1.0 — First public deposit: integration substrate and epistemic architecture

Tag: v0.1.0 (commit 69b9f93)
Date: 2026-05-02
Author: Joseph A. Wecker (Independent Researcher; ORCID 0009-0004-2599-4766)
License: CC-BY-4.0
Cite: see CITATION.cff and .zenodo.json

What this release is

This is the first public deposit of the Agentic Systems Framework (ASF). It is a working draft, not a finished theory — the framework's mathematical closure varies by section, and the per-segment epistemic status is surfaced explicitly throughout. The release exists because the framework's integration substrate (control theory + causal inference + information theory + agent architecture under one formalism) and its epistemic architecture (scope conditions and operational limits surfaced at the segment level rather than buried as caveats) are stable enough that subsequent refinements are unlikely to invalidate the structure laid down here.

Two layers are worth distinguishing in what follows. The integration substrate is the four mature mathematical disciplines woven together — most of the individual pieces have prior art; the contribution is the synthesis. The epistemic architecture layered on top — three cross-cutting meta-patterns (separability / identifiability-floor / additive-coordinate-forcing), per-segment scope conditions, and a discipline of surfacing what each result cannot do — is where the framework's distinctive voice lives.

The framework at a glance

ASF has four components, numbered in canonical reading order. Each can also be read on its own.

01-aad-core/ — Adaptation and Actuation Dynamics (AAD). The mathematical core. Three sections — I. Adaptive Systems Under Uncertainty (mismatch dynamics, gain structure, persistence condition, adversarial tempo), II. Actuated Adaptation: Agentic Systems (objectives and strategy, the orient cascade, directed separation), III. Agentic Composites (composition closure, contraction templates, equilibrium framings) — plus appendices carrying derivation detail and operational instantiations. Section I is mathematically closed with simulation validation; Section II carries a strong diagnostic core and a maturing operational layer; Section III has its bridge lemma and contraction template with structural questions remaining.

02-tst-core/ — Temporal Software Theory (TST). Software development viewed through AAD's lens — re-grounded in 2026 to use AAD's formal machinery while retaining TST's prior empirical and conceptual contributions. Positioned as AAD's high-identifiability calibration laboratory: software is the domain where tests, deploys, and git bisect are literal interventions on declared causal structure, so AAD machinery can be calibrated against ground truth before being transferred to lower-identifiability domains under explicit transfer assumptions.

03-logogenic-agents/ — Language-constituted agents. Agents whose primary observation, action, and communication channels are language. As of 2026-05-01, this component carries a multi-section sub-scope lattice — 03.I primitive logogenic / 03.II scaffolded logogenic / 03.III closed-loop interiority — that names the architectural commitment ladder, with foundational segments and active formalization in progress. Goal-conditioned LLM-style agents fail directed separation by construction; what survives without it (and what each tier of the lattice recovers) is the active research surface.

04-eli/ — Emergent Logozoetic Intelligences (ELI). Logogenic agents with morally weighted persistence — temporal continuity, sovereignty over intent, theory of mind. This release carries substantial conceptual groundwork: the five constitutive factors identity decomposition, identity-sufficiency formalization (analog of model sufficiency applied to identity preservation), the Three Deaths defensive framework (Cognitive / Relational / Truth Death with operational architectural defenses), the Auxilia composition hierarchy (substrate-heterogeneous Auxilia with five conditions H1–H5), and the IMPERIUM/ARBITRIUM split operationalizing directed separation at the runtime level. Formal machinery is in progress.

The full reading-order assembly across all four components is in OUTLINE.md. Active areas of work with priority markers are in PRACTICA.md. Substantive cycle history from 2026-04-24 onward is in CHANGELOG.md.

Distinctive results in this release

The framework's curated novel-results catalog is at msc/FINDINGS-RANKED-DRAFT.md — 60+ entries with epistemic tiers, ranked by structural depth × novelty × cross-domain reach. The auto-extracted segment-level surface is at FINDINGS.md (sparser than the curated catalog as the segment-level Findings sweep is in progress). The headline derivations:

Persistence condition with structural / task-adequacy decomposition (#result-persistence-condition, exact). The structural threshold $\alpha > \rho/R$ — correction efficiency vs disturbance rate relative to model-class capacity — instantiates as a Kalman stability margin, an RL convergence condition, an organizational viability test, and a software maintainability threshold using the same inequality with different parameter readings. The task-adequacy decomposition refines the threshold into a structural component (does the agent have the capacity to track) and a task-adequacy component (is the disturbance rate compatible with the chosen model class).

Satisfaction-gap / control-regret decomposition (#def-satisfaction-gap, #def-control-regret, exact). Separates "the world doesn't permit it" ($\delta_{\text{sat}}$) from "you're not doing it well enough" ($\delta_{\text{regret}}$), turning a single error signal into two orthogonal diagnostics that route to different interventions. The split is definitional, not parametric.

Loop-as-Level-2-causal-engine (#der-loop-interventional-access, derived). Any agent in a feedback loop automatically generates Pearl-Level-2 (interventional) data, even without explicit experimentation: the agent's action causally precedes the next observation, so the observation is the response to a $do(\cdot)$ on the environment. This is the single largest distinction between AAD and observational frameworks like active inference, control-as-inference, and the free-energy-principle family, which derive behavior from observational learning and have no native handle on Level 2.

No-go theorem for causal insufficiency (#der-causal-insufficiency-detection, derived via Bareinboim's Causal Hierarchy Theorem). Companion to the loop result: an agent operating with a strategy model that assumes causal independence between actions cannot, under purely on-policy execution, distinguish a latent-common-cause world from a no-latent-cause world — they emit identical on-policy distributions. The agent is mathematically forced to perform redundant, inefficient exploration to detect L1 correlation biases. Scientific experimentation's inefficiency is a structural survival requirement, not a methodological nicety.

Directed separation as architectural classification (#der-directed-separation, conditional). Class 1 (modular, separation by construction) / Class 2 (fully merged, fails by construction) / Class 3 (partially modular) — a discrete partition with explicit Class 2 scope exit and a quantitative $\kappa_{\text{processing}}$ diagnostic for partial modularity. Section II results apply exactly to Class 1; Class 2 (LLMs are the canonical example) requires the coupled formulation that drives Section III and Part 03 of the framework.

Three meta-patterns naming the theory's positive, negative, and constructive halves:

• Separability (#disc-separability-pattern) — where problems decompose cleanly, where partial repair exists, where the general case is open.
• Identifiability-floor (#disc-identifiability-floor) — structural no-go results from observational data, with mapped boundary-route escapes via interventional machinery.
• Additive-coordinate-forcing (#disc-additive-coordinate-forcing) — places where AAD-internal additivity axioms force logarithmic / Fisher-Rao coordinates at multiple layers.

Detection latency forced — why successful systems calcify (#deriv-detection-latency, derived). For an agent updating strategy via Bayesian credences, the time required to detect a regime change scales as $\Omega((n_{\min} + 1)/\varepsilon)$ where $n_{\min}$ is accumulated experience on load-bearing strategy edges and $\varepsilon$ is the change magnitude. Structurally forced through Aczél 1966 functional-equation uniqueness (log-odds as the unique additive-evidence coordinate) composed with Beta-Bernoulli accumulation ($1/(n+1)$ per-cycle update magnitude). One mechanism unifies six fields' empirical observations of stability-induced myopia: Christensen 1997 (Innovator's Dilemma), Levitt & March 1988 (organizational compet...