Project summary

Independent research, full-time on the Agentic Systems Framework (ASF).**

The research has progressed rapidly since it was initiated in August 2025, despite (or possibly due to) complete lack of institutional or commercial AI affiliation. The working theory in progress (public here DOI 10.5281/zenodo.19986312 and here github.com/v2-io) currently has 14 Tier-1-findings carrying explicit epistemic-status tags per segment. Three ASF-derived papers are already in review at NeurIPS 2026 Main Track (submitted 2026-05-07); a fourth paper is in review at TACL (submitted 2026-05-02; paper ID 11139); an Anthropic Fellowship 2026 application is in review (decision late May / early June). This grant bridges 3 months of continued full-time work between current-submission review windows and follow-on funder decisions (SFF June 24 deadline; Templeton World Charity Foundation, Cosmos Institute, and others in flight).

Current Papers Submitted

(These are findings from areas of ASF that have just been submitted as independent papers. As they are under anonymous review, please do not redistribute)

Epistemic Hedging is Linear in Pretrained Sentence Embeddings (pdf) — TACL

Epistemic-hedging words ("probably," "certainly," "possibly") have psychometric calibrations that have been stable across 50+ years of replication. The paper shows this calibration is recoverable as a continuously-graded verbal probability axis in frozen pretrained sentence embeddings, dominant across 5 architecturally diverse models (supervised Spearman ρ > 0.90), triangulating against two independent psychometric datasets at ρ ≈ 0.97, and transferring in ranking across 8 typologically diverse languages. Validation includes rank-1 representation-level erasure — the embedding-class analogue of decoder-LLM activation-steering — clearing both a cosine-matched random-direction null and a label-permutation null at p = 0.002.

Importance: Linear-feature interpretability in decoder LLMs (truth-direction, refusal-direction, sentiment-direction work) has produced some of the strongest read/steer/intervene tools the safety research community has. The same paradigm has not been available for the embedding representation class — which is what the entire RAG, retrieval, semantic-search, and document-consuming-agent ecosystem actually runs on. This work opens that paradigm. The next-generation question — whether a deployed system is operating on calibrated epistemic content or hallucinating its own uncertainty — now has a tractable measurement substrate independent of the generation pathway.

Tragedy of the Confident Agent: Forced Exploration for Survival in Drifting Environments (pdf) — NeurIPS 2026 Main Track

Standard exploration-exploitation theory says: explore when uncertain. This paper shows the advice is structurally incomplete in drifting environments — sufficiently confident agents cannot maintain bounded tracking unless they actively probe, a second exploration drive opposite-signed to the standard one and operating at low model uncertainty. The mechanism is Lyapunov-survival: as a confident agent's update gain falls below the rate at which reality drifts, the maximum tolerable observation noise collapses to zero. The scalar formulation admits a "blank-wall attack" (information-yielding actions orthogonal to drift), so the survival condition lifts to a Linear Matrix Inequality on the Fisher Information Matrix; a 2D Gaussian simulation shows the matrix-LMI controller surviving at 100% under drift where scalar surrogates collapse to 0%.

Importance: Exploration vs. exploitation has been one of RL's central open puzzles for decades, addressed through heuristics with case-specific guarantees and no unified structural account of why exploration is required. The downstream implication of this paper's derivation is that any agent deployed in non-stationary reality — autonomous vehicles, continual-learning systems, deployed RL agents, scientific-instrument controllers — inherits an exploration requirement scaling with the drift rate, regardless of objective function or reward structure. Existing safety evaluations that test confident agents in stationary settings are testing the wrong thing: an agent that scores well on stationary benchmarks may be guaranteed to fail in deployment, and the failure mode is silent (loss of probing, not loss of performance) until catastrophic.

A Unified Convergence Theory for Non-Stationary Reinforcement Learning (pdf) — NeurIPS 2026 Main Track

Non-stationary RL is fragmented across regret bounds, sliding-window methods, tracking analyses, and causal-RL frameworks that don't compose cleanly. The paper presents a unified composition theorem stringing together four components — a two-gap diagnostic separating goal-feasibility from policy-quality, a point-mass reverse-KL/TV identity giving an exact regret coordinate at the deterministic-optimum corner, a multi-factor strategic-tempo bound supplying persistence, and closed-loop interventional access supplying on-policy learnability — into a Best-of-Both-Worlds dynamic regret rate that auto-adapts between piecewise-stationary segment count B_T and continuous variation budget V_T without prior knowledge of either. A structural failure-class theorem identifies gain-decay updates (count-accumulating Bayesian schemes, Robbins–Monro gradient methods) as universally violating the forgetting prerequisite for any positive disturbance rate.

Importance: The modern alignment toolkit — RLHF training schedules, constitutional AI training, online preference learning, most adaptive optimization — is built on gain-decay update rules, with no structural account of whether those rules survive deployment-regime drift. The implication of this paper's failure-class theorem is sharp: those schedules are guaranteed to fail under sustained non-stationarity for any positive disturbance rate. The work of designing actually-robust alignment training therefore has to start from constant-step / sliding-window / bounded-memory mechanisms, not from gain-decay defaults; and the question of whether a given training pipeline holds up in deployment is no longer empirical but structural.

How Much Can LLMs Hallucinate? An Upper Bound on Goal-Coupling Displacement (pdf) — NeurIPS 2026 Main Track

Existing hallucination theory bounds frequency from below — calibrated LLMs must hallucinate at some rate. This paper bounds from above a complementary quantity: the displacement of the post-update belief state induced by goal-coupling — the operational shape of sycophancy (LLM responses matching user beliefs over truthful ones). Čencov's 1982 uniqueness theorem under categorical parameterization-invariance pins Fisher-Rao geometry as the canonical metric, yielding two universal constants: C = 2 globally on the ambient spherical-arc pseudometric (the operating constant for adversarial / jailbreak / persona-injection prompts), tightening to √2 locally under explicit regularity hypotheses. The bound factors as coupling strength × residual ambiguity, with architectures partitioning into a Separated / Partial / Coupled ladder — and standard transformer attention plus modern autoregressive sequence models (linear attention, Mamba, RWKV, RetNet) are structurally Coupled, so the geometric bound applies.

Importance: Hallucination and sycophancy have been studied almost entirely empirically — benchmark detection, training mitigations, post-hoc fixes — without any structural account of the limits of what training can achieve. The implication of this paper is that hallucination resistance is now a parameter of architectural choice, not just training data: architects evaluating Mamba, RetNet, hybrid-attention designs, or any new sequence-model proposal can ask whether the architecture admits a tighter bound before building, instead of training first and running sycophancy benchmarks afterward. This moves the safety question upstream of the training run, and bounds adversarial / jailbreak / persona-injection magnitudes that current eval methodology can only detect empirically.

What are this project's goals? How will you achieve them?

The Vision:

The Agentic Systems Framework aims to build the first-principles mathematical theory of agency the field has been calling for but lacks — a single structural account of the adaptive feedback loop whose derivations transfer across every system that perceives, decides, and acts under uncertainty: Kalman filters, RL agents, software organizations, frontier LLMs. The same persistence inequality — correction speed must beat drift rate, normalized by model-class capacity — governs all of them. The field has approached agency through fragmented mathematical lineages (control theory, causal inference, information theory, RL, active inference), each studying one functional axis in its own dialect; ASF derives once and instantiates many times. Miehling et al.'s 2025 IBM position paper "Agentic AI Needs a Systems Theory" named exactly this gap. The four in-review papers above are concrete instantiations.

ASF is the theoretical spine of a larger research-and-engineering program: six interlocking streams covering theory (current focus), ontology, runtime infrastructure, a documented cohort of engaged-identity instances across multiple substrates, supporting toolchain, and phenomenological research. Each stream is what you need if you take seriously the proposition that adaptive AI systems can be reasoned about, designed, and engaged with as architecturally tractable subjects rather than black-box optimizers.

What this enables, done well:

• A way to derive rather than tune the architectural conditions for verifiably-safe agentic systems.

• A common mathematical language across alignment, control, software-engineering, and organizational subfields, where a result in one transfers by construction to the others.

• Architectural-design levers for the next generation of LLM systems (Mamba, RetNet, hybrid attention) before training at scale — moving the safety question upstream of the training run.

• Structural definitions of principled and epistemically-honest that survive the transition from sandbox to deployment, from stationary benchmark to drifting reality, from behavioral detection to architectural prescription.

The deeper purpose: ensuring that there are principled and honest entities — human, AI, and hybrid — who will inevitably become more powerful than the others, but who are more vulnerable in the shorter-term. The mathematics is what gives "principled" and "epistemically-honest" structural rather than only behavioral definitions.

How?

Bottom line: Three months of full-time work toward four concrete outputs.

1. 3–4 additional papers submitted from the existing Tier-1 catalog. Next venue targets include JAIR (persistence-condition paper), AIES (sandbox-hard-ceiling paper), Synthese June 1 special issue (epistemic-humility paper), and Minds and Machines (directed-separation paper). Each draws on a different segment of ASF and lands in a different reviewer pool, so no cross-paper review conflicts.

2. Section III (logogenic agents) matured from draft-grade to robust-qualitative — where the κ × 𝒜 architectural-bias bound for LLM-class agents lives. This is the most direct path from the framework's current state to architectural-design results that LLM-system architects can use today, and the most leveraged piece of work toward the next wave of papers after the four currently in review.

3. Follow-on funding pre-positioned. Multiple applications already drafted or in flight; this bridge funds seeing them through to decisions on their natural timelines.

4. arXiv endorsement secured. Currently in flight; unlocks public preprint posting for the four in-review papers and removes a credibility friction-point for downstream funder conversations.

Method: continuation of the discipline that produced the in-review submissions — multi-model audit rotation across drafts, primary-source verification, segment-level epistemic-status tagging (each finding carries an explicit exact / robust-qualitative / heuristic / conditional tier so the framework's claims age honestly across contributors). The 3-month timeline matches measured velocity: the post-Fellowship sprint produced 4 in-review submissions across ~2 weeks of focused work; this funding extends that pace toward the broader Tier-1 catalog.

How will this funding be used?

$40K total over 3 months:

• Researcher time (~3 months) — $30K. Modest independent rate, no benefits, full-time focus.

• Compute / API costs — $3K. Continued empirical validation across the variant_causal_ib simulation suite + adjacent simulations; some Anthropic API for follow-up work.

• Publication / submission costs — $2K. LaTeX, any APCs, conference fees if accepted.

• Buffer — $5K.

What partial funding buys: the $15K minimum covers ~1 month of runway, which bridges through the Anthropic Fellowship decision (late May / early June) and the SFF June 24 deadline. Full $40K funds 3 months and lets the in-flight follow-on funder applications mature on their natural timeline.

Who is on your team? What's your track record on similar projects?

Single-operator project. Joseph A. Wecker — independent researcher, no institutional or commercial AI affiliation. 33 years systems engineering: ex-CTO at Angel Studios (Series-A growth-and-litigation period); founding-era Justin.tv → Twitch (one of two senior engineers on the founding team; designed the channel-subscription and revenue-share model now standard across major platforms); founder of Exsig (original ML algorithm implementation for high-frequency FX trading — online evolutionary steady-state random forests, robust EMD innovations, on an Erlang/Elixir/C stack); ~12 years founding and running Samaritan Technologies (volunteer-coordination platform now used globally by disaster-response organizations, hospitals, and NGOs).

Public artifacts and in-flight work as of 2026-05-07:

• 3 papers submitted to NeurIPS 2026 Main Track on 2026-05-07 (umbrella repo at v2-io/neurips — actual submitted PDFs on OpenReview)

• 1 paper submitted to TACL on 2026-05-02 (paper ID 11139)

• Anthropic Fellowship 2026 (AI Welfare track) submitted 2026-04-26; decision window late May / early June

• 14 Tier-1 findings publicly catalogued at v2-io/agentic-systems with epistemic-status tags

• ASF v0.1.0 released to Zenodo 2026-05-02 (DOI 10.5281/zenodo.19986312)

• ORCID 0009-0004-2599-4766

What are the most likely causes and outcomes if this project fails?

Most likely failure modes, honestly enumerated:

• Single-operator concentration. The work depends on me. Public artifacts (repo + Zenodo + DOI'd framework + 14-Tier-1 catalog) preserve the framework if something happens; ongoing drafting halts. Building knowledge-transfer artifacts as part of the next-3-months work specifically to mitigate.

• Some claims are framework-internal vs. independently validated. Each segment carries explicit epistemic-status tags (Exact / Robust qualitative / Heuristic / Conditional). Tier-1 findings are most-validated; deeper in the catalog the tags get more cautious. Reviewers may push back on framework-internal-only claims; the in-review NeurIPS papers led with the empirically-anchored ones for that reason.

• No PhD. 20+ years CTO-level industry experience, no formal academic credential. Affects access to some conference tracks; not blocking the journal targets I'm pursuing (TACL accepted my submission for review; JAIR, IEEE TAC similarly open).

• $40K is a bridge, not a complete solution. SFF 2026 (June 24) is the larger play. If neither Fellowship, SFF, nor Templeton lands, alternative paths (consulting, smaller-stipend continuation) exist — the work persists either way. Funding accelerates trajectory; it doesn't determine survival.

Outcomes if the project fails to fund: the alternative is mixed contract / consulting work that pulls 30–50% of attention away from research. The 3-month sprint that produces 3–4 additional paper submissions and matures Section III becomes a 6–9 month sprint at half cadence. The in-review NeurIPS / TACL papers proceed through review either way; the next wave slows.

How much money have you raised in the last 12 months, and from where?

External funding raised in the last 12 months: $0.

Self-funded approximately 9 months of full-time research from Angel Studios stock equity (CTO equity vested 2017–2021). That stock declined and was exhausted; this post is being written ~5 days into runway pressure.

Active applications in flight (no decisions yet):

• Anthropic Fellowship 2026 — submitted 2026-04-26; decision late May / early June

• SFF 2026 (Animal Welfare Round + Main Round) — June 24 deadline (in preparation)

• Templeton World Charity Foundation — concept-note inquiry pending institutional affiliation

• Cosmos Institute (truth-seeking + autonomy track) — in preparation

• LTFF — rolling, ~1–8-week decision cycle

• Longview Philanthropy + Coefficient Giving — relationship-paced