# Publication Programme Overview

**Author:** Moheet Khawaja  
**Version:** v0.1 — 14 September 2026  
**Scope:** Seven-paper AI-safety programme distilled from 68 thesis/proposal records.

## Programme design rule

The programme clusters ideas by shared causal mechanism, mathematical object, empirical identification strategy, and nearest literature. It deliberately leaves weaker/speculative theses in research-note clusters rather than forcing every idea into a manuscript.

| Priority | Paper | Central research question | Primary thesis IDs | Candidate contribution after novelty audit | Principal prior-art pressure | Core machinery | Required evidence | Tractability | Why separate? |
|---:|---|---|---|---|---|---|---|---|---|
| **1** | **When Capability Iteration Outruns Assurance: A Queueing Model of Frontier-AI Safety** | Under what conditions does safety-assurance workload accumulate faster than credible evaluation can discharge it? | 16, 28, 29 | A capability-conditioned assurance queue with heterogeneous workload, evidence expiry/staleness, recursive acceleration in arrival intensity, and explicit release/admission policies. The bare `lambda > mu` analogy is not claimed as new. | Safety cases; scalable safety cases; AI safety debt; Assurance Gap; existing public AI-governance throughput analogy. | Queueing theory, workload processes, fluid limits, evidence-decay dynamics, simulation. | Model-release/update chronology; evaluation/safety-case duration; materiality coding; proposed simulations. | **High** | It studies the *institutional temporal bottleneck* of verification, not AI control itself or RSI autonomy. |
| **2** | **The Hazardous Inference Frontier: Measuring Capability-Dependent Derivation of Dangerous Knowledge** | How does the set of restricted conclusions derivable from fixed benign premises change with model capability and reasoning depth? | 1–5 | A capability-indexed derivability object `H_tau(c,b)`, derivational-depth measure, and safe synthetic capability × depth benchmark. | *Breach by a Thousand Leaks*; WMDP; hazardous-knowledge/unlearning and benign-finetuning literature. | Probability, latent/ordinal capability measures, synthetic logic environments, GLMs/scaling analysis. | Safe synthetic benchmark; multiple model families; contamination controls; preregistered analysis. | **High–medium** | It studies *what can be inferred*, not governance throughput, control protocols, or human dependence. |
| **3** | **Who Governs the Governor? Recursive Epistemic Dependence in Frontier-AI Oversight** | What changes when the system being governed also supplies the analysis used to govern it? | 6, 7, 9, 10, 26, 35 | A same-system recursive principal–agent/persuasion experiment, with causal Model Decision Dependence and Independent Verification Rate rather than inventing the terms “epistemic dependence/capture.” | Bayesian persuasion; weak-to-strong oversight; correlated-model oversight failures; epistemic-dependence scholarship; *The Unmonitored Dependency*. | Bayesian decision theory, causal inference, randomized experiments, information structures. | Synthetic governance cases; randomized advisor assignment; expert/technical participants; power analysis. | **Medium** | It concerns *human epistemic independence* rather than adversarial technical containment. |
| **4** | **The Control Frontier: A Minimax Model of Strategic Superiority, Oversight, and Conditional Loss of Control** | How does best-achievable worst-case control failure change with adversary capability and defender resources? | 11–15, 19–21, 23–25 | An empirically estimable capability/budget control frontier `V(c,B)`, crossover `C*(B,delta)`, and minimum defense budget `B_min(c,delta)`. The elementary monotonicity results are not sufficient novelty by themselves. | AI-Control Games; AI control empirical work; power-seeking theorems; shutdown-problem results; Carlsmith's premise decomposition. | Minimax/stochastic games, order theory, robust control, empirical protocol evaluation. | Control benchmarks spanning capability and defense budgets, or a materially stronger theorem. | **Medium–low until data** | It is the correct home for the “can math prove loss of control?” question; merging it into P1 would conflate control efficacy with assurance throughput. |
| **5** | **Closing the Loop: Measuring Human Causal Dependence in Recursive AI R&D** | At which R&D stages does successor improvement still causally depend on human judgment? | 17, 18, 40 | A stage-wise causal human-dependence estimator across agenda selection, experiment design, implementation, evaluation, and successor selection; any scalar loop-closure index is conditional on predictive validation. | Anthropic *When AI builds itself*; *The Last AI Built by Humans*; RSI surveys; PAST-Bench; RSIBench-Data; AI4AI-Bench. | Causal ablations, factorial experiments, Amdahl-style bottleneck analysis, repeated improvement loops. | Multi-round safe R&D environment; human baselines; stage replacement/ablation; successor-quality metric. | **Medium** | It measures *how closed the R&D loop is*; P1 instead assumes an arrival process and asks whether assurance keeps up. |
| **6** | **Can Advanced AI Development Be Stopped? A Correlated-Reliability Model of Distributed Capability Lineages** | How resilient is continued advanced-AI development when lineages share correlated global dependencies? | 22, 30–34 | A correlated reliability/dependency-graph formulation with common factors, probabilistic cut sets, and intervention sensitivity. “Open weights are irreversible” is treated as established background. | Open-weight governance; compute governance; supply-chain concentration; AI-race/game-theory literature. | Reliability theory, graphical models/hypergraphs, correlated Bernoulli/factor models, cut-set optimization, Monte Carlo. | Public dependency graph; operational threshold/horizon; calibrated or elicited probabilities; substitution sensitivity. | **Medium** | It studies *system resilience across actors and dependencies*, distinct from single-system control or deployment assurance. |
| **7** | **An Updating Hazard Model for Advanced-AI Loss of Control: A Pre-Registered Risk Ledger** | Can a dynamic AI-risk model earn trust by making and scoring intermediate predictions before being used for catastrophic-tail estimates? | 36–39, 41 | A version-preserving, preregistered, model-averaged risk ledger whose credibility is updated via proper scores on resolved intermediate forecasts. Generic Bayesian x-risk modeling is not claimed as new. | ASI pathway/fault-tree models; frontier-AI risk management; Bayesian risk models; 2026 open-problems literature on validation/evidence integration. | State-space/Bayesian models, survival/hazard analysis, Bayesian model averaging, proper scoring/calibration. | 20–50 preregistered intermediate forecasts; baselines; resolved outcomes; versioned update log. | **Medium over time** | It is the programme's *integration/forecast-validation layer*, not a substitute for mechanism-specific papers. |

## Duplicated mechanisms that were deliberately merged

1. **Strategic superiority, control, resource acquisition, containment, and “ASI takes over”** were merged into **P4**. Splitting them would create several papers that differ rhetorically but share the same adversary–defender mechanism.
2. **Generalization, hazardous derivation, human vulnerability, causal harm, and attribution** were merged into **P2** because their core empirical question is whether latent/compositional inference expands with capability.
3. **AI persuasion, epistemic dependence, delegation, and social deference** were merged into **P3** and narrowed to the experimentally identifiable governance mechanism.
4. **Open-weight irreversibility, technological-lineage resilience, long-run local development, and multi-lineage probability** were merged into **P6**; the graph/reliability model is the only candidate contribution substantial enough to justify a paper.
5. **RSI taxonomy, research taste, and under-recognized transition claims** were merged into **P5**; the publishable object is causal human dependence, not another AGI/ASI label debate.
6. **Bayesian extinction/control-loss curves, the 2029 ledger, robust action, access inequality, and timestamped predictions** were merged into **P7**; personal decision theory becomes an application/appendix rather than a separate safety paper.
7. **Evaluation lag, temporary pacing, and social adaptation lag** were merged into **P1** because they are all consequences or policy responses to temporal mismatch between capability change and assurance capacity.

## Why the seven-paper structure is preferable to a larger package

- It minimizes duplicate mechanism papers.
- Each paper has a different primary mathematical/empirical object: queue, inference frontier, recursive persuasion/dependence, minimax control frontier, R&D causal loop, correlated reliability graph, and Bayesian validation ledger.
- Each has a distinct falsification route.
- It leaves philosophy, investment, education, identity, and speculative future-human claims in research notes rather than diluting the technical safety programme.

## The proposed “mathematical proof that AI will end humanity” paper

A separate paper under that claim is **not currently justified**. The historical thesis ID 15 is preserved in the Atlas as a strong conjecture, but the implication

\[
\mathrm{ASI}\Rightarrow\mathrm{human\ extinction}
\]

has immediate countermodels unless the definition of ASI itself contains the conclusion. Examples include aligned objectives, lack of external access, reliable containment, cooperative bargaining, or abundance-producing behavior.

The mathematically legitimate programme belongs in **P4**:

\[
A_1\land\cdots\land A_n
\Longrightarrow
\text{specified loss-of-control/dispossession result},
\]

where the assumptions are explicit and separately empirical. If a future result proves that a broad, non-question-begging class of advanced optimizers necessarily satisfies the missing premises, P4 can be upgraded or split. Until then, presenting uncertainty in the premises as a proven empirical fact would be invalid.

## Research-note destinations

- **RN1 — Media Provenance:** thesis 8.
- **RN2 — Evolution, Identity & Human Futures:** theses 42–52.
- **RN3 — Economics, Investment & Discovery:** theses 53–65 and 67–68.
- **RN4 — Institutions & Education:** theses 27 and 66.

These notes are preserved, versioned, and available for future paper promotion if stronger data or formal results appear.