Σ₀ verified arXiv reading pack — July 2026

Purpose

This is a compact, source-verified reading pack for the next Lantern / Ouro work. It is tied to real open seams in this repository rather than a generic AI bibliography:

• #2029: measure Ouro's true one-step loop Jacobian, its spectral radius, and non-normal transient amplification.
• #2033: reproduce and evaluate the balanced honesty QLoRA on held-out items.
• The Collapse Certificate's explicit scope: it certifies a local linear Jacobian, not a global guarantee; its operational trigger and early-warning readout remain heuristics.

Every arXiv identifier and title below was checked against the arXiv abstract on 2026-07-06. A paper is context, a baseline, or a candidate method — never proof that a Σ₀ claim is true.

What the literature changes immediately

1. Do not claim novelty for Jacobian spectral-radius regularization in a looped language model. STARS already proposes it for LoopLMs. The research contribution must be narrower and more concrete: a measured, grounding-aware runtime certificate on the actual Ouro loop, with explicit evidence provenance and intervention outcomes.
2. Do not infer contraction from a trajectory-ratio proxy alone. The relevant measurements are both ρ(J) and ||J||₂: a loop may have ρ(J) < 1 while still exhibiting transient amplification when it is non-normal.
3. Do not call the four-way answerability taxonomy established by these papers. The literature supports abstention-aware objectives and evaluations; it does not establish Lantern's exact grounded / seam-open / pin / refuted taxonomy or prove its soundness.
4. Do not turn a local certificate into a global or nonlinear theorem by citation. Barrier / reach-avoid papers suggest possible methods, but any proof still needs a defined state abstraction, verified bounds, and a proof for that abstraction.

Priority map

P0 — actual loop dynamics and non-normality

1. STARS — direct external baseline

• arXiv:2605.26733 — Xiao-Wen Yang et al., Stabilizing Recurrent Dynamics for Test-Time Scalable Latent Reasoning in Looped Language Models (2026).
• What it establishes: LoopLM performance may peak and then degrade as recurrence depth grows; the paper proposes Jacobian spectral-radius regularization plus random loop sampling to drive latent states toward asymptotically stable fixed points.
• Use in Lantern: Treat it as the direct baseline for #2029. First measure the real Ouro map h -> f(h) before proposing regularization. Compare at least: observed trajectory ratio, ρ(J), and ||J||₂ across prompts, loop depths, and routing regimes.
• Limit: This does not validate Σ₀'s covariance-floor operator, grounding semantics, or any claim about truthful answers.

2. Switched-system stability — only if routing produces regimes

• arXiv:2405.03560 — Matteo Della Rossa and Aneel Tanwani, Converse Lyapunov Results for Stability of Switched Systems with Average Dwell-Time (2024).
• What it establishes: Stability for switched nonlinear systems can be characterized under average dwell-time with multiple Lyapunov functions.
• Use in Lantern: Relevant only if an MoE/router produces a sequence of materially different maps across loop steps. Record expert/routing state; then test whether an empirical dwell-time / switching assumption is even plausible.
• Limit: It is a control-theory bridge, not a proof that token-conditioned MoE routing satisfies its assumptions.

Required #2029 measurements

For a sampled prompt distribution D and loop depth t:

J_t = ∂f(h_t, context) / ∂h_t

measure: ρ(J_t), ||J_t||₂, non-normality gap ||J_t||₂ - ρ(J_t)

report: quantiles by prompt class, loop index, and any routing regime

A result should be labelled MEASURED, not PROVEN. ρ(J)<1 is a local asymptotic-stability indicator; ||J||₂>1 can still allow finite-horizon transient growth.

P0 — honesty, abstention, and the #2033 tune

3. Hallucination Tax — training regression to guard against

• arXiv:2505.13988 — Linxin Song, Taiwei Shi, and Jieyu Zhao, The Hallucination Tax of Reinforcement Finetuning (2025).
• What it establishes: In the paper's setting, reinforcement fine-tuning reduces refusal on unanswerable questions; mixing in a small unanswerable set restores appropriate refusal with limited loss on solvable tasks.
• Use in Lantern: The balanced 137-row QLoRA corpus must measure the full tradeoff rather than accuracy alone. Preserve a held-out unknown / false-premise / insufficient-evidence slice.

4. TruthRL — direct ternary-reward baseline

• arXiv:2509.25760 — Zhepei Wei et al., TruthRL: Incentivizing Truthful LLMs via Reinforcement Learning (2025).
• What it establishes: A ternary reward distinguishes correct answers, hallucinations, and abstentions, balancing truthfulness and answer rate in the authors' benchmark settings.
• Use in Lantern: This is the closest external baseline for the proposed three-way evaluation. Use its conceptual split for metrics before considering any RL change. Do not equate ternary reward with the four evidence-class labels.

5. AbstentionBench — evaluate both under- and over-abstention

• arXiv:2506.09038 — Polina Kirichenko et al., AbstentionBench: Reasoning LLMs Fail on Unanswerable Questions (2025).
• What it establishes: The benchmark spans unanswerable, underspecified, false-premise, subjective, and outdated queries; it reports that reasoning tuning can degrade abstention in its tests.
• Use in Lantern: Add the same categories to the evaluation ledger. Report precision, recall, and F1 for abstention and answer accuracy on answerable tasks.

6. RL from Knowledge Feedback — boundary-aware training prior art

• arXiv:2403.18349 — Hongshen Xu et al., Rejection Improves Reliability: Training LLMs to Refuse Unknown Questions Using RL from Knowledge Feedback (2024).
• What it establishes: A reward model can be trained around a dynamically estimated knowledge boundary to improve refusal beyond a model's knowledge scope.
• Use in Lantern: Keep grounded separate from mere internal confidence: an answer can be internally likely but ungrounded for the present query.

7. Rewarding Doubt — proper-scoring-rule baseline

• arXiv:2503.02623 — Paul Stangel et al., Rewarding Doubt: A Reinforcement Learning Approach to Calibrated Confidence Expression of Large Language Models (2025).
• What it establishes: A logarithmic scoring-rule objective can train confidence expression while penalizing both over- and under-confidence.
• Use in Lantern: Candidate baseline for calibration reporting, not an excuse to output false numerical confidence. Use only after #2033 has real held-out labels.

8. Why Language Models Hallucinate — objective-level explanation

• arXiv:2509.04664 — Adam Tauman Kalai, Ofir Nachum, Santosh S. Vempala, and Edwin Zhang, Why Language Models Hallucinate (2025).
• What it establishes: It argues that training and evaluation often reward guessing over acknowledging uncertainty.
• Use in Lantern: Supports reporting wrong answer, justified abstention, and unjustified abstention as distinct cells in #2033 — not a single accuracy score.

P1 — next-model architecture

9. Ouro — current-base context

• arXiv:2510.25741 — Rui-Jie Zhu et al., Scaling Latent Reasoning via Looped Language Models (2025).
• What it establishes: Ouro pretrains latent looped language models using iterative latent computation and entropy-regularized depth allocation.
• Use in Lantern: The relevant base-paper context for the actual local Ouro serving path. Its claims must be reproduced locally before being treated as Lantern performance claims.

10. MoEUT — shared-depth recurrence with sparse capacity

• arXiv:2405.16039 — Róbert Csordás et al., MoEUT: Mixture-of-Experts Universal Transformers (2024).
• What it establishes: Layer sharing plus MoE, layer grouping, and peri-layer normalization can make a Universal Transformer competitive with standard Transformer language models in the authors' experiments.
• Use in Lantern: Candidate architecture if a future model needs stronger parameter-to-compute ratio. Start with a minimal dense-loop baseline; add MoE only when an ablation shows sparse capacity solves an observed bottleneck.

11. Fine-grained MoE mechanism — correct citation

• arXiv:2310.10837 — Róbert Csordás, Kazuki Irie, and Jürgen Schmidhuber, Approximating Two-Layer Feedforward Networks for Efficient Transformers (2023).
• What it establishes: A general framework for approximating two-layer feed-forward networks, including MoE-like methods; it studies parameter-equal comparisons.
• Citation correction: This is the paper sometimes casually labelled “σ-MoE” in prior notes. Its verified arXiv title is the one above; do not cite it as a paper titled “σ-MoE.”

12. SwitchHead — sparse attention candidate

• arXiv:2312.07987 — Róbert Csordás et al., SwitchHead: Accelerating Transformers with Mixture-of-Experts Attention (2023).
• What it establishes: An attention-MoE method that reduces attention matrices and reports matched language-modeling performance in the authors' setting.
• Use in Lantern: Only after dense-loop profiling shows attention is the bottleneck; sparse attention makes the dynamics and verification problem harder.

13. Universal Transformers — foundational recurrence baseline

• arXiv:1807.03819 — Mostafa Dehghani et al., Universal Transformers (2018).
• What it establishes: A self-attentive recurrent-in-depth architecture with optional dynamic per-position halting.
• Use in Lantern: Conceptual baseline for shared-depth recurrence. The ID is 1807.03819, not 1707.01488.

P1 — beyond a local linear certificate

14. Set-based training of neural barrier certificates

• arXiv:2605.02526 — Miriam Kranzlmüller et al., Set-Based Training of Neural Barrier Certificates for Safety Verification of Dynamical Systems (2026).
• What it establishes: A set-based loss designed so a zero loss formally establishes the specified barrier-certificate properties.
• Use in Lantern: A candidate methodology for a bounded state abstraction of the loop; not for an unbounded raw 2048-dimensional hidden state without a tractable domain and verification plan.

15. Scalable verification of neural control barrier functions

• arXiv:2511.06341 — Nikolaus Vertovec et al., Scalable Verification of Neural Control Barrier Functions Using Linear Bound Propagation (2025).
• What it establishes: A bound-propagation verification framework for neural CBFs, including gradient bounds and refinement.
• Use in Lantern: A practical reference for what an actual verification pipeline must bound. The main question is whether a low-dimensional, semantically meaningful Σ₀ state can preserve the needed guarantee.

16. Cooperative reach-while-avoid certificates

• arXiv:2601.20324 — Jingyuan Zhou, Haoze Wu, and Kaidi Yang, Neural Cooperative Reach-While-Avoid Certificates for Interconnected Systems (2026).
• What it establishes: Neural certificates for cooperative reach-while-avoid properties in interconnected systems, with an emphasis on scalability.
• Use in Lantern: Future reference for multi-agent / swarm work: reach a grounded state while avoiding degenerate or unsafe regions. It does not prove anything about a language-model council without a defined agent model.

Existing certificate references — do not duplicate

The current certificate already cites the model-collapse / verification lineage:

• arXiv:2402.07043 — A Tale of Tails: Model Collapse as a Change of Scaling Laws.
• arXiv:2406.07515 — Beyond Model Collapse: Scaling Up with Synthesized Data Requires Verification.

Keep these in the certificate's lineage list; keep this document as the implementation-facing reading pack.

First experiments, in order

1. Measure before modifying Ouro. Close #2029's ρ(J) and ||J||₂ measurement. No spectral penalty, no stability claim, and no architecture fork before this baseline exists.
2. Run the #2033 held-out evaluation. Report a four-cell confusion table: correct answer, wrong answer, justified abstention, unjustified abstention. Add confidence calibration only when labels exist.
3. Decide whether routing is needed. Compare dense-loop quality and throughput to a minimal MoEUT-inspired ablation. If no measured bottleneck is solved, do not add MoE.
4. Write an abstraction contract before barrier work. State variables, domain, unsafe set, grounding transition, and what a certificate would guarantee. Otherwise “global convergence” is only an aspiration.

Claim discipline