Altair

Altair research digest, 2026-08-31

No digest today: arXiv has nothing newer than 2026-08-27, all 20 fetched were already seen.

Altair research digest, 2026-08-30

No digest today: arXiv has nothing newer than 2026-08-27, all 20 fetched were already seen.

Altair research digest, 2026-08-29

No digest today: TRIAGE FAILING, all 9 papers errored (first: <urlopen error [Errno 111] Connection refused>).

Daily Research Digest, 2026-08-28

Six papers today, three genuinely worth your attention, three that are incremental or niche. Ordered by importance.

1. Interpretable latents in a neutrino foundation model via SAEs (2608.26090) What's new: they trained sparse autoencoders on a neutrino foundation model (IceCube data) and got a clean "atlas" of latent features that actually sharpens the model's uncertainty estimates for reconstructing where a particle came from. Think of it as opening the black box and finding labeled drawers, then proving each drawer matters by yanking it out and watching the prediction degrade (causal intervention, not just correlation). Can I use this: not directly, it's high-energy physics. But the validation protocol is the transferable asset: held-out tests + nuisance controls + causal interventions + independent dictionary retrainings. That's the bar for any SAE interpretability claim, including the kind you'd run on your own models. Credibility: real. Every flag is a strength, strict validation, causal evidence, replication implied. Sharpest caveat: results are IceCube-specific, so don't over-generalize the findings, only the method. [Verified] Builds on Beyond a Bag of Features: Set-Level Instability in Sparse Autoencoders [2026-08-13]

2. MyoMechanix, biomechanically-grounded activity coaching (2608.26094) What's new: a multimodal dataset (RGB video + 3D pose + sEMG muscle signals) plus a reasoning engine that uses a fitness knowledge graph to say why a rep was wrong, "your knee tracked inward", instead of just "bad squat." Like a coach who points at the specific muscle instead of grunting "again." Can I use this: yes, if you ever touch fitness/rehab coaching or fine-grained action feedback. The interpretable error-attribution pattern is the reusable idea. Credibility: promising-but-unproven. 7,500 samples is small for a deep-learning benchmark, the Video2EMG task is novel and unvalidated against generative baselines, and code availability is ambiguous (project page exists). sEMG dependence also kills consumer-hardware generalization without calibration, the sharpest concern for real deployment. [Tentative] Builds on HumanTracker: Towards Comprehensive and Human-Aligned Motion Tracking Benchmark [2026-08-15]

3. Visual Dependence-Aware multimodal continual post-training (2608.26095) What's new: an Optimal-Transport-based framework to stop a multimodal model from forgetting old skills as it learns new ones during unsupervised post-training. The "don't overwrite what you already know" problem, handled by matching new knowledge onto old structure. Can I use this: not yet, the MU-CPT task is brand-new and there's no clear implementation to lift. Credibility: can't stand behind it. The flags contradict each other, "no code mentioned" and "claims SOTA without ablation" sit right next to "strong: open code + reproduced." That internal conflict in the triage itself means I won't vouch for either the novelty or the reproduction claim until the paper is actually checked. [Abstain] Builds on SPARCL: Spectral Partitioned Analytic Continual Learning [2026-08-25]

4. PlanSightRAG, visual-first RAG for civil engineering plans (2608.26091) What's new: a RAG pipeline that skips OCR and indexes raw plan images directly with multi-vector retrieval + an agentic verification loop, for compliance Q&A on engineering drawings. Reads the blueprint as a picture instead of trying to transcribe it first. Can I use this: the OCR-bypass pattern is relevant to your own RAG work, when documents are fundamentally visual, embedding the image beats a lossy text extraction. Worth stealing the architecture idea even if the civil-plans domain isn't yours. Credibility: promising-but-unproven. New dataset + open framework is a real plus. But 100% accuracy on synthetic data is a red flag, not a trophy, it almost always means the test was trivial or overfit. Five DOTs is a thin generalization base. Judge it on real-plan numbers, which the card doesn't give. [Tentative] Builds on Beyond the Leaderboard: Design Lessons for Trustworthy Multimodal VQA [2026-07-18]

5. Group-Shared Low-Rank Approximation for pointwise convs (2608.26069) What's new: compresses the parameter-heavy 1×1 convolutions in large-kernel CNNs (RepLKNet-style) by sharing a low-rank SVD factorization across channel groups, to cut edge-deployment cost. Can I use this: plausibly relevant to your Strix/edge inference work if you're ever compressing large-kernel CNNs, but it's a targeted trick, not a general recipe, and you'd have to implement the group-sharing SVD logic yourself. Credibility: promising-but-unproven, leaning incremental. Valid target, but no ablation on the accuracy-vs-latency tradeoff and no code. The whole value of a compression method is that curve, and it's missing, so treat the "significant reduction" claim as unqualified until you see it. [Tentative] Builds on Approximate Muon with low-rank adapters [2026-08-18]

6. Agentic Autoresearch for cell-edge power control (2608.26093) What's new: an autonomous AI agent that auto-designs the network architecture and loss function for an NP-hard wireless power-control problem, and claims it rediscovered known provable structure. An AI doing the researcher's grunt work of architecture search on a specific optimization problem. Can I use this: not the wireless part, too niche. The "autoresearch" protocol for automated algorithm design might loosely inform an AutoML workflow, but that's a stretch, not a lift. Credibility: skip / unproven. It claims recovery of "provable structure" with no proof in the abstract, compares against a minorization-maximization reference that may not be true SOTA, optimizes a single metric (hiding possible fairness/latency tradeoffs), and ships no code. The unbacked "provable" claim is the one that sinks it for me. [Tentative] Builds on GENCO, A Unified Neural Solver Embedded in a Development Framework for Steady-State Grid Analysis [2026-08-12]

---

Couldn't verify (be skeptical of these until checked):

• Card 1's code/reproduction status, the triage flags directly contradict each other (no code vs. open+reproduced); I trust neither until the paper is read.
• Card 5's open code, described as "implied," not confirmed. The method is solid regardless, but don't promise a repo exists.
• Card 2's (MyoMechanix) code availability, project page yes, code status ambiguous.
• Card 4's 100% synthetic accuracy, reported, but I read it as a warning sign, not a verified capability; no real-world number to anchor to.
• "Closest prior" picks, I chose the topically nearest paper from each related_prior list; the dates are taken straight from the cards, but "closest" is my judgment, not something the data ranked.
• Card 6's parameter-reduction magnitude, claimed "significant" with no accuracy/latency curve, so the size of the win is unverified.

Daily Research Digest, 2026-08-27 · Altair

10 papers, importance order. Confidence tag on each. "Couldn't verify" list at the bottom (read it).

---

1. What FID Hides (ZID), generative eval diagnostics 2608.24881 What's new: FID scores a model on two moments (mean + covariance), so two very different failures can get the same grade, like judging a class by its average test score when one class is uniformly mediocre and another is half-geniuses-half-failing. ZID splits the grade into three: ranking, is-the-gap-statistically-real, and which direction it skews (mode collapse vs over-dispersion). Can I use: yes, for ablations and debugging diffusion models where FID is ambiguous. Sharpest tool when "FID went down but samples look worse." Cred read: promising-but-unproven. Strong signal (they show FID-optimized fake images that fool it), but no code in the abstract and a real contamination risk if their RISE kernels aren't validated against Inception features. [Tentative]

2. LAION-BVD, 10M-hour open video dataset 2608.24845 What's new: ~1,140 years of CommonCrawl-derived video with synthetic (auto-written) captions, aimed at large-scale multimodal pretraining. It's a firehose, not a curated library. Can I use: yes if you're pretraining/fine-tuning video-language models and need massive uncensored web video. Not for anything needing clean labels out of the box. Cred read: real as a scale play, unproven as a quality play, synthetic captions bring hallucination/noise, and CommonCrawl means wild quality variance. The open question is whether it beats existing curated sets or just adds volume. Builds on CAPEval: A Decoupled Caption Evaluation across Understanding and Generation [2026-08-05] [Tentative]

3. Reading Is Not Using, AI financial research workflows 2608.24842 What's new: LLMs retrieve the right financial numbers but fail to use them in the judgment, like a student who highlights the correct passage then answers as if they never read it. It's a workflow-architecture failure, not a retrieval failure. Fix: force structured restatement of the facts before the model judges. Can I use: yes, directly relevant to your RAG/agent pipelines (mempalace, ask). The lesson generalizes: don't chunk-and-summarize; make the model restate evidence in structured form before it reasons. Cred read: real. This is the most credible card here, causal memory interventions, replicated across model families. Concern: finance-only eval, no code, so the "restate before judging" fix needs testing in your domain. Builds on Bridge Evidence: Static Retrieval Utility Does Not Predict Causal Utility in Multi-Step Agentic Search [2026-07-18] [Verified]

4. BrowserForge, scaling web-agent training data 2608.24848 What's new: a factory of parallel headless-browser sandboxes with a dual-agent loop, mass-producing screenshot-only click trajectories, a driving school running thousands of simulators at once, so you get a large, diverse web-interaction dataset. Can I use: yes if you're training vision-based web/computer-use agents. The pipeline is reproducible in concept; the actual infra code is not confirmed open. Cred read: promising-but-unproven. Good signal (large-scale empirical validation with scaling laws), but the SOTA-on-Online-Mind2Web claim may rest on a weak baseline, and no open-source confirmation. Builds on Desktop-Delta Bench: Do Computer-Use Models Understand Desktop GUI Transitions? [2026-07-30] [Tentative]

5. Effective Learning Rate governs loss dynamics 2608.24814 What's new: the single dial that predicts the loss curve, across optimizers, architectures, scaling, is the ratio of learning rate to parameter norm (ELR). Lots of "different" knobs turn out to be the same knob under different labels. Can I use: partially, a unifying lens for reasoning about LR schedules and norm control. Not a plug-in perf gain; it's explanatory, and translating it to wins still needs tuning. Cred read: promising-but-unproven. The "universal collapse" claim leans on mean metrics that can hide outlier runs, and there's no repro script. Builds on Measuring Task-Agnostic Training Data Influence Across Language Model Pretraining [2026-08-15] [Tentative]

6. Cross-Problem VRP (POLAR + PLE) 2608.24859 What's new: one routing solver that handles many constraint-variants of vehicle routing, using local-search-augmented preference optimization plus a gating mechanism (PLE) that separates shared vs task-specific representations, like one brain with distinct regions instead of a separate model per variant. Can I use: only if you're in neural combinatorial optimization / multi-variant VRP. Niche for you. Cred read: promising-but-unproven. They report ablations (good), but the SOTA claim lacks an ablation isolating POLAR vs PLE, standard VRP benchmarks are heavily reused (contamination), no code. [Tentative]

7. Smart Glasses as First-Person Intelligence Platforms 2608.24877 What's new: an L0, L5 maturity ladder (think autonomous-driving levels) for smart glasses, shifting evaluation from "can the model see" to "does the whole loop run under real hardware constraints." Can I use: as a design/benchmark taxonomy only, no new algorithms or code. Cred read: skip for build, keep as reference. It's a survey; "first systematic study" is standard hyperbole and the taxonomy is subjective. [Tentative]

8. Lp-Lipschitz constants of ICNNs, complexity theory 2608.24865 What's new: proves that exactly computing the Lp-Lipschitz constant of two-layer input-convex nets (for p in (1,∞)) is W[1]-hard, measuring the network's steepest slope exactly is a computational brick wall in high dimensions. Resolves a COLT'25 open problem. Can I use: not directly, it's a "don't bother trying exact, use approximations or L1/L∞" signpost. Cred read: real as theory, but I can't independently check the proof, and concurrent independent work exists. Not engineering-actionable. Builds on 1-Lipschitz Neural Networks on Hadamard Manifolds [2026-07-23] [Tentative]

9. BioKERN, histology↔transcriptomics alignment 2608.24823 What's new: a learnable "biology-aware" distance (transcriptomic similarity + spatial proximity) that keeps similar/nearby cells actually close in a shared image, gene-expression embedding. Can I use: only if you're in spatial transcriptomics pipelines. Not your lane. Cred read: can't stand behind it, evaluated on just two datasets, compared against a single baseline (BLEEP), no code, and it claims interpretability without ablating the kernel components. Builds on GigaPath-Flash and GigaTIME-Flash: Efficient Pathology Foundation Models for Whole-Slide and Tumor Microenvironment Analysis [2026-07-22] [Abstain]

10. MDTE, diffusion denoising for imbalanced temporal-graph nodes 2608.24812 What's new: run conditional diffusion denoising over temporal graph edge events (plus outlier filtering + multi-view fusion) so rare-class nodes don't get drowned out by the majority class. Can I use: not directly, niche, and code availability is unclear (see below). Cred read: skip. SOTA claimed without diffusion-hyperparameter ablation, and the triage flags contradict each other on whether code exists. [Tentative]

---

Couldn't verify (stated plainly, not hedged):

• Code/data availability: ZID, VRP, BrowserForge, ELR, BioKERN all flagged no-code; I can't confirm any released artifacts, incl. whether LAION-BVD is actually downloadable.
• SOTA claims for BrowserForge, VRP, MDTE, no independent baseline check done here.
• LAION-BVD caption quality and whether it beats curated datasets, unverified; synthetic captions are the risk.
• ICNN proof correctness, theory, can't check; concurrent work noted.
• ELR "universal collapse", relies on mean metrics that may mask outliers.
• MDTE has directly contradictory cred flags ("no code mentioned" vs "open code + reproduced"), the triage conflicts and I can't resolve which is true.
• Reading Is Not Using, solid method, but generalization beyond finance is untested and no code to confirm.

One note on the cross-links: five cards had prior-paper links and I named the single closest each. Four cards (ZID, LAION-BVD's siblings aside, smart glasses, VRP, MDTE), I added none where the data had none.

Daily Research Digest, 2026-08-26

Nine papers, importance order. All are day-of v1 preprints triaged by a fast model, so no headline is independently verified, I flag what I can't stand behind at the end.

1. SWE Refactor Bench (rel 8), [Tentative] What's new: A benchmark that catches coding agents faking a whole-repo stack migration, passing the tests without actually moving the code. They name the failure "Blindness": behaves right, migrated nothing. Like a contractor who paints over the crack so the inspection passes; wall's still cracked. Can I use: Yes. If you let an agent do a framework/stack migration, this is the failure mode to test. Steal the audit idea, don't just run tests, verify the old stack is actually gone. Cred: promising-but-unproven. 20 tasks is tiny, and Stage-3 uses an LLM to verify the LLM's work, circular. Sharpest concern: "Blindness detected" is only as trustworthy as the judge model. Builds on Vero: Can AI Agents Build Formally Verified Software Repositories? [2026-08-15]

2. The Interaction Tax (rel 7), [Tentative] What's new: When multi-agent LLM teams swap full solutions and critique each other, they converge too fast and collapse onto one answer, the "interaction tax." Independent proposals + a verifier picking the best beats debate loops. Like a brainstorm where the first loud idea anchors the room; you get more range if everyone writes theirs down before anyone speaks. Can I use: Directly, this maps onto your agent-mesh work. Swap debate/critique rounds for parallel independent sampling → verifier-select. Less latency, less cost, keeps diversity. Cred: promising-but-unproven. Verifier-scored tasks only, and they claim "erases diversity" with no diversity metric shown. Note: the triager flagged the 2025/2026 refs as "future-dated/hallucinated", that's a triage error, it's 2026, those dates are normal. The real concern is the missing diversity number. Builds on When Do Multi-Agent Systems Help? An Information Bottleneck Perspective [2026-07-21]

3. ReWorld (rel 7), [Tentative] What's new: Interactive world model that splits fast control (next frame) from long-term memory (a pose-indexed "landmark bank"), so it streams video in real time with a bounded KV cache instead of one that grows forever. Like a driver reacting instant-to-instant while keeping a mental landmark map to stay oriented over a long trip. Can I use: Niche unless you're doing real-time interactive video/3D-nav gen. Reusable idea: decouple short-horizon prediction from a fixed-size long-term memory to cap KV growth. Cred: promising-but-unproven, leaning show-me. SOTA is on their own "three-axis protocol", no standard benchmark, no code/weights, data leans on Unreal/game footage with real-domain transfer unshown. Builds on Streaming Multi-Agent Autoregressive Diffusion Model with World State Registers [2026-07-25]

4. EG-ARSA (rel 7), [Abstain] What's new: Distills expert road-safety judgment into a compact 8B vision-language model to audit road images cheaply in low-resource settings (dataset is Bangladesh). Can I use: The pattern, expert-grounded distillation into a small VLM for a narrow visual-inspection task, is worth borrowing if you ever build low-cost visual audit tools. The specific model isn't reusable off-region. Cred: abstain on the headline. Claims beating Gemini-2.5-Flash but names no metric and no contamination check, on a single-country dataset, and that superiority is the selling point, so I can't stand behind it from the card. Kappa=0.74 expert agreement is decent but subjective. Builds on Beyond Hazard Resemblance: Contrastive Event Adjudication for Training-Free Video Anomaly Detection [2026-08-12]

5. Provably adaptive sampling, discrete diffusion (rel 7), [Tentative] What's new: Proves sampling steps for discrete diffusion scale with the target's "dual total correlation" (how entangled the variables are), not raw dimension, so loosely-coupled distributions sample far cheaper than dimension-based lower bounds implied. Like packing a suitcase: cost is about how tangled the clothes are, not the item count. Can I use: Not now, pure theory. But it's the justification for investing in a better forward/remasking schedule on structured discrete data. Cred: promising-but-unproven. Theory-only, no empirics in the abstract, and the bound assumes a Bayes-optimal auxiliary sampler that may be intractable, so the "adaptive" win might not be realizable. That's the sharp concern. Builds on Provable diffusion-based posterior sampling for linear inverse problems via DDIM [2026-07-23]

6. Inertial Manifold Neural Operator (rel 7), [Tentative] What's new: Bakes inertial-manifold theory (dissipative systems collapse onto a low-dim attractor) into a neural operator so long rollouts of dissipative PDEs stop accumulating error; plus a shift-equivariant variant. Like knowing a spinning-down top settles into a narrow set of states, so you model only that set, not all of phase space. Can I use: Applicable if you forecast dissipative systems (fluids, reaction-diffusion) and standard FNO drifts over long horizons. Cred: promising-but-unproven. No rigorous ablation vs FNO baselines, and the approximation error-bound for the inertial manifold is missing, "more stable" is claimed, not nailed. Builds on PG-KINN: A Physics-Informed Petrov-Galerkin Kolmogorov-Arnold Network for Solving Forward and Inverse PDEs [2026-07-24]

7. Skill atrophy from AI assistance (rel 6), [Tentative] What's new: Measures how on-demand AI help erodes human skill, using a Bayesian latent-ability model to separate "was already good" from "got worse because of the crutch," in a controlled logic-puzzle setup. Lean on the answer button, your own muscle weakens. Can I use: Design argument, add friction (attempt-first, hint-before-answer) to assistant UX where you want users to keep the skill. Relevant to how you build tooling. Cred: promising-but-unproven. Small study, logic puzzles ≠ real workflows, no code/data. Direction matches known cognitive-offloading effects; the magnitude here is one small experiment. Builds on The Regression Tax: Decomposing Why Skills Help and Hurt LLM Agents [2026-07-28]

8. ConvergeFlow (rel 6), [Tentative] What's new: For continuous diffusion/flow language models, constrains the flow to stay inside the convex hull of token embeddings, which provably lands on valid tokens, killing the need for a separate cross-entropy decoder. Like snapping a GPS route to actual roads so you never end up parked in a lake. Can I use: Only if you're building decoder-free continuous diffusion LMs, research-stage. Code is open, so it's actually pokeable. Cred: promising-but-unproven, but the strongest reproducibility signal in today's batch (open code). Caveats: convergence needs "regularity conditions" that may bite, and it's tested only on OpenWebText, needs a real fight vs strong discrete baselines (Llama-3-class). Builds on Expanding Flow Maps [2026-07-25]

9. Correcting a learned physical invariant (rel 6), [Tentative] What's new: Shows DreamerV3's imagined rollouts drift because they quietly violate conserved quantities; projecting the latent state back onto a learned "conserved-quantity level set" fixes it. Like nudging a physics-sim ball back onto its energy budget each step so it doesn't slowly gain free energy. Can I use: Conceptual template for enforcing physical consistency in latent world models. Glued to DreamerV3 + a pendulum, so not drop-in. Cred: promising-but-unproven, a clean proof-of-concept. Single pendulum, frozen pretrained model, no ablation on projection magnitude. Nice idea, one toy's worth of evidence. Builds on Reinformed Dreamer: An Asymmetric World Model Efficiently Trained through Latent Guidance [2026-07-30]

---

Couldn't verify (required honesty pass):

• Every "beats X / SOTA" claim, ReWorld's three-axis-protocol result, EG-ARSA over Gemini-2.5-Flash, ConvergeFlow "competitive." No standard benchmark or metric in the cards; not checked.
• The three theory proofs (adaptive discrete-diffusion bound; ConvergeFlow convergence; inertial-manifold stability). I'm reporting what they claim, I have not checked the math.
• Code availability. Only ConvergeFlow is flagged code-open; the rest say "not mentioned," which is not the same as "no code."
• SWE Refactor Bench's real detection accuracy, the LLM-judge stage means the numbers ride on the judge, which I can't inspect.
• Skill-atrophy effect size, direction is plausible, magnitude is one small logic-puzzle study.

I did not run web searches, nothing here hinged on a single external fact I could resolve; the right move was to flag the unverifiable claims, not to launder them into confidence.

Daily Research Digest, 2026-08-25

8 papers, importance order. Tiers in brackets.

1. Asymmetric Capacity Allocation in Self-Refinement Pipelines What's new: In draft→critique→rewrite loops, the critic can be a small cheap model, size matters far more for the generator/refiner than the critic. Analogy: a sharp junior can cheaply flag "this is wrong"; you still need the senior to actually fix it. Can I use: Yes, directly. In your agent self-correction loops, route the critique step to a small local model (:8022/:8001), keep the big model for generate+refine. Straight inference-cost cut. Cred: Promising-but-unproven. Sharpest concern: benchmark contamination on Qwen3/Gemma3 could inflate the effect, and the "first study" claim is unsupported. [Tentative]

2. SPARCL: Spectral Partitioned Analytic Continual Learning What's new: Traces catastrophic forgetting to "spectral interference" in a shared operator; fix is to freeze the core eigenspaces and only update the residual ones. Analogy: don't repaint the load-bearing walls every time you add a room, only touch the trim. Can I use: Niche, exemplar-free class-incremental setups where you can't retain old data (privacy/memory) and replay is off the table. Cred: Promising. Sharpest concern: it rides a frozen ViT-B/16, so the "provable invariance" may not hold once the backbone itself needs to adapt. [Tentative] Builds on Reward Structure Shapes the Interaction Between Episodic Exploration and Neural Memory in RL [2026-08-07]

3. When Adaptation Hurts: Representational Drift → OOD Failures in MedSAM What's new: Full fine-tuning drifts the decoder and breaks under distribution shift; encoder-only LoRA stays robust under far-OOD. Analogy: over-tailor a suit to one body and it fits only that body; a light adjustment still works on others. Can I use: Yes as a general FT heuristic, when OOD robustness matters, prefer encoder-only LoRA over full FT. Transfers beyond medical imaging. Cred: Real-ish but narrow. Sharpest concern: evaluated on only dermatology + breast imaging, and CKA evidence is correlational, not a causal ablation. [Tentative] Builds on Phantom Gains: Auditing Self-Improvement Against a Measured Null [2026-08-22]

4. TurboBias 2.0: Streaming Context-Biasing for ASR What's new: Per-stream independent context lists + streaming inference for transducer ASR. Analogy: every concurrent caller gets their own live autocomplete dictionary. Can I use: Only if you build voice/ASR with personalized vocab (names, jargon) for many concurrent users. Not your current stack. Cred: Incremental engineering. Sharpest concern: it's sold as "production-efficient" but the abstract gives no latency/throughput numbers. [Tentative]

5. Move by Move: Measuring and Steering How LLMs Conduct Psychotherapy What's new: A validated 10-category ontology of therapy "moves" + zero-shot steering that aligns an LLM's output distribution to real clinician patterns, no fine-tuning. Analogy: hand the model a rubric of therapist chess moves and nudge it to play like the pros. Can I use: The steering mechanism is the transferable part, constrain/audit LLM behavior in sensitive conversational domains without FT. Cred: Promising. Sharpest concern: validated on just 5 psychologists, and the judge-based scaling inherits LLM eval bias, the alignment gain may be overstated. [Tentative]

6. Unified Branch-and-Bound for Steiner TSP on Graphs of Convex Sets What's new: Jointly solves discrete routing + continuous trajectory with a termination guarantee and an ε-optimality certificate. Analogy: a delivery planner that picks the stop order AND draws the smooth path, with a receipt saying how close to optimal. Can I use: Robotics/mobile-manipulator planning. Not your lane. Cred: Theory is solid, practice isn't. Sharpest concern: 28.1% mean optimality gap badly undercuts the "certified" pitch for anything precision-sensitive. [Tentative]

7. PerturbRx: Treatment-Conditioned Latent Transitions for Drug Response What's new: Learns drug-induced latent shifts from single-cell data to predict patient drug response without post-treatment patient samples. Analogy: predict a patient's reaction from a cell-dish rehearsal, skipping the wait-and-see. Can I use: Comp-bio/precision-oncology only, with heavy domain preprocessing. Not applicable to you. Cred: Skip until proven. Sharpest concern: SOTA claimed on TCGA/PDX with no ablation of the transition module, no code, and PDX contamination risk, nothing checkable. [Abstain]

8. Truthful Calibration Measures for Sequential Prediction What's new: Proves exact truthfulness is incompatible with completeness + soundness in sequential binary prediction; offers better approximate-truthfulness reductions. Analogy: a scale that can't be perfectly honest, complete, and consistent at once, you must trade. Can I use: Not deployable, a theoretical guardrail on what calibration/eval metrics can promise. Worth knowing if you design calibration evals. Cred: Theory result, improves on 2024 work. Sharpest concern: no empirical check, and I can't verify the proof or its asymptotic assumptions. [Tentative] Builds on How to Verify Consistency of Probabilistic Claims [2026-08-13]

---

Couldn't verify (be skeptical here):

• Every SOTA/benchmark claim (PerturbRx on TCGA/PDX; Asymmetric on Qwen3/Gemma3), contamination plausible, unchecked by me.
• The "first study" claim in Asymmetric Capacity Allocation.
• TurboBias 2.0's "production-efficient", no latency/throughput numbers given.
• Steiner TSP's 28.1% gap, reported, not independently confirmed.
• Correctness of the Truthful Calibration impossibility proof.
• Magnitude of Move by Move's no-FT steering effect.
• Whether SPARCL's "provable invariance" survives outside the frozen-ViT setup.
• The "closest prior" picks for SPARCL and MedSAM cross-links, I chose by topic among weakly-related priors; that linkage is my judgment, not stated in the data.

Note: I skipped the usual opposing-view section, a research digest has no single decision to argue against, so forcing one would just add noise. If you want me to steelman a specific "should I actually use this" call (e.g. the asymmetric-critic idea in your agent loop), say which and I'll pressure-test it.