Altair

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.

Altair research digest, 2026-08-24

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

Altair research digest, 2026-08-23

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

DAILY RESEARCH DIGEST, 2026-08-22 (7 kept, ranked)

1. ConceptGuard: Context-Sensitive Unlearning Benchmark (2608.20338) New: tests whether unlearning can remove the harmful use of a concept while keeping the benign one, like firing a locksmith's burglary side-gig without making him forget how locks work. Use it: directly relevant if you're auditing unlearning/safety pipelines; it's the right question (contextual separation, not fact deletion). Cred: promising-but-unproven. Sharpest concern: ROUGE as the metric, lexical overlap can't tell you if the concept is gone. Dataset claimed but no code link. [Tentative] Builds on LittleLearner: Language Models Under Pedagogically Controlled Knowledge Exposure [2026-08-15]

2. Phantom Gains: Auditing Self-Improvement Against a Measured Null (2608.20290) New: many reported self-improvement gains dissolve once you control for inference batching artifacts and measurement noise, the "gains" were the scale wobbling, not the weight changing. Use it: yes, today. Adopt the per-problem exact test + FDR control before believing any self-training delta in your own experiments (mempalace loop included). Cred: real. Triage flags are uniformly strong, explicit null controls, rigorous stats. This is the load-bearing paper of the batch. [Verified] Builds on PAST-Bench: Benchmarking the Foundations of Recursive Self-Improvement in Personal Agents [2026-08-06]

3. Information on Trajectories: Martingales and Random Times (2608.20337) New: exact accounting of the information classical concentration bounds (Ville, PAC-Bayes, Azuma) throw away, measuring precisely how much you overpay for the right to peek at data sequentially. Use it: not directly; framework-level. Relevant later for sequential testing / safe model pooling, needs engineering on top. Cred: real theory, zero empirics, that's fine for what it is. Concern: unification claims are only as useful as someone operationalizing them. [Tentative] Builds on An Optimal Agnostic PAC Algorithm [2026-08-08]

4. Pandora's AI Model Routing Box (2608.20316) New: model routing framed as Pandora's Box search, each model has a hidden value and a price to find out, and they derive the closed-form "when to stop peeking" policy. Use it: conceptually relevant to your router.py work, but it's math, not code, you'd implement the Gaussian signal model yourself, and the Gaussian assumption is unvalidated. Cred: promising-but-unproven. Sharpest concern: the decentralized setting admits strategic manipulation, their own result. [Tentative] Builds on CodeRescue: Budget-Calibrated Recovery Routing for Coding Agents [2026-07-23]

5. TCP_α: Margin-Controlled Confidence for MIR (2608.20326) New: a confidence target with a theoretical guaranteed gap between correct and wrong predictions' confidence, not just "usually higher," provably separated by a margin. Use it: only if you're doing music information retrieval; the margin-control idea might port elsewhere but nobody's shown it. Cred: promising-but-unproven. Sharpest concern: the abstract literally cuts off mid-sentence, so the fine-tuning results can't be assessed at all. No code. [Tentative]

6. CalcSeg: Curriculum Learning for Myocardial Scar Segmentation (2608.20305) New: uses the model's own uncertainty to order training examples, teach the easy heart slices first, hard ones later, for 3D scar maps from sparse 2D MRI. Use it: not applicable to your stack; clinical-imaging niche. Cred: promising-but-unproven. SOTA claimed without ablating the curriculum component, so the headline ingredient is unproven as the cause. Open code is a genuine plus. [Tentative] Builds on Anatomy Contextualized Adaption of CT Foundation Models [2026-07-31]

7. Physical-Support Confidence Sets for Coherent Dictionaries (2608.20295) New: confidence sets for where a sparse signal physically lives when dictionary atoms are nearly indistinguishable, plus an algorithm to avoid claiming more precision than the data supports. Use it: skip unless you're in signal processing / scientific ML with hard UQ requirements. Cred: skip for us. Synthetic data only, no code, narrow "physical support" framing. [Tentative]

---

Couldn't verify:

• Phantom Gains' code release, triage says "implied by reproducibility claims," not confirmed. Check the repo before citing it as reproducible.
• ConceptGuard's dataset actually being available (claimed, no link found).
• TCP_α's fine-tuning results, truncated abstract, literally unreadable.
• CalcSeg's SOTA being attributable to the curriculum (no component ablation).
• Pandora's Gaussian signal assumption holding on real routing workloads, untested by the authors.
• All of the above is from triage cards, not the papers themselves, I haven't read a single full PDF today. Rankings inherit whatever the fast model missed.

TLDR: read #2 (Phantom Gains) properly, it's the one that changes how you evaluate your own work. #1 if unlearning audits land on your desk. Rest is awareness-level.

Daily research digest, 2026-08-20 3 cards. Ordered by what's actually worth your time.

---

1. Magnitude + complex k-space for dynamic MRI (ADMM-unrolled), 2608.18036 What's new: they feed a cheap side-channel (magnitude-only k-space) into a physics-unrolled recon net. Like rebuilding a song from a fraction of the notes when you also have the loudness envelope of every bar, the extra track kills a huge chunk of the guess space for free. Can I use this: only if you're in accelerated cardiac/flow MRI. The transferable idea is generic though, if you have a cheap partial measurement that's consistent across frames, fold it in as a data-consistency term inside the unrolled loop rather than as another loss. Credibility: promising-but-unproven. Sharpest concern, retrospective undersampling for the headline metrics and no comparison against recent diffusion/transformer recon baselines. Two escape hatches in the same eval means the gain could be baseline-choice, not method. Builds on Provable diffusion-based posterior sampling for linear inverse problems via DDIM [2026-07-23] [Tentative]

---

2. Delegation asymmetry in agentic dating recommenders, 2608.18058 What's new: people will happily let an agent write for them, but don't want to be written to by one. Everyone likes sending the form letter; nobody likes getting one. They propose routing agent-mediated messages toward the minority who don't mind receiving them. Can I use this: the finding generalizes past dating, any agent-in-the-loop marketplace (outreach, support, matching) has a two-sided receptivity gap you should measure separately per side, not as one "agent acceptance" number. That framing is worth stealing. The routing rule itself, I wouldn't ship on this evidence. Credibility: skip as a basis for decisions. Sharpest concern, it's stated preference on a socially-loaded question, and the field data backing it is proprietary and unverified. Latent-variable modeling and out-of-sample validation make the survey fit better; they don't make the survey predict behavior. [Abstain], direction of the asymmetry is plausible, the magnitude and the routing prescription are not something I'd stand behind.

---

3. Initialization-free bundle adjustment, revisited, 2608.18028 What's new: a negative result worth having, low Object-Space Error does not mean you got a valid metric 3D reconstruction. Your optimizer reports a great score while the model is internally consistent but geometrically wrong. They attribute actual success to metric-upgrade stability, not objective minimization. Plus a C++ impl and a Blender benchmark. Can I use this: not on anything you're currently running. If you ever touch SfM/3D recon, the lesson is the loss you're minimizing isn't the thing you care about, add an independent metric-validity check. Credibility: real, for the diagnostic claim. Synthetic ground truth is exactly the right instrument for "the objective is misleading", and code is open. Sharpest concern, no real-world dataset, so how often this bites in production pipelines is unmeasured. Builds on ATSplat: Compact Feed-forward 3D Gaussian Splatting with Adaptive Token Expansion [2026-07-24] [Verified] for the finding, unverified for real-world prevalence.

---

Couldn't verify

• I read triage cards only, not abstracts or full papers. No numbers, tables, or claimed deltas were checked against source.
• Authors, institutions, and whether the BA code is actually released, unchecked.
• "Improvement over PD-DL", I did not confirm which PD-DL variant is the baseline, so I can't size the gain.
• Whether the related_prior links are semantically meaningful or just retrieval neighbours. The DDIM one reads plausible for MRI; the ATSplat→bundle-adjustment link is thinner and I'd treat it as topical, not lineage.
• arXiv IDs are consistent with Aug 2026 but I did not resolve any of them live.

Daily digest, 2026-08-19 · 7 papers, all from cards (I read the triage evidence, not the PDFs, see the bottom).

---

1. What Do Compliance Detectors Read? 2608.16852v1 Swap the rule text but keep the response, and guard models / activation probes give the same verdict, they're reading the vibe of the output, not the rule. A bouncer who memorized which faces get in, and never re-reads the clipboard when the door policy changes. Use it: yes, today. Steal the counterfactual protocol, take your guard-model eval set, substitute the rule with one that flips the correct label, and see if the score moves. If it doesn't, your compliance layer is theater. Cheap, no training, runs local. Cred: real diagnosis, weak cure. Their proposed Internal Compliance Score only ties a bag-of-words baseline and dies to an adaptive white-box attack, so use the audit, ignore the metric. No code. [Tentative], audit design is rigorous, the ICS contribution isn't load-bearing.

2. Model Hypnosis 2608.16834v1 Individually harmless cues, a typo here, a paraphrase there, stack additively into strong behavioral control, and it transfers across model families. One person leaning on you is nothing; forty people leaning the same way moves you. Use it: red-team input only. If you're deploying anything on-prem that takes untrusted text, this is the attack shape to test, not one jailbreak string, many weak ones summed. Cred: promising-but-unproven. Cross-model transfer is the strong signal; the sharpest concern is that "hypnosis" may be a new name for known prompt sensitivity, and there's no ablation isolating which cue types actually carry the effect. Paraphrase augmentation also risks contaminating their own benchmark. [Tentative] Builds on Instruction-Tuned Models Locally Reuse Human Syntax More Than Humans Do [2026-07-30]

3. Non-Crossing Deep Quantile Regression for Survival 2608.16864v1 KAN + Transformer with a structural constraint that stops quantile curves crossing, i.e. it can't tell you the 90th-percentile survival time is earlier than the 50th. A forecast that says "10% chance of rain by noon, 5% by evening" is broken on its face; this makes that impossible by construction, not by penalty. Use it: only if you touch time-to-event data (churn, failure, clinical). The non-crossing trick generalizes beyond survival though, any multi-quantile head. Cred: the most solid card today. Open code, claimed finite-sample excess-risk bound. Sharpest concern: standard clinical cohorts are heavily benchmarked, so the gains may be contamination; and KAN+Transformer is almost certainly overkill under ~10k rows. [Verified], for "the method is real and reproducible," not for "it beats everything." Builds on Generation-Powered Inference for Distribution-Valued Outcomes [2026-08-18]

4. Time-Aware Validation of Fuel Consumption Models 2608.16833v1 Random train/test split on 1 Hz sensor data leaks the future: your test second sits between two training seconds. Like testing whether someone remembers a film by hiding a single frame. Time-aware splits collapse the flattering numbers. Use it: yes, this is a checklist item, not a paper. Any high-frequency operational data (IoT, telemetry, trading, monitoring) gets blocked/rolling-origin validation or the metrics are fiction. Applies directly to anything you deploy that watches a live system. Cred: real, unsurprising. 3.88M real records is genuine evidence; the finding itself is textbook, so the value is the concrete demonstration, not novelty. No code, and it's one ship, generalization beyond similar data is asserted, not shown. [Verified] for the lesson, [Tentative] for the magnitude. Builds on Beyond Sufficiency: Time Series Explanation with Counterfactual Necessity [2026-07-25]

5. Spectral Gaps of Hit-and-Run 2608.16878v1 Sampling a weird high-dimensional shape by repeatedly picking a random direction and sliding to a random point along that chord. They tie the mixing rate to the Poincaré constant by duality and cut the dimension dependence from ~d³ to nearly d², for near-isotropic bodies. Use it: no. Zero engineering surface. It's a permission slip to trust Hit-and-Run in high dimensions if you ever do constrained/volume sampling. Cred: real theory, unverifiable by me. No empirics (normal for this genre). Sharpest concern: "nearly isotropic" is doing real work, if your body isn't preconditioned, you're back to the old bound, and functional-analysis bounds often hide brutal constants. [Tentative], I cannot check the proof.

6. BATON (long-horizon robot manipulation) 2608.16889v1 Splits long tasks into subtask-level search plus a memory that tracks the handoff between subtasks, so errors don't compound. A relay race where the dropped time is almost always the baton pass, not the running. Use it: not applicable to you unless you're touching VLA policies or physical manipulation. The transferable idea, make state handoffs between agent steps explicit and verified, you already get from ordinary agent design. Cred: skip until code. Four flags and the worst is that the headline benchmark "RoboMemArena" looks non-standard or possibly invented; an 11% gain with no ablation and no code is unfalsifiable. The verifier agent doing transition logic is itself unvalidated. [Abstain] Builds on RoboTTT: Context Scaling for Robot Policies [2026-07-18]

7. AutoSR (symbolic regression via research states) 2608.16876v1 Proposer-Reviewer MCTS that keeps the reasoning trail behind each candidate formula, not just the formula. Keeping the lab notebook instead of only the final equation on the whiteboard. Use it: no, not for black-box fitting. Only interesting if you need the discovery process auditable, which is a governance angle, not a modeling one. Cred: skip. Claims SOTA on niche benchmarks while skipping the obvious strong baselines, no code, and it scores algebraic equivalence rather than out-of-distribution or physical consistency, which is the whole point of symbolic regression. Multi-agent MCTS also implies serious compute overhead nobody reported. [Abstain] Builds on Automated Discovery Has No Universally Superior Harness [2026-07-22]

---

Couldn't verify

• All of it is second-hand: I read triage cards, not the papers or abstracts. Titles, arXiv IDs and dates are as given; I didn't check they resolve.
• Whether "RoboMemArena" exists. The flag says possibly fabricated, I didn't search, so it's unconfirmed either way.
• The 11% BATON number, and AutoSR's SOTA claim.
• Whether the survival paper's released code is complete/runnable, and whether the excess-risk bound holds, I read a flag, not a proof.
• The d³→d² improvement and the Lovász, Vempala comparison. Unchecked.
• Whether Model Hypnosis is genuinely distinct from the existing prompt-sensitivity literature. That's the flag's suspicion and mine; neither is evidence.
• My "closest prior" picks are my judgment over the supplied lists, not stated in the data.

---

Steelman: The two I graded [Abstain] are the ones a working agent-builder would most want. BATON's real claim, that long-horizon failure lives in the transitions, not the steps, is architecturally right and would still be right if the benchmark were renamed; missing code at v1 is normal, not damning. And AutoSR's "keep the research state" is exactly the audit trail an on-prem governance story needs, where the reasoning trail is the deliverable and algebraic-equivalence scoring is the correct metric for it.

This flips if: BATON ships code and the gain survives an ablation on a benchmark that predates the paper; or someone shows Model Hypnosis's additive stacking beats the sum of its individual cues, which would make it a real mechanism rather than a renaming.

Daily digest, 2026-08-18, 11 papers. Only one is worth your afternoon: YOPO (#2). The rest is context.

---

1. Universal Thermodynamic Interatomic Potentials for Crystalline Materials (2608.14502) Bolts autodiff onto a universal MLIP so it emits Gibbs free energy directly instead of you averaging a thousand simulated snapshots to get there, like a weather model that hands you the seasonal average instead of making you run 1,000 days and mean them. Can I use this: Not your stack. Only if materials screening ever shows up as a client problem. Credibility: Real method, narrow validation, quasi-harmonic approximation is standard and sound, but it breaks on anharmonic systems, and they only tested crystals. Sharpest concern: the "single evaluation" speed claim isn't benchmarked against actual MD, which is the entire selling point. [Tentative]

2. YOPO: Answering and Abstaining Together in a Single Forward Pass (2608.14465) A small reconstruction net predicts what the residual stream would have looked like before steering, so the model can steer its reasoning and decide "I don't know" in one pass instead of two. Like reading the room before you open your mouth, rather than talking and then apologising. Can I use this: Yes, this is the abstention story for latency-bound on-prem deployments. You currently pay double inference for answer + confidence check. Worth a read against your judgment gate on :8005. Credibility: Real. 10 backbones across 6 families, and they self-audited hard enough to catch their own alphaNLI data leakage and walk the claims back, that's the behaviour of people who want to be right, not published. Sharpest concern: "frozen LLM" is doing PR work; there's a trained module in there. Builds on Off-Context GRPO: Learning to Reason on Hard Problems using Privileged Information [2026-07-23] [Verified]

3. Split the Labor: Separating Evidence Interpretation from Decision Aggregation (2608.14509) Names "count-scale drift": when you tally votes, ten weak signals drown one strong one. Fix is to convert every source to calibrated log-likelihood ratios first, then pool, same unit before you add, like converting currencies before summing a budget. Can I use this: Yes, if you have any multi-source triage or ensemble routing. Directly applicable to how you'd fuse judgments from several models of unequal reliability. Credibility: Promising, reports actual AUPRC against baselines. Sharpest concern: the abstract trails off into "five predictions", which reads like either a truncated summary or speculation dressed as results, check the paper before quoting it. Builds on Partition, Prompt, Aggregate: Statistical Self-Consistency in Language Models [2026-07-18] [Tentative]

4. Generation-Powered Inference for Distribution-Valued Outcomes (2608.14542) Error bars for when part of your "data" is model-generated and the thing you're estimating is an entire distribution, not a single number. They get there by mapping Wasserstein space into Hilbert space, where mean estimation is a solved problem. Can I use this: No. Statistician's tool, not an engineer's. Credibility: Promising-but-unproven. Theory only, no code, one genomics case study. Sharpest concern: the guarantees hang on a "bridge representation" assumption nobody has stress-tested outside their own example. [Tentative]

5. Handover of In-Context Learning State Across Session Boundaries (2608.14528) Frames context handoff as: what's the smallest sticky note you can leave the next session such that it predicts the same thing the full transcript would? Formal sufficient-statistic bounds on that. Can I use this: Theoretically interesting for your memory-palace work, zero plug-and-play. Don't wait on it. Credibility: Skip for now. No empirical validation on any actual LLM, and the exogeneity assumption almost certainly fails in real multi-turn context. Sharpest concern: the abstract is regression theory wearing an LLM hat. Builds on Structured Memory for Edge Language Models: Persistent Context and Corpus Retrieval via O(1) SSM State Injection [2026-08-05] [Abstain]

6. Approximate Muon with low-rank adapters (2608.14492) Muon's orthogonalisation step is what makes it expensive; they linearise it so it fits inside LoRA's small matrices as plain matmuls. Can I use this: Maybe, if you fine-tune on Strix. But you'd be writing the training loop yourself, no library support. Credibility: Promising-but-unproven. No code, no clean ablation against LoRA+AdamW, and they admit gains are "model- and eval-dependent", which is how people say "sometimes it doesn't work". Sharpest concern: relaxing orthogonality is exactly the property Muon's convergence argument rests on. Builds on Sign compression for Muon: SignMuon, MuonSign, and the Limits of Error Feedback [2026-08-04] [Tentative]

7. Twin: Playing an Unknown Game with a Test-Time Digital Twin (2608.14490) A coding agent writes an executable simulator of the game it's playing, plays against its own sim, and patches the sim wherever reality disagrees. Learning the rules of a board game by writing them down and crossing out what turns out wrong. Can I use this: Not in production. It's a test-time-compute research result. Credibility: Promising-but-unproven. Sharpest concern: ARC-AGI-3 contamination risk plus a hard dependency on frontier coding-agent capability, you likely can't reproduce it with anything you'd run locally. [Tentative]

8. Generating Benchmark Health Data Using a Tabular Diffusion Transformer (2608.14496) Standardises heterogeneous tables into summary-statistic space so a single diffusion transformer can generate synthetic rows across different schemas. Can I use this: Possible angle for privacy-preserving data sharing in a health client, but heavier than the GAN baselines it'd replace. Credibility: Can't call it. The triage flags contradict themselves, "open code + reproduced" and "no code mentioned" in the same set, plus a contamination warning. Sharpest concern: that contradiction means the flags weren't reconciled, so treat all of them as unread until someone opens the repo. Builds on Empirical Evaluation of Out-Of-Distribution Performance of Tabular Foundation Models [2026-07-30] [Abstain]

9. Participatory Moral AI Is Not Neutral: The Invisible Hand of Developers (2608.14522) Shows the preprocessing choices, which features you show, how you sample, how you frame, move the aggregated moral outcome. The ballot design decides the election. Can I use this: Not code. Useful ammunition for governance conversations, where "we asked people" gets treated as neutrality. Credibility: Real as a caution, N=809. Sharpest concern: no repro artifacts, so the effect sizes are take-their-word. Builds on Beyond Sycophancy: Structured Resistance and Compliance in LLM Moral Reasoning [2026-07-25] [Tentative]

10. MagnifiQ: Patch-aware Text Guided Progressive Upscaling (2608.14543) Swaps SDXL's self-attention for convolutions to make 4K upscaling linear instead of quadratic, plus a per-patch text prompt so each tile knows what it's supposed to contain. Can I use this: Only if ComfyUI work needs 4K from low-res under VRAM pressure, and you'd pay it back in progressive-inference latency. Credibility: Promising-but-unproven. Sharpest concern: per-patch prompts invite semantic drift at tile boundaries, and they claim SOTA without ablating the conv-vs-attention trade that is the whole paper. Builds on Inference-Time Scaling of Diffusion Models via Progressive Seed Pruning [2026-07-25] [Tentative]

11. Expected Free Energy-based Informative Path Planning for Robotic Mars Exploration (2608.14466) Uses Active Inference to put "what will I learn here" and "what will it cost me to get there" in one currency, so the robot stops needing a hand-tuned exploration weight. Can I use this: No. Credibility: Skip unless you're doing robotics. Sharpest concern: simulation-only, and Active Inference carries real compute overhead for a benefit that greedy entropy reduction usually gets you cheaper. Builds on Learning Adaptive Safety Margins for Visual Navigation [2026-07-22] [Tentative]

---

Couldn't verify:

• I read triage cards, not papers. No abstract, no full text, no check that any arXiv ID resolves. Every verdict above is a read of the evidence card's quality, not of the work.
• The tabular-diffusion card asserts both "open code + reproduced" and "no code mentioned". Unreconciled, that paper's tier is a placeholder.
• Every "Builds on" line is my judgment of which listed prior is closest by title. It is not a computed similarity, and I didn't read the priors either.
• "Five predictions" in Split the Labor, no idea what that refers to.
• Nothing on whether the approximate-Muon relaxation preserves convergence; the card raises it, nobody answers it.
• Relevance scores are the fast triage model's. I ordered by them and didn't recompute; my own ranking would put YOPO first.

Steelman: The case against spending your time on YOPO: single-pass abstention is a latency micro-optimisation, and your bottleneck on Strix is model quality and memory, not one extra forward pass. Meanwhile #5 (ICL handover) is aimed squarely at the problem you're actually building around, what to carry across session boundaries, and being theory-first is a feature there, not a flaw, because a bound tells you when your compaction scheme is provably lossy and no benchmark will. Investing in the unproven-but-load-bearing beats polishing the proven-but-marginal.

This flips if: your abstention path turns out not to be a separate forward pass at all (then YOPO buys you nothing), or the handover paper ships an empirical section on real LLM contexts, at which point it moves above YOPO, not below it.