How this triage was run
Checked against primary sources on 22 October 2026. These are research-funding recommendations, not authorization to launch. Public availability does not mean successful local reproduction. No experiments were run. Completing a replication or obtaining a null result does not guarantee publication.
SP-002 — Grothendieck bounds
Strong premise, but separate author-verified theorems from system-tested claims. The case study: arxiv.org/abs/… dates from August, not October, and documents a withdrawn numerical upper bound. Its reported approximately $5,400 model expenditure, four-GPU search node and human steering are historical costs, not a quote for this project.
The mathematical paper: arxiv.org/html/… and certificate repository: github.com/trishullab/grothendieck-bounds provide an Arb/python-flint route. The repository lists a complete cubic–quintic certificate for 1.7818666069360661, but the stronger headline endpoint has an explicitly incomplete construction. Script contents and environment compatibility were not fully verified.
Cheapest useful test: replay one complete certificate, then compare AI-selected search choices with a fixed strategy under equal budgets. Numerical optimization alone cannot certify a universal bound. Reproduction tests the pipeline; unsuccessful search proves no impossibility.
A new certificate or controlled strategy comparison could contribute something. A failure diary largely duplicates the existing case study.
SP-003 — BB(6) Cryptids
The paper: arxiv.org/abs/… supplies formalized accelerations, not eleven halting decisions. It explicitly leaves those machines unresolved and their Cryptid classification provisional. The public Lean repository: github.com/google-deepmind/cryptids exists, but a complete autonomous discovery harness was not verified.
Replaying a proof validates that artifact, not the ability to discover a new acceleration. Any new result must distinguish acceleration, reduction and actual halting/nonhalting resolution. Targets also require deduplication against the evolving repository catalogue: github.com/google-deepmind/cryptids/blob/…/README.md and community ledger: wiki.bbchallenge.org/wiki/BB(6).
A genuinely new verified reduction could be valuable. The proposed months-long rebuild presently lacks a distinct method and fresh target set.
Re-entry: a deduplicated target list, a specific methodological improvement, a Lean-capable owner and a successful bounded proof-reproduction gate.
SP-004 — AI-tractable mathematics
The developments are substantially real, but “published successes versus still-open problems” is not a clean tractability dataset. Unsolved problems may never have received comparable effort. Certificate length measured after discovery leaks the outcome. Ten forecasts do not, by themselves, validate a general signature.
The KLS paper: arxiv.org/abs/… and formalization project: github.com/HongruZhao/KLS15 exist. The OpenAI repository: github.com/openai/math now reports 719 manuscripts in 372 families; its history: raw.githubusercontent.com/openai/math/main/history.md records withdrawals and repairs. These are not 719 independent accepted resolutions. Complete comparable attempt/failure/effort records were not verified.
A versioned, independently adjudicated prospective dataset could be useful. Retrospective separation alone cannot establish AI tractability, and existing tractability projects create duplication risk.
Re-entry: ex-ante measurable features, time-stamped verified outcomes, an attempted-problem denominator or equal-budget prospective panel, effort/model/field controls, and a preregistered scoring and power plan.
SP-005 — qLDPC correlated-noise benchmark
The barbell paper: arxiv.org/html/… is real, but its best headline figures should not be combined into one hardware claim. The near-1,000× example compares equal data-qubit counts under circuit-level depolarizing noise. The 8× code-parameter figure concerns a different construction whose additional circuit depth hurts simulated performance.
Infleqtion’s release: infleqtion.com/newsroom/announcements/infleqtion-advances-fault-tolerant-quantum-computing-software-with-nvidia-cuda-q-logical describes approximately six physical data qubits per logical qubit, not complete hardware overhead. It identifies syndrome extraction, noise modeling, decoder benchmarking and physical execution as subsequent work.
Stim: github.com/quantumlib/Stim, qLDPC: github.com/qLDPCOrg/qldpc, Relay-BP: github.com/trmue/relay and PyMatching: pymatching.readthedocs.io/en/stable are public. A runnable vendor barbell circuit release was not verified. Stim supports correlated Pauli events, not arbitrary leakage or coherent dynamics.
The “empty independent-comparison lane” is overstated: geometry-correlated qLDPC research: arxiv.org/abs/2604.01040 and its code archive: zenodo.org/records/19337541 already exist.
Bound the first test to one barbell construction and one correlated-Pauli mechanism. Require circuit validation and baseline reproduction before a sweep. Compare equal logical tasks, total resources, timing assumptions and decoder budgets. Output a sensitivity/crossover map, not a verdict on real hardware. Reconstruction and rare-event decoding costs remain unbenchmarked.
SP-006 — Solar-wind helicity
Strong source and artifact availability. The primary paper: arxiv.org/html/… studies reduced radial polarization and opposite-signed spectral peaks, not a universal observer-independent clockwise/anticlockwise rule. Its distinctive result is radial spectral migration that does not collapse onto local ion scales.
The Zenodo archive: zenodo.org/records/22779133 provides a reachable 20.8 MB release with plotted data and code. The repository: github.com/Institute-of-Theoretical-Physics/helicity-evolution_psp warns that raw preprocessing is not turnkey and intermediate files are missing. Official PSP data access exists, but individual raw-file retrieval was not tested.
A raw-data independent implementation with predeclared held-out intervals can decisively test a specific spectral statistic. Replotting the authors’ arrays cannot. Single-spacecraft spectra also cannot alone establish a causal nonlocal-transfer mechanism.
Potential output: a controlled replication dataset and methods note. Detailed pilot below.
SP-007 — K2-18 b multiverse
Substantial duplication. Stevenson et al.: arxiv.org/html/… already performed independent reductions, bin-width/pixel-shift tests and retrieval comparisons. The 2026 ten-planet survey: arxiv.org/html/… uses existing spectra, not new K2-18 b observations, and explicitly omits several alternative-bin reductions.
The approximately 6–7% survey figure concerns believing extraterrestrial life had probably been found, not accepting DMS detection; responses were collected in 2025. See the survey paper: nature.com/articles/s41550-026-02876-9.
Machine-readable spectral tables: cdsarc.cds.unistra.fr/viz-bin/…/257?format=html&tex=true, Eureka: github.com/kevin218/Eureka and POSEIDON: github.com/MartianColonist/POSEIDON are available. Raw-download access, complete configurations and retrieval costs remain untested.
A fraction of configurations supporting DMS depends on arbitrary configuration selection and weighting; it is not a calibrated probability of a molecule or life.
Re-entry: an unperformed, bounded injection/recovery or independent-transit test, with fixed hypotheses, configuration weights, false-positive calibration, recovery power and convergence criteria. Another uncalibrated percentage is insufficient.
SP-008 — Brain–LLM layer timing
A credible narrow opportunity, but the card overlooks existing controls. The target article: nature.com/articles/s41467-025-65518-0 appeared November 26, 2025. It already includes symbolic comparisons, shuffled layer indices, interpolated pseudo-layers and additional statistical checks. The inspected encoding source: github.com/DVader96/GPTology/blob/…/tfsenc_utils.py uses non-shuffled folds. Do not characterize it as an uncontrolled or shuffled-split study.
Hadidi et al.: nature.com/articles/s41467-026-72253-7 study generic confounds on other datasets. Azizpour et al.: elifesciences.org/reviewed-preprints/110085v1 revisit pre-onset prediction on this Podcast dataset, not the complete layer-order endpoint.
The narrow test is incremental layer-to-latency ordering after a joint acoustic, lexical and positional nuisance model. The pinned GPTology repository: github.com/DVader96/GPTology/tree/7e76c49333a0…, OpenNeuro dataset: doi.org/10.18112/… and dataset paper: nature.com/articles/s41597-025-05462-2 provide a real substrate. Payload downloads and preprocessing equivalence still require checking.
Potential output: a timing-specific robustness benchmark, not proof that brains implement transformers. Detailed pilot below.
SP-009 — Protein watermark robustness
The Nature paper: nature.com/articles/s41586-026-10965-y and official repository: github.com/google-deepmind/synthidbio exist. However, the paper already studies robustness, so “empirical and untested in public” is incorrect. Sequence and structural watermarks are also different objects and should not be pooled into one survival claim.
The proposed computational endpoint cannot establish retained biological function. Structural confidence and binding proxies are not experimental function measurements. Recreating a different watermark on an open model would not validate claims about the original system.
An unrestricted removal-optimization benchmark would also need explicit biosecurity review and a responsible disclosure/release plan. This triage provides no evasion procedures.
Re-entry: a defensive, non-evasive detector-calibration or false-positive study on approved benign data, with a defined threat model, distinct novelty, validated functional labels where needed, and an appropriate safety review.
SP-010 — New-target protease design
The cysteine-protease paper: nature.com/articles/s41586-026-11143-w is real and reports experimentally measured activity. Its data and code availability statements are public. The metalloprotease work is a separate companion article, not a single interchangeable recipe.
The proposed in-silico pass rate is not a success rate for catalysis or substrate selectivity. The paper’s experimentally validated rate cannot fairly be compared with confidence-filter pass rates on new targets. No actual benign target or matching assay is specified, and “plastic/pollutant” does not identify a protease substrate.
Public source code is not identical to unrestricted model access or a working environment. AlphaFold 3’s repository: github.com/google-deepmind/alphafold3 distinguishes code licensing from model-parameter terms. Compute, dependency access and validation costs remain unverified.
A ranked candidate list might generate hypotheses, but would not establish the transfer claim. No engineering procedures or sequences are recommended here.
Re-entry: a clearly benign substrate, appropriate biological safety review, a retrospective benchmark with genuine activity labels, verified model/dependency access, and a qualified experimental collaborator or a strictly limited computational claim.
SP-011 — UFO humanoid robustness
The official repository: github.com/Roboparty/UFO is public, but the current implementation uses mjlab, not the card’s proposed Isaac Lab replication substrate. It uses motion datasets; unsupervised RL does not mean no demonstrations, motion references or engineered constraints.
A released G1 checkpoint: huggingface.co/xuewang/ufo-g1-policy and robot-specific motion data: huggingface.co/datasets/xuewang/UFO-MotionData are reachable. The checkpoint lacks optimizer state and replay buffer, so it is not an exact training-resume package. Code/policy and underlying motion-data licenses have noncommercial restrictions; the data also retain no-derivatives terms.
Cheapest useful test: validate the released policy in its native simulator, then measure prespecified perturbation recovery against a suitable matched baseline. Freeze robot, tasks, initialization, observation/action constraints, disturbance definitions and success criteria. A checkpoint test evaluates behavior, not training reproducibility or real-world safety.
Full training is not automatically cheap: the authors report approximately five hours on eight H20 GPUs for a 192M-step run. That is a historical allocation, not our runtime or price. No suitable GPU allocation has been verified.
Duplication is meaningful: BFM-Zero: lecar-lab.github.io/BFM-Zero and P-BFM: arxiv.org/abs/2610.15833 already address recovery and disturbances. The contribution must be independent, resource-matched measurement, not claiming to invent push testing.
SP-012 — Interfacial heavy fermions
The 2026 article: nature.com/articles/s43246-026-01332-5 is real. It presents interfacial hybridization evidence, but its raw experimental data are by request. Its slab DFT excludes electron correlation and does not itself predict heavy quasiparticles.
Lattice matching and band-alignment proxies can nominate hypotheses, not establish a coherent heavy-fermion interface. Growth feasibility, valence, magnetic competition and many-body effects remain unresolved. The 2023 YbCu2 work: nature.com/articles/s41467-023-43662-9 and earlier computational heavy-fermion searches: arxiv.org/abs/2206.00076 also narrow the novelty.
Materials Project: docs.materialsproject.org/downloading-data/using-the-api/getting-started, JARVIS: jarvis.nist.gov and DFTTools: triqs.github.io/dft_tools/latest/documentation.html are genuine resources. Their existence does not establish a validated project workflow or available compute.
Re-entry: reproduce a known positive interface and discriminate negative controls with an appropriate many-body workflow; secure reference data, an experimentally plausible candidate family, a growth/measurement collaborator and a compute budget.
Portfolio recommendation, excluding SP-001
- SP-006: strongest combination of a precise observable, accessible archived artifacts and a bounded independent test.
- SP-008: real public data and a narrow residual-timing question, with substantial controls and access caveats.
- SP-002: reserve, especially if the team has mathematical certification expertise.
SP-005 is valuable but less launch-ready because circuit reconstruction, resource accounting and correlated-decoder fairness are substantial prerequisites. SP-011 should begin with inference, not full retraining.
Final count: five WORTH recommendations for explicitly bounded pilots; six NOT WORTH recommendations with re-entry conditions. Neither verdict establishes a scientific result or publication probability.
SP-006 outline — Solar-wind helicity
Question
Does a predefined opposite-signed spectral-peak statistic survive independent raw-data processing and qualified held-out PSP intervals? Secondary: does normalized peak position retain a radial trend after controlling for local plasma conditions?
Verified inputs
- Paper v2: arxiv.org/html/…, revised October 13, 2026.
- Zenodo v0.1.5: zenodo.org/records/22779133, a listed 20.8 MB archive with plotting code and extracted figure data.
- Author repository: github.com/Institute-of-Theoretical-Physics/helicity-evolution_psp, which documents missing intermediate files and non-turnkey raw preprocessing.
- Official FIELDS archive information: fields.ssl.berkeley.edu/data and NASA’s PSP_FLD_L2_MAG_RTN record: catalog.data.gov/dataset/psp-fields-fluxgate-magnetometer-mag-magnetic-field-vectors-radial-tangential-normal-rtn….
- PySPEDAS PSP support: pyspedas.readthedocs.io/en/stable/psp.html as an independent access route.
Access is verified at the documentation/archive-metadata level. Individual raw CDF retrieval, exact version coverage and end-to-end execution have not been tested.
First gate
Proposed cap: two analyst-days. Establish a file manifest covering versions, checksums, intervals, cadence, coordinate conventions and quality flags. Recreate selected published spectra from raw CDFs, not archived plot arrays. Pass synthetic circular-polarization and sign-convention checks.
Stop if raw versions, interval selection, instrument corrections or paper-to-data lineage cannot be reconciled. That outcome is an access/reconstruction failure, not a refutation.
Bounded study
- Freeze estimator, frequency bands, exclusions and primary peak statistic before held-out analysis.
- Control field polarity, sampling direction, mean-field angle, stationarity, instrument noise bands, sampling changes, gaps, interpolation, anti-aliasing and current sheets/switchbacks.
- Assess where the Taylor hypothesis is credible.
- Apply real sampling masks to synthetic signals to quantify artifact rates and detection power.
- Use nonoverlapping units and uncertainty clustered by interval/encounter, not independent frequency bins.
- Report effect sizes, uncertainty, exclusions and failed detections. Specify multiple-testing treatment before inspecting extensions.
- Begin with another PSP encounter. Cross-mission Solar Orbiter comparison is a later calibrated extension, not an assumed interchangeable dataset.
Decision rule
Advance only if raw reconstruction and synthetic controls pass, and held-out coverage supports the registered detection threshold. Report attenuation, replication or inconclusive evidence according to those criteria. Do not interpret a single-spacecraft result as proof of causal inverse transfer.
Deliverable
Versioned raw-file manifest, independently implemented analysis, control results, held-out spectra, uncertainty estimates and a discrepancy log. A useful replication artifact is achievable; publication remains contingent.
Cost uncertainty
The proposed access gate is two analyst-days, followed by roughly one to two weeks of bounded analysis if it passes. These are planning caps, not measured estimates. Raw transfer volume, preprocessing effort, runtime and specialist-review costs remain unbenchmarked. No paid compute has been priced or authorized.
SP-008 outline — Brain–LLM layer timing
Question
Does IFG layer-to-latency ordering retain incremental predictive value beyond a jointly modeled acoustic, lexical and positional baseline?
Verified inputs
- Goldstein et al.: nature.com/articles/s41467-025-65518-0, published November 26, 2025.
- GPTology snapshot 7e76c49: github.com/DVader96/GPTology/tree/7e76c49333a0…, containing electrode-file entries, ROI information and GPT2-XL embedding entries. An inspected embedding entry is an LFS pointer, not a verified downloaded payload.
- Podcast OpenNeuro v1.0.2: doi.org/10.18112/…, described in the Scientific Data paper: nature.com/articles/s41597-025-05462-2.
- Official dataset tutorial: hassonlab.github.io/podcast-ecog-tutorials/html/notebooks/01-qualitychecks.html, documenting audio, aligned transcript and associated neural/linguistic features.
Novelty boundary
The original study already has several controls. Hadidi et al.: nature.com/articles/s41467-026-72253-7 address generic confounds elsewhere, and Azizpour et al.: elifesciences.org/reviewed-preprints/110085v1 reanalyze pre-onset prediction on this dataset. The remaining question is specifically residual timing across layers, not whether adding acoustic features or variance partitioning is new.
First gate
Proposed cap: one analyst-day to retrieve payloads and check hashes, participant/ROI mapping, word exclusions, timestamps and sampling. Resolve the eight participant directories in the code tree versus nine participants reported overall. Do not assume OpenNeuro and GPTology preprocessing are identical.
Stop if stimulus alignment and sample equivalence cannot be established. Then reproduce original IFG curves and the timing effect within a preregistered numerical tolerance before interpreting any extension.
Primary analysis
- Use all words as primary; original predictability subsets are secondary.
- Compare nuisance-only and nuisance-plus-layer models using acoustic spectrum/envelope, onset/rate/duration, lexical frequency, surprisal, position, phonemes, static embeddings and previous-word features.
- Use nested time-blocked validation, boundary buffers, training-only transformations and matched regularization.
- Retain interpolation controls. Relabeling layers alone is not a competing mechanistic model.
- Primary endpoints: held-out incremental predictive performance and residual peak-latency slope across normalized layer depth.
- Estimate uncertainty by participant and contiguous stimulus blocks. Use autocorrelation-preserving nulls that repeat the entire lag-selection procedure.
- Report weak or unstable peaks rather than forcing a latency estimate. Account for endpoint multiplicity.
Decision rule
Stop after the preregistered analysis matrix. Require both incremental prediction and residual ordering to satisfy the registered uncertainty/null criteria before claiming robustness.
Attenuation can identify a limitation but does not prove the original effect was meaningless. Nuisance adjustment may also remove genuinely shared signal. Persistence does not establish identical brain/model computation or a causal transformer mechanism.
Deliverable
A timing-specific reproducibility and robustness report, frozen configurations, data-lineage record, uncertainty/null results and an executable benchmark. Publication depends on novelty and information gained.
Cost uncertainty
The authors report approximately 2.5 hours for core encoding on four CPUs/30 GB RAM; embedding regeneration used one GPU, four CPUs, 192 GB and six hours. These are historical allocations, not our quote. Reusing verified embeddings may avoid regeneration. Payload transfer, preprocessing reconciliation and repeated null fits could dominate. No local runtime, storage footprint or compute price has been established.