Krishi Attri

Aligning independently captured 3D Gaussian-Splatting scans usually means falling back to point-cloud registration that throws away the splat structure. splatreg registers natively on the Gaussian representation (a Gaussian-SDF residual with a closed-form Jacobian over SE(3)/Sim(3)), then merges, or aligns without merging: the CLI bakes the recovered pose into the source so both scans stay separate PLYs in one frame. Baked-in transforms rotate the higher-order spherical-harmonic colour with the splat (real-basis Wigner-D); photometric refinement with exposure compensation handles the poses geometry cannot see; every builtin solve reports pose covariance for pose-graph weighting, never faked. Recall matches the GeoTransformer point-cloud baseline on official 3DMatch (91.5% mean / 93.5% pooled) and 3DLoMatch (72.5% / 74.4%) while adding the Sim(3) scale DoF the baseline lacks, and rotation lands 2.9× tighter than existing splat tools (5.2° vs 15.3°). The MAC maximal-clique seed handles contaminated correspondence sets, with the honest measured verdict kept: a wash on the official 3DMatch split, a decisive win on structured decoys.

official 3DMatch registration recall, 1279 pairs

91.5%

SH Wigner-rotation error vs an independent evaluator

2.4e-15

splat-merge Chamfer vs naïve concat

10.3 → 2.0 mm

photometric refine where geometry under-constrains

5° → 0.36°

A robot replanning through a world whose costs drift never knows how good its current route is once the map goes stale; classical planners silently trust the stale map. CERT-FLOW answers with a proof every round: a high-probability certificate LB ≤ OPT ≤ UB on the optimal route cost, built from age-weighted non-exchangeable conformal prediction over drift-adjusted residuals, and it spends paid sensing exactly where the certificate says the gap shrinks fastest. When the certificate proves the map tight, that proof licenses ns-to-µs preprocessed queries that self-expire the instant drift exceeds tolerance. Seven theorems (coverage through an impossibility result on lower bounds), validated on 17 synthetic regimes, game maps, and real traffic (METR-LA, PEMS-BAY); the failed hypotheses stay documented in the record.

coverage on every condition ever run; classical replanning 0.02 – 0.59

0.95 – 1.00

certificate-gated static cost query; 3.7 ms p50 full certified round, one CPU core

269 ns

lower sensing regret than freshness, uncertainty, or random at equal budget

2 – 3×

Feature caches skip the network on most denoising steps and forecast the cached features instead. HiCache++ ships the exponential (Dynamic Mode Decomposition / Prony) basis, exact on the local feature-ODE class where polynomial bases (TaylorSeer, Hermite) diverge, and the honest finding the benchmarks forced: no single forecast basis wins across diffusion families. The exponential basis wins on flow-matching 3D generators; polynomials hold DiT-class denoising. So the product is the selector: backend auto backcasts a held-out snapshot with both bases at every compute step and serves whichever demonstrably wins, at zero extra model calls. The same study surfaced a benchmark-integrity find: a one-character sign error that left a published baseline anti-extrapolative yet invisible to every end-to-end metric, answered with directional closed-form regression tests. The exponential-basis DMD calibrator is now upstream: it was merged into vipshop/cache-dit (PR #1053, 2026-06-14) and ships in cache-dit v1.5.0, behind SGLang-Diffusion, vLLM-Omni, and ComfyUI. Also deployed through per-model adapters across TRELLIS, Hunyuan3D, and SAM 3D, plus three GPU-validated ComfyUI nodes (Hunyuan3D, TRELLIS, TRELLIS.2).

F-score at skip-interval 5, exponential vs polynomial arm, Hunyuan3D-2.1

0.860 vs 0.735

geometry-lossless (F1 = 1.000) through interval 6, SAM 3D Objects

1.56×

holdout auto detects basis misfit and serves the winning arm

120/120

A robot policy can emit the right numbers through the wrong interface and move confidently in the wrong direction. ActionABI attacks that failure offline: it recovers undocumented action-tensor semantics from logged trajectories, retains calibrated equivalence sets, and refuses to emit a converter when the evidence cannot identify one. ActionShift attacks it online: frozen ManiSkill policies face hidden permutations, signs, scales, target modes, frames, lag, and gripper conventions, then adapt through bounded probes. The pair shares one grammar and one scorer, with ActionABI's C++ evidence core running inside ActionShift's belief loop.

ambiguous synthetic cases correctly abstained, zero false uniques

25 / 25

PickCube success, no adaptation to pool-belief adaptation

0.000 → 0.987

fewer probe steps for DualABI at matched success

52%

Language models hallucinate theorems, and prose is not verification. mathlas is an MCP server of 12 deterministic, data-returning tools an AI agent drives: search over a 3.68M-document index (dense + BM25 + rank fusion), PSLQ constant identification, OEIS sequence lookup, and real Lean-kernel checks that now verify full proofs, returning the kernel's error verbatim so the agent can repair and re-call. A quantized laptop tier serves the same index from 1.9 GB at 2.4 s/query on 4 CPU threads, measured recall-lossless. The discipline is airtight-or-nothing: every verification tier returns an independently checkable fact or an honest nothing, with zero false positives measured across all tiers.

the same agent on 18 math tasks, with vs without mathlas

18/18 vs 15/18

Hit@20 vs TheoremSearch, on its own 110 human queries

59.1 vs 45.0

documents indexed; dual-channel R@10 0.999; 1.9 GB quantized tier

3.68M