Known limits

What rang-core does not do, and where it is known to differ from the tools it was checked against. Each item links the report that measured or decided it (owner decisions: research/finish-plan.md section 7). No novelty is claimed: rang-core is checked for parity with CASA, QuartiCal, Tricolour and WSClean, not shown to beat them.

Important

Acceptance on real data, as of 2026-11-18 (research/v030-readiness.md). E2-B (LUNO, full polarisation) and E2-A (ESO 137-006, Stokes I) have reported: the gates they miss are listed below and in CHANGELOG.md (research/e2b-misses.md, research/e2-acceptance.md 9.13). E2-A is not passed as frozen; its gate 3b is a limitation with its cause (research/e2a-3b-k.md). E3 (CARACal’s shipped MeerKAT strategy on the same data) has measured E2-B (CARACal 6,926 s against rang 6,195 s; branch e3-caracal) and was still running on E2-A at the release (0.3.1 reports it). E2-N (native data, issue #7) is not met as frozen: 17 of 25 rows (research/e2n-native.md). Absolute polarisation angle is stated as a limit, not solved (below).

Scope

  • MeerKAT L band only. UHF and S band are not validated: a stage given channels outside 856-1712 MHz runs with one warning that names this page. Data from other telescopes runs with a warning, untested (research/meerkat-scope.md).

  • Native MeerKAT data: tested once, not met as frozen. E2-N (SARAO 1735268787, L band, observatory-phased) runs end to end but meets 17 of 25 rows: 3a (73 flag cells), Xf p95 (2.085° against 2.0°), the K/G/B composites, 3c, the J1744−5144 χ and the target’s compact scatter (13.3 % against 1.5 %) miss. The data are observatory-phased, so K, KCROSS and Xf are near-null and their real-size solve is untested. The CASA reference flagged 19 of 62 antennas in its delay solve because the oracle gave CASA the post-extendflags flags, not those rang solved on, so the K-downstream misses are partly an oracle defect; a corrected rerun is in progress. The polarisation claim still rests on LUNO (processed data) plus simulation (research/e2n-native.md, issue #7).

  • No direction-dependent calibration and no facets.

  • No learned flagger. rang-core flag runs the Tricolour port only (research/f3-flagger.md).

  • GPU paths need a source build with the core’s gpu feature; the wheels are CPU only (Installation).

Flagging

  • The Tricolour port differs from Tricolour on 530 cells of A1 (413 of them from the port’s float64 transform against Tricolour’s complex64 pocketfft, about 2 ulp at a threshold); accepted, not ported exactly (research/e2-acceptance.md).

Calibration

  • Reference antenna. CARACal’s polarisation strategy 1 is kept with its limit: on A1’s truth every reference antenna but m000 makes the X-Y phase drift 13-22° rms (research/e2-acceptance.md).

  • K offsets are not written to the K table; G absorbs them, so a K table applied alone lacks them (research/c2c-bandpass.md). The delay error estimate is low by a factor 1.5-2 (research/c2b-delay.md).

  • Solution TIME differs from CASA’s by up to 0.054 s (0.41 s with combine); it is reported, not gated, until CASA’s centroid weighting is ported (research/c2a-gain.md).

  • Absolute polarisation angle χ is a stated limit, not gated. On LUNO’s 3C 286 the RM-corrected χ(λ²=0) is 29.4°, 3.6° from 33°, rang equal to CASA (research/chi-offset.md section 11). On LUNO’s calibrator product use FEED’s stored receptor angle 0 (leave feed_angle_rotation unset): the fitted angle is +8.2°, which excludes CARACal strategy 1’s −90; −90 suits stores holding −90 such as ESO137. On held-out 3C 286 simulations rang’s Xf differs from CASA’s by 0.054° (median; issue #33). The cause is CASA’s calwt weighting of the Xf solve; the opt-in calibrate.xf_calwt: true takes the gap to 0.012° (research/xf-gap.md; the other solve types: research/calwt-terms.md). E2-B’s 1830−211 χ, transfer χ and 0.30 % leakage misses are in research/e2b-misses.md sections 5-7.

  • Parallactic angle in memory only. feed_angle_rotation is applied in memory and FEED is not rewritten: CASA reading a rang-calibrated MS whose FEED holds 0 applies a different P with parang=True (research/c6w-polwire.md section 7).

  • Pre-averaging, off by default. calibrate.average_time and selfcal.average_time average in time before a solve, as CASA does, which differs from CubiCal’s estimator. Gpol, KCROSS, Xf and Df with the parallactic angle applied never average across scans or more than 5° of it (research/xf-avg-fix.md); calibrate.average_freq is CASA’s frequency pre-averaging of the G, Gsec, F and Gpol solves (K, B, KCROSS, Xf, Df and selfcal unchanged; research/e2-average-freq.md) (research/e2-acceptance.md). calibrate.casa_phase_weights (needs average_freq) gives the calmode p solves (Gpol) CASA’s phase-only weighting instead of QuartiCal’s (research/e2b-misses.md 9.26).

  • Options refused for now: solnorm: true, calmode: a, gainfield on pre-applied tables, complex terms mixed with K, G, B, A, F; complex gains are never transferred to a target (rang-core config check --all CONFIG lists what is refused).

Self-calibration

Imaging (GeMClean)

  • The imaging floor, in part unexplained. On the real pilot the residual floor is 40-60 × thermal in every imager tried, WSClean included. Steady antenna gains are its main cause: on floor-3’s subset one phase-only self-calibration round takes the floor from 1,073 to 379 µJy (share 0.88 under floor-3’s standard; floor-4 gives 0.83 under imaging-floor’s full criterion). Of the remaining 379 µJy, 8.6 × thermal, deconvolution depth and spectral modelling are confirmed on a held-out scan (D_h 0.201 and 0.474), and masked Högbom to 0.3 mJy after the round explains 0.717 of the held-out excess; the rest is not explained, and no default changes yet (research/floor-3.md, research/floor-4.md, research/floor-5.md, research/floor-6.md, research/imaging-floor.md).

  • Memory is 2.0-2.4 × WSClean’s on the pilot at 512 channels in Stokes I (research/g7-defaults.md). Joined IQUV at 2048², 8 channels on E2-B’s fields peaks at 7.7 GiB on the large fields and 5.7 GiB on J1833−2103: -reorder is the default with -join-polarizations, and writes temporary files beside -name (or in -temp-dir), about 28 GB on the large fields, which a killed run leaves behind. -no-reorder keeps everything in memory (16.5-16.9 GiB on the large fields, about 5-8 % faster) (research/gemclean-memory.md, research/gemclean-memory2.md, research/gemclean-memory3.md).

  • Multiscale over-estimates compact sources: a median +9 to +18 µJy on sources ≥ 20σ_ap on DF5-O’s field, in WSClean 3.7 and GeMClean alike; no tested setting fixes it on both fields. Automatic scales come nearest; Högbom is unbiased at 7-11 % higher residual RMS (research/multiscale-flux-fix.md 8.1).

  • A FITS mask with multiscale confines the component centres, not the flux, as in WSClean 3.7 and radler. The scale shapes of centres inside the mask reach outside it. With extended emission outside a mask of small islands, the large scales take that emission up: after one major iteration on a simulated sky, 2.7 Jy lies outside a 3,394-pixel mask in GeMClean and 2.9 Jy in WSClean. Cleaned deep (about 1.6σ), both imagers can diverge. When nothing may be cleaned outside the mask, use Högbom, cap the scales, or erode the mask (research/gemclean-ms-mask.md, issue #35).

  • Joined polarisations peak on the linear sum, as WSClean 3.7 and radler do: with -pol QU -join-polarizations, components whose Q and U cancel are not cleaned (research/g9-stokes.md, research/v030-readiness.md section 4).

  • Cleaned images against WSClean (gate 3c): compact-source scatter 1.18 % on E2-B with the mask (bound 1 %), and 3.6 % on E2-A, where 512- against 4,096-channel smearing lowers the ratios with radius (research/e2b-misses.md 9.29, research/e2-acceptance.md 9.14).

  • Images are not byte-identical across platforms: native Windows and Linux builds differ in every file, and no tolerance across them has been measured (research/g4-masks.md).

  • Refused options include -fit-spectral-log-pol, squared channel joining, XY, YX and circular polarisations and -link-polarizations (research/g3-spectral.md, research/g9-stokes.md).

  • gemclean writes the model column by default, as WSClean does; pass -no-update-model-required when imaging a shared MS. rang-core image writes MODEL_DATA only with image.write_model: true (research/g6-predict.md).

Data safety and platforms

  • Table lock on Linux and Windows only. Linux takes casacore’s fcntl lock, Windows a LockFileEx lock; on macOS nothing stops a second writer to the same MS, and on Windows a casacore process inside WSL does not see rang’s lock (research/ms-lock-windows.md). rang-core refuses a second command on the same run directory on every platform (Data ingest: Measurement Sets, zarr and HTTP range fetch).

  • Memory limit. data.memory_limit_gb is 8 GiB by default; flag, calibrate and apply read in chunks within it. A self-calibration Sp or Sa solve whose solution interval does not fit (the default solint: inf on a long scan) is streamed through a scratch file beside the MS, 92 bytes per unflagged visibility, with the same solution (research/selfcal-memory.md); Sd and the CASA-rule solves still refuse such an interval. Channel-chunked solves on whole-array data are slow: 30-52 min per arm in W3b, before the channel sliced read, which is 2.1-2.2 × faster on the polarisation sets (research/w3b-chainchunks.md, research/sliced-read.md).