Each crystal's (psi_x, psi_y) tilt was refined by a Ceres problem rebuilt for every inner
iteration: one auto-differentiated residual block and one Cauchy loss per reflection. With the
scaling reference cleaned up it was 70% of the stills scale/merge. The same objective - Cauchy
on the sigma-weighted intensity residuals plus the Gaussian prior on the tilt, the tilt bounded -
is now minimised by a small Levenberg-Marquardt on the rho'-reweighted 2x2 normal equations,
the tilt applied to all reflections through one rotation matrix and the Jacobian taken by
central differences.
It reaches the same answer to optimizer precision: on the new synthetic test
(StillsPartialityRefine_RecoversTheTilt) partiality error 0.0559 -> 0.03019 with Ceres and
0.03016 with this, mean tilt 0.02065 vs 0.02067 deg; on two 25k-frame XFEL runs (--mode scale)
CC1/2, I/sigma, R_pim and the CC against a deposited lysozyme model agree to 0.0005, mean
|dpsi| 0.023 deg on both. Scale+merge 6.9 -> 3.0 s (--mode scale, loaded box).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The final stills merge drops every crystal whose cell is not close to the consensus cell
(MergeOnTheFly::Mask), but the scaling reference (MergeAll) and the partiality post-refinement
were made over all of them. On two 25k-frame XFEL runs of a tetragonal crystal those were 26-30%
of the indexed frames and 60% of the integrated reflections: weak frames (median 8 indexed
spots) indexed in a 2x/3x/4x supercell or an unrelated lattice, whose supercell-only reflections
are pure noise and whose indices do not name the true reflections. They went into the
reference every merged crystal was scaled against, and through the tilt refinement.
MergeAll and StillsPartialityRefine::Run now take the reference cell and use only the crystals
it accepts - the same test, with the same tolerances, as the merge.
Measured (both runs, mx path): CC1/2 0.9599 -> 0.9598 and 0.9645 -> 0.9644, I/sigma +0.05 and
+0.02, R_pim unchanged, space group and resolution unchanged; --mode scale against a deposited
tetragonal lysozyme model CC 0.9452 -> 0.9477 and 0.9461 -> 0.9473. Scale+merge 2.1x faster on
--mode scale (14.6 -> 6.9 s, loaded box).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
hide_gpus() in CUDAWrapper makes get_gpu_count() answer 0 without calling
cudaGetDeviceCount, so every GPU decision in the process (indexer Auto ->
FFTW, preprocessing/spot finding/decode, pre-scan, scaling engine, model
validation, MR) takes the path it takes on a machine without a GPU, and no
CUDA context is created. rugnux reads --no-gpu in main, ahead of the parser,
because set_gpu_blocking_sync() already asks CUDA for the devices. The
banner says 'CPU only (--no-gpu)'. The 'run it on the CPU' advice in the
out-of-GPU-memory messages and in the install docs now names --no-gpu
instead of CUDA_VISIBLE_DEVICES=.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The stills scale/merge after the image loop was almost entirely serial:
- every scaling-reference merge (MergeAll: the per-image scaling reference and the two
partiality post-refinement references) ran French-Wilson single-threaded on its export,
about 3 s each, and assigned free-R flags - neither is read by a reference, which uses I
only. MergeAll now exports intensities alone (MergeOnTheFly::ExportIntensities).
- the merge accumulated into a std::map; it is now an unordered_map exported in key order,
which is the std::map order.
- AddImage, the error-model collection and the R_meas pass of MergeStats did the
per-observation work (systematic-absence test, ASU key, filters, error-model sigma,
outlier test, merged-I lookup) in one serial loop. They now reduce a block of images in
parallel and consume it in image order (ReduceImagesInOrder), so every floating-point
sum, half-set choice and hash-map insertion is made in the serial order.
Measured on a 25k-frame XFEL stills run (box loaded, load 9-14): merge P1 19.0 -> 7.1 s,
final-group merge 27.7 -> 13.2 s; on a second run 23.4 -> 7.7 s and 29.7 -> 13.6 s.
mtz/unmerged mtz data, hkl, cif and report identical on both runs, mx and --mode scale.
New test MergeOnTheFly_AddImagesIsTheSerialMerge checks the parallel merge against the
serial loop across a block boundary.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The stills per-image loop spent 99% of its time indexing (28.4 of 28.7 ms per frame on a 16 Mpx
serial set; read, decode, spot finding and integration together 0.33 ms), and only four threads
were ever busy:
- The 16 image workers each block on IndexerThreadPool, which had the default 4 threads, while the
FFT candidate post-processing is CPU work. Stills now size the pool to min(-N, 8); 8 gets 94% of
what 16 does (83 / 125 / 133 frames/s at 4 / 8 / 16 with the changes below) at half the card
memory - each GPU FFT indexer holds ~0.5 GB, 16 of them took the run to 12.7 GB.
- FFTIndexer::ReduceResults (widened path, taken on ~2/3 of calls) deduplicated ~2900 candidate
cells per call by a linear is_close scan: 46% of all CPU cycles. The accepted cells are now kept
ordered by a and only the 2%-window of a is asked; the answer (any close cell) is the same.
- FilterFFTResults built a 16k-entry multimap per call (11% of cycles); now a partial sort of
indices in the multimap's reverse order (magnitude descending, later direction first on ties).
- calculate_fft_result reads magnitudes from a bin-major transposed copy (in the histogram buffer,
no new memory): neighbouring lanes no longer read a spectrum apart. Bit-identical results
(checked against the old kernel on 4000+ calls, 0 mismatches); not faster on an idle card, but
67 -> 98 frames/s with the pool's streams sharing it.
- The stills min-pix choice indexes a frame at 3/2/1 and then the winner again; the seed of
strongest spots is often the same list each time (31% of indexer calls repeated an earlier call
of the same frame). IndexAndRefine answers a repeated list from a per-frame FrameIndexerCalls
memo kept by MXAnalysisWithoutFPGA for the duration of one frame.
Measured (RTX 5080, 32 threads, data in page cache), first 2000 frames, ABAB: 35/35 -> 127/127
frames/s, p.hkl md5 identical. Full 24960-frame run: loop 1081 s -> 277 s, wall 1132 s -> 328 s,
p.hkl md5 identical (P41212, 33234 unique); with a pool of 16 the loop is 222 s (the full run has
more spots per frame than its first 2000 frames, so the CPU side weighs more there). With a known cell (ffbidx path) 715 -> 920 frames/s on
2000 frames, p.hkl identical. Tests: jfjoch_test [Indexing] pass.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
- MolecularReplacement: an engine whose making throws anything (std::bad_alloc, a gemmi error), not
only a JFJochException, is refused and its reservation given back, so a thread waiting for the
same engine is woken instead of waiting forever.
- RotationScaleMergeGPU::Alloc waits for the GPU work beside itself (GPU_BUSY_TIMEOUT), so its
buffers are allocated plainly Synchronous; SynchronousAfterWorkBeside added a second 10-minute wait
in front of it.
- CPU_DATA_ANALYSIS_INDEXING: an offline run decides its near ties on the whole sweep from the start;
only a live run (broker URL or --simulate-live) decides early and re-checks.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Last-bit change, not bit-exact: every partial whose frame does not contain its exact diffracting
condition gets a partiality - and so a slice-centroid walk along its Debye ring - that differs in
the last bits.
The partiality was (erf(b) - erf(a)) / 2 in float. Where the frame's slice lies in one tail of the
rocking curve both erf are near +-1, and their difference keeps only a few ulps of 1: glibc erf and
CUDA erff, one ulp apart, gave partialities ~1e-3 apart (relative) at p ~ 3e-5. The slice centroid
(RockingSliceCentroid_rad) divides by that partiality, and at zeta ~ 0.05 the walk it scales is
hundreds of pixels, so CPU and GPU placed such a reflection up to 0.1 px apart -
BraggPrediction_CPU_GPU_consistency_large_cell failed on one of them once --fmad=false (7ed970503)
moved the ulps. A slice in one tail now takes the same difference from erfcf, which keeps its relative
precision there; one helper (RockingSlicePartiality) serves both predictors.
Reflections placed more than 0.01 px apart by CPU and GPU in that test: 65 -> 0. BraggPrediction*,
BraggPredictionRot*, BraggPredictionRotGPU*, RockingSlice* pass. 7qij image passes 5.68/7.90 s ->
5.26/7.50 s (one lock hold; no cost).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
Molecular replacement's structure-factor score engines were made one per group, copy count and
resolution, each admitted against a quarter of the card's total, and all kept to the end of the
search. On a large cell they summed to well over the card: traced with nsys on the rotated 4nwv
model (which failed on every run, base included), seven engines of 1.36 GB were held beside the
two rigid-body pools (5.9 GB) when a 288 MB buffer did not fit.
Each engine is still admitted, or the search scored on the CPU, exactly as before, on the card's
total memory. Now the engines kept on a card also stay within that quarter together: an engine
that does not fit beside them takes the place of the least recently used one no score is using,
and waits for a score to finish where all are in use; one taken out is made again when next asked
for. A score holds its engine for as long as it uses it (shared_ptr lease). Engines compute the
same numbers whenever they are made, so only when the card holds them changes.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
The intermittent "out of GPU memory" on the 3600-frame cbf.gz set (about one run in three on a 16 GB
card, base included) was caught with allocation backtraces: FitErrorModelGPU's 42.8 MB in
MergeAndStats failed while the speculative geometry probe (GPUWorkBeside) held its FFT indexers on
the same card. The engine's own arrays (RotationScaleMergeGPU Impl::Alloc) wait for such work to end
and ask again; the merge's helpers - ErrorModelGPU, WilsonOutliersGPU, PairSortGPU and the cub
scratch of the engine itself - allocated directly and failed at once.
CudaAlloc::SynchronousAfterWorkBeside: Synchronous, and where the card is full after the pool is
trimmed, it waits for the GPU work beside the main line (wait_for_gpu_work_beside, 10 min as for
the engine) and asks once more. The four merge files use it. Only the probe passes register
GPUWorkBeside and they index only, so none of them merges and nothing waits for itself. When a
buffer is allocated is all that changes; what is computed is not.
Catch2 CudaDevicePtr_AfterWorkBesideWaitsForItsMemory: with the card filled by work beside, a
Synchronous buffer fails and a SynchronousAfterWorkBeside one gets its memory once the work ends.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
A screw candidate whose absent class no reflection contradicts, refused only because its
evidence could not reach min_screw_absence_evidence even with every predicted-absent axial
reflection at the floor (one such reflection reaches 6.9 nats; a row without a control class
reaches nothing), left the symmorphic group adopted and the report saying the group follows from
the absences. Such a row is now SPACE_GROUP_SCREW_UNDETERMINED, with the screw group in
SPACE_GROUP_ALTERNATIVES. Report only: the result is kept apart (unreached_screws) and folded in
after every decision, so the written group and files do not change.
Measured on a monoclinic set whose merge holds five 0k0 reflections at noise level, one of them
predicted absent: P2 was reported as decided by the absences against a deposited P21; it is now
P2 with the b screw open and P21 offered. Over the full battery's candidate tables the condition
holds on that set and one private set (the same P2/P21 case) only. ci tier: no change.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
- All 84 DOIs checked (82 on Crossref, 2 DataCite): title, journal, volume, pages and first author
match; Matthews 1968 and Yeates 1997 links percent-encode their parentheses; the MOSFLM link
points at a page that answers.
- autoPROC: an explicit acknowledgement of Clemens Vonrhein (Global Phasing) for the principle that a
pipeline should verify the metadata it is given - beam centre, geometry, goniometer - and correct
it where it is wrong; Vonrhein et al. (2011) Acta Cryst. D67, 293-302 (DOI checked).
- Zhang, Sauter et al. (2006) DISTL, already credited in SpotFindingSettings.h, gets its entry.
- Molecular replacement: tNCS looked for in the native Patterson as Read, Adams & McCoy do; the same
credit at DetectTranslationalNcs (comment only).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
- CPU_DATA_ANALYSIS_IMAGE §0: CBF byte-offset and marCCD widening are decoded on the card; the host
decoders are the CPU build's path.
- CPU_DATA_ANALYSIS_DECISIONS §14: the null runs for every model the rigid body is tried on; the
device choice is made once on the card's total (structure factors and rigid body a quarter, scale
fits an eighth) and a CUDA failure ends the validation, deciding nothing; a subgroup the model fits
is written as <prefix>_mr; new §14.10 on molecular replacement (model kind, rotation and
translation functions, likelihood, Matthews, tNCS and P1 rules, null, what it may change).
- CPU_DATA_ANALYSIS: scope and references gain molecular replacement.
- CPU_DATA_ANALYSIS_INDEXING: the near tie decided early and re-checked once every frame is in.
- RUGNUX: merging is the default and _process.h5 optional; live following of a collection with a
broker URL; --model's MR fallback.
- RUGNUX_INSTALL, RUGNUX_OVERVIEW, RUGNUX_TUTORIAL, RUGNUX_REPORT: GPU work list, <prefix>_mr outputs,
MR in --mode mx for rotation and stills, rigid body only in --mode scale.
- RUGNUX_ADVANCED: -N default as the usage states it.
- battery README: the ci tier has 28 sets.
- ModelValidation.cpp: a comment still said a GPU failure reran the validation on the CPU (comment only).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
The ingest term was 90 bytes per reflection and 64 per run. Measured with cudaMemGetInfo at each
step of IngestUpload/IngestSort (relative to the free memory before the observation arrays):
- the second radix sort is the peak where runs are few: the 66 B of observation arrays, by_image,
perm, key, key_sorted and cub's alternate key/value buffers of a plain SortPairs - 102.3-103.0 B
per reflection on 7qij (51.1 M, 52.2 M, 78.1 M reflections) and 4nwv (32.7 M);
- the cut is the peak where runs are many: 98 B per reflection plus 56 per run (4nwv, 15.4 M runs:
125.8 B per reflection, 4106 MB against 3923 MB from the old formula);
- the upload holds the 66 B beside the two 64 MB staging buffers (~150 MB measured).
Now max(103 B per reflection + 58 B per run, 66 B per reflection + 160 MB): 2-4% above every
measured peak, where the old formula was 10% under on 7qij (7170 MB against 7986 MB for the main
engine) and 4.5% under on 4nwv. Its only consumer is whether the all-observation arm goes beside
the main engine, which changes when that merge runs, not what it gives.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
A UINT16_WIDENED frame (marCCD) was page-locked where it lay with PinInputRegion, which keeps the
registration until the engine moves on to another frame. The pre-scan hands its kept frames to the
spot measurement and to the defective-pixel search at once, each with its own engines, so one
engine could unregister a frame while another engine's copy from it was still in flight. That
takes the context down ("unspecified launch failure"), seen as a rare crash at the end of the
pre-scan on a 3072x3072 marCCD sweep (1 run in 30 reproduced it).
The stored bytes now go through the engine's own page-locked staging buffer, as the CBF and BSLZ4
decoders already do. The pixels decoded are unchanged.
Test: ImagePreprocessorGPU_SharedStoredFramesManyEngines - four engines upload four shared marCCD
frames concurrently; it aborted with the launch failure before the change and passes after it.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
The first device decoder walked 256-byte tiles, one thread each, straight from global memory: on a
PILATUS 6M frame (~6 MB of stream) that is ~24k threads walking alone, and the two walks took
~0.5 ms of the card per frame - with the CBF decode gone from the workers the card is what bounds
the 7qij loops (kernels 67-78 % busy). Now each thread walks a 68-byte part that its block has read
into shared memory together (68, not 64, so the threads of a warp do not read the same banks), the
scan runs over tiles of 8 parts whose maps are composed first, and the write kernel stages a block's
pixels in shared memory and writes them out coalesced.
Measured on 36 real frames of 9 CBF sets (nsys, average per frame): 0.41 ms against ~0.6 ms before.
Output unchanged: the same tests, and the same 36 frames identical to the host decoder.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
Every merge downloaded the fulls' frame and group keys and built both CSRs by serial counting
sorts on the host - four passes over the fulls, two of them scattering into an array as long
as the distinct hkl (15 M on 4nwv) - and uploaded them again. They are now stable radix sorts
of the resident fulls on the card, with the counts summed there; the host gets the counts per
frame and the group CSR it walks itself. The same permutations (stable sorts list each frame's
and each group's fulls in index order, as the counting sorts did): p.hkl identical on cytc
and 4nwv.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
On a set too large for the quality guard's engine to sit beside the main one (8qaw: 123 M
observations; 8a1a), the main engine was ingested, found not to leave room, dropped and
ingested again after the guard - several seconds of ingest for nothing. The pre-pass's
engine merged the guard's very frames earlier in the run, so its card memory per reflection
is now kept (prepass_engine_bytes_ / prepass_engine_reflections_) and both engines are
reckoned from it before either ingests: the guard at its frames' reflections, the main
engine at all of them (observations add in proportion; rocking events and distinct hkl at
most in proportion). Beside where the two fit the card less the 3 GB reserve, otherwise
the guard first. The re-ingest path is gone.
7qij, 9v16, cytc: beside; 8qaw, 8a1a: guard first. Same merges, same numbers.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
The rest of the resident ingest moves to the card: the delta_phi re-solve against each
frame's smoothed lattice, the one exact-Bragg angle per rocking event, the partiality
recomputed from the smoothed mosaicity (with corr brought to it) and the histogram of the
frames an event spans. The host keeps the per-frame work (the smoothed lattices, the
mosaicity window) and reads back only corr; nothing per observation crosses back, and the
host arrays the smoothing worked on are no longer built on this path.
The integer work is the host's to the bit (the event-span histogram). The floating-point
work differs in the last bits: every rounding on the card is explicit (no contraction, so
the same bits on any GPU), the host's compiler fuses some of the same products into sums,
and erf/atan2 come from the CUDA library. This is the one place in the lever where results
change; see the hand-back report for the sets and metrics it moves.
Measured on 7qij (main engine, 78 M reflections): the whole ingest 6.5 s host + 1.3 s device
before this lever -> 2.9 s, of which 2.4 s is the upload (memory-bound, beside the
post-refinement gather) and 0.25 s this smoothing. Same space group, error model,
resolution and reflection count; IMEAN differs at the 1e-7 level except on near-zero
intensities.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
With the GPU resident, the host ingest swept the source reflections four times (background
and h histogram, sort keys, the in-range build, the device staging), bucket-sorted the keys
and cut the raw-hkl runs: 5-6 s of host time per engine on a 78 M-reflection sweep, run
twice beside the quality guard. Now the frames' reflections are copied once, through two
pinned buffers, to the card, where the per-frame background is summed, the reflections
are ordered by two stable radix passes ((image, index), then raw hkl - the one order the
host's total comparator gives), cut into runs by run-length encoding, filtered by the
resolution range and compacted. The observations, the runs and their order are the host
ingest's to the bit; a new GPU test checks the layout against the host's sort and cut.
The geometry smoothing still runs on the host in this step, on arrays read back.
The all-observation arm's engine is cloned after the main engine's device ingest (its
DeviceBytes are known only then) and copies the runs on the card as well. DeviceBytes now
also covers the ingest's own peak (90 bytes per reflection, 64 per run before the cut).
Measured: p.hkl and MTZ data identical on cytc and 7qij.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
The marCCD reader read each frame's 16-bit scanlines into scratch and widened them to int32 on the
worker thread (marccd::ReadInto, ~20 % of the image-loop worker CPU on 4nwv), and the 4-byte image was
uploaded. ReadRawImage now reads the scanlines straight into the image's buffer and hands them over as
a new read-only CompressionAlgorithm::UINT16_WIDENED (mode Int32): two bytes a pixel cross PCIe, from
page-locked memory, and the preprocessing kernel widens them itself.
It is not CompressedImageMode::Uint16, which calls 65535 an error pixel; a marCCD's 65535 is its
saturation marker. So the kernel compares each pixel with int32's error marker and the int32 saturation
cap - the comparisons are made on the pixel's value, in int64, rather than in the storage type - and
decides every pixel exactly as for the int32 image the host would have widened it to. The host widening
(JFJochDecompressPtr) is the CPU path.
Test: ImagePreprocessorGPU_StoredEncodingsMatchCPU, with a frame carrying 65535 at a saturation of
65535, against the CPU preprocessor on the host-widened image.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
The CBF reader decoded every frame's x-CBF_BYTE_OFFSET stream on the worker thread - 19-30 % of the
image-loop worker CPU on CBF sets (7qij, 4mtk) - and the 4-byte image was then uploaded. ReadRawImage
now hands over the stream as the file stores it, in the image's own buffer, as a new read-only
CompressionAlgorithm::CBF_BYTE_OFFSET (mode Int32, what it decodes to), and ImagePreprocessorGPU
decodes it on the card into the same int32 image and runs the same preprocessing kernel on it.
The device decode (CBFDecoderGPU) cuts the stream into 256-byte tiles. A token is 1, 3, 7 or 15 bytes,
so a tile can be entered at 15 offsets; for each one thread works out where the decode leaves the
tile, how many pixels it emitted and their sum modulo 2^32 (the walks from different entries meet
within a few bytes, after which they are one walk). An exclusive scan composing these maps gives each
tile its true entry, first pixel and running value, and a second walk writes the pixels. A stream
that ends before the image is refused as on the host.
The host decoder moves to compression/ (DecodeCBFByteOffset) and is reached through
JFJochDecompressPtr, so every host consumer - the CPU build, the azimuthal-integration worker, the
pre-scan's shadow and gain passes - decodes exactly as before, now from the stream. It is the CPU
path, not a fallback: a frame the device refuses is not retried on the host. A process file written
from CBF input stores the decoded int32 image, as it always did. .cbf.gz is still inflated on the host.
Tests: CBFDecoderGPU_MatchesHost (every token size, 0x80 inside payloads, escape runs across tiles,
trailing CIF text), CBFDecoderGPU_ShortStreamIsRefused, ImagePreprocessorGPU_StoredEncodingsMatchCPU.
Checked locally against the host decoder on 36 real frames of 9 open-arm CBF sets (signed 32-bit,
signed and unsigned 16-bit, plain and gzip): identical.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
The rotation predictor used to download every predicted reflection into pinned memory, copy it into
the host vector, build sort keys, std::sort them, gather the 88-byte structs into scratch and copy
them back (OrderOutput), and - on a frame over output_limit - partial_sort the structs themselves
(TruncateToOutput). On a large cell that is a quarter to a third of the image-loop worker CPU
(profiles of 4nwv and 7qij: TruncateToOutput 12-16 %, OrderOutput 9-13 %, plus the memmoves).
Now the same order is built on the device with stable cub radix sorts of an index: delta_phi, then
the packed hkl (OrderOutput's h, k, l, delta_phi), then - over the cap only - partiality descending
and excitation error ascending (TruncateToOutput, whose hkl tie-break the stable sort keeps). Only
the kept rows are gathered and downloaded. Floats are keyed by an order-preserving integer map with
-0 made +0, so the key order is the comparators' order; the only possible difference from the host
is between two solutions of the SAME hkl with bit-identical partiality and excitation error, whose
relative order partial_sort leaves to its heap and the stable sort to delta_phi.
The device buffers grow with half again of slack, so a frame predicting a few more than the last
record (the supercell probe's frames above all) no longer reallocates and predicts again; the host
side is sized for what is returned, at most output_limit.
Test: BraggPredictionRotGPU_DeviceOrderMatchesHost holds the device output against OrderOutput's
order and against TruncateToOutput applied on the host, at four caps.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
One --model now does the right thing by itself:
- A prediction (AlphaFold/ModelCIF, read from the file) goes straight to molecular replacement.
- An experimental structure is placed by the rigid body as before. The null now runs on every
--model run (no NOT_TESTED): "placed" = the fit beats its null (4.51 sd) in the answer's merge
or in the twin hypothesis' merge. Placed: behaviour unchanged (hand and reindex may go into
p.mtz); a subgroup the model fits and prefers is written as <prefix>_mr.
- Not placed (or no crystal cell): molecular replacement with the experimental-model path. It may
relabel the hand into p.mtz past the anomalous veto, never reindexes; another group or the twin
subgroup goes only to <prefix>_mr. Maps, R-free and CC come from validating the MR placement with
every decision switched off; MODEL_FIT stays the rigid body's verdict.
- The twin hypothesis' merge is made in memory with a model where a metric re-ask promoted the
group or the model's group has fewer rotations than the adopted one; written and reported only
where the model chooses it.
- --mr is kept as "--model, searched for by MR straight away".
- Report: MODEL_PLACEMENT= RIGID_BODY|MR|NOT_PLACED, MODEL_DECISIONS_TAKEN gains MR_ENANTIOMORPH,
MODEL_CHOICE_FILE_PREFIX/SPACE_GROUP; the MR keys move into 5. MODEL VALIDATION; MODEL_FIT absent
where no rigid body ran. MR NOT_PLACED is never a warning.
- Battery wrong-model step: check (c) also fails a wrong model placed by MR.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
The scale fit of every model-validation fit (fit_model: the real model, the indexing probe's relabellings
and every null replicate) was FitModelScale on the host - a long-double Levenberg-Marquardt per solvent grid
point, ~80% of the validation's CPU on a 400 A cell. It now runs on the GPU through ModelScaleGPU::FitSolvent,
the engine the rigid body and molecular replacement already use, from a small pool (ModelScaleGPUPool: up to
four engines with their own streams, sized on the card's total memory, at most an eighth of it). The grid,
the fits and the winner rule are FitModelScale's; the sums are taken in double on the device, so R-factors
agree with the host's to rounding. The CPU build, and a GPU build without a card, keep FitModelScale.
The change-of-basis probe gridded each candidate frame with gemmi's serial put_model_density_on_grid, once
for the data as merged and once more for each alternative indexing. It now grids each frame once, with
ModelGrid's parallel PutModelDensityOnGrid (the same grid, bit for bit), from threads outside the pool so the
gridding spreads over it, and scores every labelling on that F_calc. Exact.
4nwv (122k atoms, I222, 400 A cell): model validation 33.2 s -> 14.0 s; every logged R, sigma and decision
the same.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
- The option is --mr <file>. What kind of model it is is read from the file: a prediction where it
is ModelCIF (_ma_qa_metric) or has no cell and one B value per residue in 0-100 (AlphaFold PDB/mmCIF);
an experimental structure otherwise. A prediction is handled as before (pLDDT >= 70 trim, B from
pLDDT, PAE domains where the AlphaFold DB's PAE file sits beside it). An experimental structure is
taken as its polymer - no waters, ligands, hydrogens or alternative conformers, occupancy 1, isotropic
B as refined - with a coordinate error of 0.67 A (a prediction's at pLDDT 90) for sigma_A; a B column
above 100 no longer stops the search.
- The report section is MOLECULAR REPLACEMENT (EXPERIMENTAL), its keys MR_* (MR, MR_MODEL_FILE,
MR_MODEL_KIND, MR_DECISION, MR_SPACE_GROUP, MR_COPIES, MR_HAND_CHECK, MR_SECONDS, ...); the <prefix>_mr
outputs are unchanged.
- The twin hypothesis' merge that is made only for the model to test (a metric re-ask promoted the
indexed lattice's group) is written and reported only where the model chooses the subgroup.
Raw deposited coordinates of nine control sets, unconverted: every one placed and confirmed against the
deposition (all chains as one body), the groups the same as with the prediction or the prepared model.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
From the nsys/perf profiles of the copy chain (GPU busy 39% between copies, the thread mostly in host
work and syncs):
- The known-structure translation function's lists, device arrays and cuFFT plan are made once per
group (MolecularReplacementGPU::Prepare) instead of on every copy (~260 MB of uploads and a plan
per call before).
- The known copies' CopyFactors and likelihood, which rank the peaks, are taken while the GPU runs
the translation function.
- The structure-factor engines are made per entry, without one lock held across all groups' creation.
- The transform's index wrap divides only for an index outside the grid (CPU and GPU).
All exact: same arrays, same kernels, same order. 4mtk MR, alternating 3 pairs: 9.5 -> 9.0 s.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
Drops the rule of af796d0a6 that the merge's own group searched to fewer copies always leaves the
choice UNDECIDED: it also caught a correct SPACE_GROUP decision (a target set; the half-vector move
check already catches the wrong one it was meant for). Instead, the chosen copies carried into a
tNCS-related group rule that group out only where the chosen placement fits the data itself
(R-free below 0.50): a perfect deposited model placed in the wrong group at R-free 0.520 had ruled
out the deposited group by 0.036; it is UNDECIDED now. The carries that decide in the survey all
start from R-free 0.36-0.47 and lead by 0.04-0.10.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
Completes c6912e2a0 (whose message was a placeholder). What the two commits do:
- tNCS by half a lattice vector v: copies v apart weaken the reflections with h.v half-integral, and a
placement moved by (v + o)/2 (o an origin shift of the group) fits the strong ones exactly as well.
The candidate the choice falls on is compared with every such move of itself (rigid body refined,
packing) by R over the weakened reflections only; a move better by 0.01 is taken instead, and where
the two best are within 0.01 the merge is UNDECIDED.
- With tNCS, the group the merge was made in, searched to fewer copies than the chosen group, keeps
the merge UNDECIDED (not for its enantiomorph, which the hand check decides): its first copy alone
can sit at half the vector, and no placement of the chosen group, which may sit there itself, rules
it out.
- P1: the first copy sits at the origin, which no function of its position can single out (the
translation function is flat), so one copy stands out where its rotation-function Z reaches 8; the
null then judges it as before.
Checked against the deposited coordinates (placement comparison, symmetry/origin/setting matched):
two half-vector-shifted placements that had been decided (a SPACE_GROUP in the wrong setting, an
AGREES off by a/4) are UNDECIDED; two UNDECIDED tNCS sets are now AGREES with confirmed placements;
the nine confirmed tNCS placements and the ENANTIOMORPH stay; a P1 single copy with a perfect-match
prediction is placed (confirmed), a wrong PAE-domain placement in P1 (RF Z 6.4) is not.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
The null a placement must beat turned the whole assembly about its common
centroid, so with several copies it also moved them off the arrangement the
search found. Under tNCS, two parallel copies a tNCS vector apart fit the data
whatever the model is, and the null threw that fit away: wrong deposited models
reached TF Z 14-47 and 8-34 sd over the null on tNCS crystals, and two of them
decided the space group (SPACE_GROUP, a reflection file in the wrong group).
Each copy is now turned about its own centroid, all by the same rotation, so
only the model's own orientation is random. On a control of 38 open-arm sets
(deposited chain randomly re-oriented as search model, 3 unrelated deposited
models and a fold-scrambled decoy per set) the wrong models' null falls to
1.4-4.2 sd on the tNCS sets (one at 7.2, left UNDECIDED), both wrong decisions
become NOT_PLACED, and no right-model or AlphaFold-model decision changes
(nulls stay 15-107 sd). The 193-set AF survey replay loses no GREEN (two
UNDECIDED become NOT_PLACED); the four twin/tNCS targets stay correct.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
The search was CPU-bound: on a 54-chain partial complex ~80% of all CPU samples were the score's
FitModelScale grid (and libm's exp under it), with the GPU busy 13 of 37 s.
- score_placement fits its scale on the GPU where its structure factors are made there
(ModelScaleGPU's FitSolvent / Fit, as the rigid body already does); equal to rounding, the CPU
build unchanged.
- A model whose expected log-likelihood gain (sum sigma_A^4 / 2, McCoy et al. 2017; sigma_A^2 the
part of the scattering one copy is by Matthews, times Luzzati's D^2) is below 10 is not searched
for. Over the 193-set survey the lowest eLLG of a placed model is 26; every PAE domain below 24
placed nowhere, so no decision changes.
- The likelihood setups made where first needed, one serving both where the reflections are the same.
- After a candidate fails its null, the other placed candidates' nulls side by side.
- The first copy's placements refined side by side on an engine each; the groups' translation
functions side by side; CopyFactors and KnownStructureFactors over threads; the further copies'
CopyFactors and gains for the peaks walked all at once; the known-structure TF setup made while the
first copy is refined. All exact.
Survey replay (193 sets): no decision changes; two R values differ in the 3rd decimal (GPU scale).
Alternating runs, MODEL_PRIOR_SECONDS base -> new: 54-chain complex 38.3 -> 14.7, 4mtk 18.3 -> 9.7,
9v16 13.7 -> 8.3 (shared machine, indicative).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
The spot-width stage of the pre-scan copied every preprocessed frame back to the host (72 MB a
frame on a 16M detector, page-locked host mirrors per worker) for measurements that read a few
pixels about each spot. Now:
- MeasureSpotsOnTiles: the flux curves and the footprint read (2*24+1)^2 squares about each spot,
gathered on the card (ImagePreprocessorBufferGPU::GatherTiles) - no copy_image_to_host, no host
mirror. The flux curves are bounded at R_BKG_OUT = 20 px. The footprint keeps its arithmetic; a
per-iteration check that its window is inside the square, and a spot whose window walks off it is
measured again from scratch on an 80 px square, then on the whole frame. Where the squares would
be more pixels than the frame, the frame is read. Same pixels in the same order, so the same bits.
Harness over 34 rotation sets (~1500 frames): 0 differences; misses at 24 on a few spots of most
frames, at 54 on up to 80% of frames of a few sets, at 80 on none.
- The azimuthal mapping and the per-worker spot engines are built while the frames are read.
- The tiers no longer wait for each other: workers run on through the visit order, each tier's
answer is drawn from exactly its own frames, and frames measured past the tier the answer stops
at are thrown away - a frame is measured the same whichever tier it is read for.
- PreScanMemo keeps the spot measurement, the defective pixels and the background gain under the
pre-scan key plus the spot-finding and azimuthal settings, so a copy of the run that pre-scans
again on the same inputs reads no frames.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
- Likelihood (Read 2001; Z-score over random positions as McCoy et al. 2005): log-likelihood gain of
Rice/Woolfson E distributions, sigma_A from the pLDDT coordinate error with its scale refined, to
2.5 A (or the 20000 lowest-resolution reflections). Further-copy TF peaks ranked by the gain on the
search's reflections (peaks that lower it dropped); a copy whose TF Z is below 8 also stands out where
its refined likelihood Z-score reaches 8 - only where the copies placed are as many as Matthews makes
most likely. The likelihood is computed only where a translation function does not stand out.
Copy structure factors exact from one FFT of the cell's grid, or read off the molecular transform
where it reaches the likelihood's resolution.
- PAE domains: the AlphaFold DB's predicted_aligned_error JSON beside the model (same name pattern) is
read automatically; average linkage on the symmetrised PAE (10 A, domains >= 40 residues). Where the
whole model is placed nowhere, each domain is searched by itself side by side; placed domains must
agree, and a domain alone may only confirm the answer's group.
- Twin comparison: where the answer's merge chose and the twin hypothesis' merge exists, the chosen
placement read as a crystal in the subgroup (its symmetry images as copies, the subgroup's setting
found from the two cells) against the model searched for afresh in the subgroup, both against the
data twinned by the law (fraction refined on the working set, Yeates 1997); the subgroup is taken by
a 0.03 R-free lead plus its own null/margin -> TWIN_SUBGROUP, written to p_mr with law and fraction.
With --af-model rugnux makes the subgroup merge wherever a metric re-ask promoted the indexed group.
- tNCS rule: a group with fewer copies beside tNCS (screws bridged by the vector, or the enantiomorph)
no longer leaves the choice undecided where the chosen copies, carried into it and refined, fit
worse by 0.03 in R-free.
- Speed: the molecular transform's FFT on the GPU (cuFFT, serialised), no rigid body for trials that
lower the likelihood, re-search only with tNCS, PAE domains searched in parallel.
Survey (193 sets, MR replay on stored merges): single GREEN 58 -> 68, multi 48 -> 58, partial 8 -> 9;
4bwl (and, in full runs, 2wnq and a withdrawn entry of the same protein) TWIN_SUBGROUP P21 instead of
a wrong AGREES; a P21212 set with tNCS now SPACE_GROUP P21212 instead of UNDECIDED; no wrong call.
CPU and GPU builds make the same decisions on 18 demo/multi-copy/twin sets.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
The group reduction reaches every observation through the group permutation, and so fetched one
32-byte sector per field for I, sigma, partiality and corr. I, sigma, partiality and prescaling_corr
now sit together in one aligned record, so it fetches two sectors per observation; the grouping
kernels' finiteness test and the combine read the same record. The per-observation fit coefficient,
its flag and 1/sigma are no longer stored: the per-frame fit forms them as it reads each observation
(the same expressions, so the same bits), which removes PrepScaleObsKernel and 13 bytes per
observation of device memory. DeviceBytes' per-observation figure drops by those 13 bytes.
Every sum keeps its order: p.hkl md5 identical to the base on cytc, 8a1a and 8qaw.
RTX 5080, nsys, merge-engine kernels: 8a1a 2459 -> 1924 ms, 8qaw 3624 -> 3025 ms, cytc 501 -> 476 ms
(ReduceGroupMeans 1.5-1.6x faster, PrepScaleObs gone, FitPerFrameG absorbs it).
Measured and not taken: storing all per-observation arrays in group order makes the reduction 3-4x
faster on its own, but the per-frame fit then has to gather, and moving 15 arrays at every grouping
costs as much as the reduction saves (net 0 on the 5080); a record for the combine-only fields
did not speed the combine.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
The beam-stop projection (ShadowAccumulatorGPU) and the defective-pixel sums (HotPixelFinderGPU) are
now kept one per card, fed by workers pinned t % gpus like the other pre-scan workers, and folded
into the first card when they are read (cudaMemcpyPeer + an integer add / max-with-count kernel).
All of them are integer sums or a maximum with the first-value rule, so the fold is order-free and
the result is identical to one card's. With one card nothing is folded.
The pre-scan sample is read once and kept (compressed as stored; up to 2 GiB, a frame past that is
read again where it is wanted) for the spot measurement and the defective pixels, which read every
frame again each before - a CBF/marCCD/SMV frame was decoded on the host three times.
A beam-stop detection that finds no shadow leaves the mask as it was, so every probe copy of the run
repeated it on the same inputs. Its answer (shadow, ring centre, mean projection) is now kept in a
memo shared with the copies, keyed on the sampled frames, the geometry the rings are drawn in, the
polarization and the binary mask; up to 4 entries.
Debug TIMING lines for the pre-scan's parts (read and projection, beam stop, defective pixels, spot
measurement). New tests: two accumulators/finders folded match one fed every frame.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
Copy number. The hard "Vm >= 4: MULTI_COPY, no decision" rule is gone. Copies
are searched one after another up to the largest copy number with a Matthews
probability of 1% (a Gumbel distribution with Kantardjieff & Rupp's 2003 mode
and mean, no resolution dependence); COPIES_LIKELY reports the most probable.
Sequential search. Copy n's translation function carries copies 1..n-1 as a
known partial structure: the cross term 2 Re(conj(F_known) F(t)) goes on the
same FFT grid (CPU FFTW and GPU cuFFT; KnownStructureFactors, AssemblyCorrelation).
Its peaks, past those that clash, are ranked by the correlation of the whole
assembly with E^2 (the TF itself, linear in |F|^2, rewards copying a placed
copy onto an origin shift), the first that stands out (TF Z >= 8), packs and
lowers R-free by 0.01 is kept; each copy is refined beside the others held
fixed (RefineRigidBody's fixed copies), and the first copy again beside the
second, the earlier ones once more at the end.
Packing: ClashFraction counts C-alpha clashes with other copies and all their
mates, binned (7mzt: 51 s of clash checks -> 1.9 s).
Search quality: rotation-function peaks hill-climbed off the 6 deg sampling
(6ze4: RF Z 10.4 -> 16.0, 3.3 deg off the deposited orientation -> 0.2); the
first copy goes down the TF list until three placements that pack are
refined, so a group is CLASH only where no offered orientation packs; groups
are searched side by side, and every group the best one leads is searched
again with the best one's copy orientations tried first, the better search kept.
tNCS: the native Patterson of E^2 (5-10 A) off the lattice; placed copies'
orientations join every further copy's search; and where the tNCS vector turns
one group's screw axes into another's (TncsRelatesGroups), a group whose search
stopped at fewer copies does not lose to the other - UNDECIDED. Measured on an
in-house tNCS crystal (vector c/3 along the 3-fold): both hands fit with three
copies to within 0.01 in R-free, and where one hand's search failed the other
was taken as ENANTIOMORPH; now UNDECIDED.
Null: the replicates are placed by the translation function in their own
orientation, as the search placed the real one (a null at the real centroid
reads high), and the gate is 4.51 = t(8) sqrt(1 + 1/9), 3 sd of a nine-draw
null, the gate --model's MODEL_FIT is now read at.
Speed: placements scored on the GPU structure-factor engines; further copies'
trials score with the first copy's bulk solvent held, the assembly refitted at
the end. Survey of 193 sets (jfjoch_mr_replay on stored merges, RTX 5080
shared with other jobs): MR median 1.15 s, max 24 s; demos 0.8-1.1 s.
Report: COPIES and COPIES_LIKELY columns, MODEL_PRIOR_COPIES.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
An optional `fixed` model: other atoms of the crystal held where they are -
the other copies of a molecular-replacement search - whose Fcalc is added to
the body's at every evaluation and whose atoms the bulk-solvent mask covers
together with the body's (computed once per zone, with the body where the
zone starts, and held). With them the origin is no longer free, so the gauge
projection is skipped. On the CPU the fixed part is gridded once per zone; on
the GPU the fixed atoms must be copies of the body, and each one's placement
is read off by Kabsch superposition and composed on the device with the
engine's own density, transform and mask (SetFixedCopies), so a fit beside
three copies of a 90 kDa model costs what a fit alone does.
Without `fixed` nothing changes (the composition kernel's extra term is null).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
Three defects the battery's wrong-model null found (5 sets x 5 deposited models of unrelated proteins):
CC(model, data) significance. The per-shell sigma was Fisher's transform on independent reflections,
but the data and a scaled model share the falloff across a shell, the anisotropic B fitted to these
very data, the bulk solvent and the symmetry enhancement epsilon - whatever the model is. 21 of 24
wrong models read CC_MODEL_CONFIRMED_TO_D_MIN to the data's d_min (10-15 sigma in most shells).
The correlation is now formed on I/epsilon and |F_model|^2/epsilon, and its null is empirical: the
data at h against the model at h + d for the 26 nearest lattice offsets, which keeps everything smooth
in reciprocal space and loses what the model knows of h. SIGMA = (z - mean) / max(sd, 1/sqrt(N-3)) in
Fisher z over those 26, a new CC_NULL column shows the null, and a shell confirms only beside a coarser
shell that also clears 3 sigma (one shell alone did so once in 216 wrong-model shells). Prototyped on
the p_maps.mtz of a full open-arm run and the null runs: wrong models confirm 0/24 (was 21/24), the
deposited models confirm 197/200 (was 198/200 in the prototype's reading; the one lost has CC_MODEL
0.05-0.13 in every shell). The report no longer says a model cannot agree by accident.
MODEL_FIT gate. Accepted at 3.0 sample sds of a nine-draw null, as if the sd were known; an unrelated
P1 helical model on a P3(1) helical crystal was ACCEPTED at +3.05. Against 48 draws it reads +2.09 and
the best random placement fits as well as it does, and drawing the null about the cell axes instead
changes nothing (+2.15): the cause is the nine-draw sd, not the orientation. The gate is now 3 sigma
allowing for that: t(8) * sqrt(1 + 1/9) = 4.51. The deposited models of the open arm read +7.0 and up.
The indexing lead stays at 3 sigma (INDEXING_MARGIN_SIGMA): it is read only behind the fit gate, and
a deposited model settles one at +4.15.
Rigid-body GPU reserve. The finest zone's observations were counted by the data's d, but the zones
count them by the d of the index in the model's ASU, and a model's group need not keep the data's
metric (an I2(1)3 model on I222 data: 2320 in the zone against 1712 reserved), so the validation died
with "zone over the engine's reserve". Counted the zones' way, under every indexing probed.
Tests: ModelValidation_CCModelUnrelatedModelConfirmsNothing (Fisher alone reads +4.7 on an unrelated
model with an anisotropic falloff; the new sigma stays under 1), ModelValidation_WrongModelInAGroupOffTheDataMetric
(fails with the old count), ResultReport_ModelValidationSection (CC_NULL column, lone-shell rule).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
