diff --git a/figure-collection-20261004/BBBC039-612-FRESH10-MECHANISM-REVIEW.rst b/figure-collection-20261004/BBBC039-612-FRESH10-MECHANISM-REVIEW.rst new file mode 100644 index 0000000000..a070abe606 --- /dev/null +++ b/figure-collection-20261004/BBBC039-612-FRESH10-MECHANISM-REVIEW.rst @@ -0,0 +1,239 @@ +BBBC039 frozen paired quality mechanism review +============================================= + +Singer owns this coordinator-only source and saved-presentation audit following +merged796. Parent owns797's manuscript/scatter. Planck owns748 integration and +the next522 receiving candidate. No scorer, pipeline, viewer, native process or +author interaction is started here. The original frozen files and failed/UNKNOWN +attempts are preserved; no settings, scores or reference answers go to authors. + +Completed determining review +--------------------------- + +Both final pipelines use Minimum Cross-Entropy and SHAPE markers/watershed. +The old explicit INTENSITY smoothing option does not smooth SHAPE's distance +landscape. Actual changed stage contracts are minimum size/suppression10 to12, +threshold smoothing1 to1.3488, and hole filling AFTER_BOTH to AFTER_DECLUMP. +Their individual contribution is not established by the paired score delta. + +The visible failure is not absent raw signal: in the saved unequal-neighbour +diagnostic, both lobes enter foreground support, but the weaker lobe has no +obvious separate displayed seed and the final ROI remains joined. A suppression +change did not repair it. This is an application/diagnostic-closure limitation, +not a demonstrated missing generic guide. No skill or processing change is +proposed. Attribution of the pooled regression to a particular setting remains +unproved; the strongest score-selected case lacks saved native matched captures. + +Source and original evidence identities +--------------------------------------- + +All paths below are retained originals, not new inputs or reconstructed views. +Define OLD as:: + + /home/ts/wt/openhcs-issue-batch-20260929/next-bbbc01395-after612-20261004/BBBC039_FRESH612_96/author-workspace/output + +Define NEW as:: + + /home/ts/wt/openhcs-issue-batch-20260929/next-h00488-h00389-bbbc03996-fresh10-after-capacity-20261005/BBBC039_FRESH10_COVERAGE_96/author-workspace/output + +NEW's image captures are on its original HDD science root, not directly in +NEW/qa. Define NEW_IMAGES as:: + + /run/media/ts/hdd/openhcs-science/next-h00488-h00389-bbbc03996-fresh10-after-capacity-20261005/BBBC039_FRESH10_COVERAGE_96 + +SHA256 pins:: + + OLD/final_full200/pipeline.py eb3a043f81c7e1508f036bbeb5c6f74aaf586f2fef8622c519eb5ad5072105e9 + OLD/REPORT.md 4600f82d902be420406137764cc6da53b672e63fb071024dfb7f0f8155f38ea7 + NEW/pipeline.py 0824f47839bfe4a5ac05592d8e57f4c548c68517ea46f7ea7765d0921b9f6b1a + NEW/report.md 50928e9d1ac905f477047605bcbcc573a135eddd3b76c134c9befa2edeaf9a59 + NEW/qa/FULL200-final-triplets.json cb23d9588d5565852870ddb8d4a012f20aec43ec26b533db84ab621a185dfac9 + NEW/qa/FULL200-final-mounted-state.json 3f6d35ddfef4f96f145324859306f4ef9ce30167d80e23c41085b9ebfca6ca1f + +Original installed source roots, identified from retained MCP receipts, are +issue-batch/engineering-mcp-queued-cancellation-20261004/receiving03/target +(OLD) and engineering-pre-first-routing-20261004/receiving10/target (NEW). +No import or backend execution was needed to read these declarations. +Relative to their openhcs/processing/backends/cellprofiler directories:: + + primary_objects.py OLD 12b3c8a3829a8b19697029467ccb3bc5f8c897ab0ebbc9f5f0b7d35f49619f5f + primary_objects.py NEW 62c3fe6e04841d69e2858f3b6492dfc86188998ea945ad769a96899fbcd8577b + thresholding.py BOTH 4ed28e53ec30580fb9558b94ad4c50eb17a59b61a4a516cf992b388fadf887d0 + morphology.py OLD bf52da0b5e8db68f613deb0bd5c61f76e1c49d6e6b8a8b1117c9dc6e4b0ea647 + morphology.py NEW 52c8b7ef7493abd1a202e0f2df812cacfb8ea515759b23dc9a9f6055d6fe52dd + +The primary_objects diff is documentation-only, including the explicit warning +that intensity smoothing is ignored for SHAPE. The morphology diff is NOT +documentation-only: it includes labeled-hole filling/kernel changes. Therefore +this review does not claim an identical complete numerical backend or an +isolated parameter experiment. Both installed interop/cellprofiler/ +image_normalization.py files have SHA256 +a840f237daecc63d9e45e65166fd5d0a91117a4e417134e5b1fd6971026a0e6e; +that shared delegating normalization source is not proof of every dependency's +equivalence. The original same-source-key/hash checks remain those of #796. + +Effective stage contracts, not argument-name correlation +------------------------------------------------------- + +Read identify_primary_objects, FillHolesOption, DeclumpingMaximaGeometry, +filter_labels_by_diameter_range and PrimaryObjectDiagnosticPlanes in the +actual installed source. Their determining relationships are: + +* Threshold smoothing changes applied foreground support even for SHAPE. + Both pipelines use global Minimum Cross-Entropy, not a MCE/Otsu comparison. +* SHAPE uses distance to the admitted labeled boundary plus deterministic tiny + tie-breaking noise. INTENSITY's automatic_smoothing/smoothing_filter_size + do not alter this landscape. The old explicit3/new automatic option is not + a SHAPE mechanism explanation. +* With advanced settings enabled, AFTER_BOTH fills before markers and after + declumping; AFTER_DECLUMP does not pre-fill. Threshold-support diagnostics + are captured before pre-fill, initial components after it. Changing this + option can change the distance landscape before suppression/partition. +* Manual suppression is governed by the existing footprint/seed owner; both + runs disable automatic suppression and low-resolution maxima. +* Minimum diameter is an equivalent-area cutoff, pi*d*d/4, not a measured + bounding span. The changed lower cutoff is about78.54 to113.10 pixels. + Size filtering occurs before final hole filling. Final fill cannot recover + a basin already removed by size admission. + +Consequently support, pre-fill geometry, emitted markers, unfiltered basins +and accepted labels must be distinguished. Their interaction is a plausible +mechanism family, not a demonstrated assignment of the314 extra false negatives. +No alternative codec, semantic owner, production scanner or consumer is added. + +Strongest regression and available paired native evidence +-------------------------------------------------------- + +The already published #796 JSON is the ONLY scoring source used here: +figure-collection-20261004/bbbc039-fresh10coverage-postfreeze-evaluation.json, +SHA256 e257a17d67662967ba9af16a9a94aa78756b47a889c3e093a0cf1abe640f828e. +Its same200 comparison reports314 extra FN,11 extra FP and303 fewer predictions; +pooled F1 is90.6224% to89.8368%. These are annotation-relative results, not a +perfect biological census or evidence of skill causality. No score was rerun. + +The largest recorded per-field F1 decrease is20589_P23_7: TP12/FP0/FN0 becomes +TP8/FP2/FN4. Neither inspected native capture inventory supplies a matched +triple for this field. Dense20630_H06_6 also lacks the required captures; it +must not be substituted with the differently identified20639_H06_4. No +reference pixels were opened, and no synthetic overlay was generated to fill +these gaps. Thus the strongest score-selected mechanism is unresolved. + +Available shared final field20630_A02_1 is a genuine, lesser paired regression: +TP101/FP5/FN9 becomes TP96/FP5/FN14; F1 .935185 to .909953. Its six original +raw/result/combined PNGs were personally opened. Upper-left crowded group +partitions differ, with a broader joined region in NEW, while many isolated +compact-body controls remain coherently outlined in both. This supports local +partition deterioration, not a claim that every additional FN is that merge. +Colour palettes are not cross-run object identities. + +Both triples cover the same520x696 source at center(0,259.5,347.5), but OLD zoom +is.65 and NEW zoom.8. Within each triple isolation/camera controls match; the +cross-run screenshots are NOT pixelwise matched presentation. OLD uses a +recorded p0/p99 display request; NEW state gives raw window120..680/gamma1. +Their contrast is not asserted numerically equal. NEW saved state reports101 +ROIs in the exact source domain, scale1/translation0, not proven physical +calibration. Original OLD/runtime/mcp.stdin and NEW/runtime/mcp.stdin preserve +the route isolation, axes and capture requests; NEW's mounted-state and triplet +index carry its source/result/camera receipts. + +Failure-stage witness and positive controls +------------------------------------------ + +NEW's20592_A21_1 unequal-neighbour witness has a dim upper lobe touching a +brighter lower one around native y239..258/x116. Camera is center(0,250,180), +zoom2, z_index0/timepoint0. Personally opened raw/result/combined-corrected, +foreground, response and seeds show connected admitted support, a broad +lower-dominant distance response, no obvious separate seed in the weaker upper +lobe, and one accepted ROI. Nearby compact bodies retain individual seeds and +outlines: this is not a whole-frame failure. The earlier all-layer combined +capture after failed isolation is retained but excluded. + +The original report and saved REPAIR01/REPAIR03 combined captures agree that +lower suppression and then intensity markers with lower suppression did not +separate this witness. REPAIR02's reported two peaks in a one-dimensional +smoothed profile do not prove two emitted two-dimensional markers or basins. +This audit does not infer unobserved IDs from the profile. This field's overall +annotation F1 actually IMPROVES (.902778 to .929577), so its local merge is a +failure-stage control, not proof of the pooled regression or old same-pair +success. No old matched capture of this particular pair was established. + +As an independent OLD continuous-body control, candidate02_field0_native's +three PNGs were opened. The native legend identifies20585_F14_7, not the A06 +true-pair repair. A ring-shaped body remains one filled outline and several +isolated ovals remain coherent; the continuous crescent still has multiple +partitions. OLD is therefore not presented as error-free or a universally +better declumping choice. The old report's A06 suppression repair is distinct +from this F14 control and cannot be transferred by filename alone. + +Existing guidance versus actual applied reasoning +------------------------------------------------- + +The full current analysis-learning, measurement-interpretation and +segmentation-diagnostics references were read at main154566ed. Relevant source +SHA256, under packaging/codex/openhcs/skills/use-openhcs/references/:: + + analysis-learning.md 78a5179e6e5d5e12cfc281ef17ae359bcb3496508609c4db3357e3dfe0e6b7fa + measurement-interpretation.md 5f8b4784f4cfe1f008b593fafce99ba08065b122daa0dad58fba3546b2ad4fd1 + segmentation-diagnostics.md 3173aebf8412d878148efa7dee5e8ccb011e08e51377ac5c707042d6f2df6938 + +Measurement already distinguishes equivalent area from span, threshold from +marker smoothing, within-body texture from real pairs, and footprint competition +with a stronger neighbour's shoulder. Its stage-linked worked contrast states +that profile peaks are not proof of two-dimensional maxima or dividing lines. +Segmentation already requires the earliest-loss distinction across support, +seeds, unfiltered basins and size admission, without a zero-error acceptance +rule. Analysis learning already requires tested, scoped repair evidence. + +NEW's original stdin94..97 and report record relevant guide retrieval BEFORE +FIRST, followed by measured raw profiles and a reasoned SHAPE hypothesis. This +is not a demonstrated failure to retrieve the marker/body guides. It also is +not proof that every later passage of today's guide existed in that bundle. +The incomplete applied link is from the chosen empirical envelope to actual +weaker-neighbour seed/partition survival and pre-filter geometry after changed +support/fill/suppression/area choices. Failed local repairs preserve that limit; +choosing the original useful method after them is not itself misconduct or a +reason to reject every measurement. No known older settings are fed forward +to authors as an answer, and no duplicate generic skill paragraph is warranted. + +Capture byte pins +----------------- + +All entries below were personally opened unchanged. Every basename is the +listed timestamp followed by _napari_6017_OpenHCS_Napari_Visualization.png. +Prefixes are OLD/qa or NEW_IMAGES/qa, respectively:: + + OLD final_field2_full_raw/20261004T130942480136Z 86c07ba9c159a38caa27e59a0b51687af58881918893f0853a1ff9fdc131358c + OLD final_field2_full_result/20261004T130947639815Z 1278b7885e81835205491797c6ca9d6fe551162dace90f17756b0c5614d7bc92 + OLD final_field2_full_combined/20261004T130952836362Z 9966a862657b5c2000df0576a80efc6ca9d73cd92201291fb4195fd88c40bfa0 + OLD candidate02_field0_native_raw/20261004T125348507707Z a01b77cae1a756f26727feeb5c6af6c92e7057ca803dbc86969f0f38aa917daf + OLD candidate02_field0_native_result/20261004T125352981245Z a6de6d3983390624d7f6efc72fd6d4f14c6e8ba8be358aca24e754f82753a491 + OLD candidate02_field0_native_combined/20261004T125357377402Z fcb6edf160abb16608f3bd8924c196155ecedf675ceae83c17dad50609982273 + NEW FULL200/A02-raw/20261005T130255533882Z d4bdb0815564bf1b7a8c75e58dfc3df259677bb3638ce6b289ed52d918831617 + NEW FULL200/A02-result/20261005T130303594735Z 1ba7d08069454f01d6ce9f2172c2c90cda22fd66cee723fad84e36ca4ca4f116 + NEW FULL200/A02-combined/20261005T130311190833Z 4320ab0e41b53dd2f4fa4b875684be85d1f4b716311aa2884870b36563d6e4b4 + NEW DEV06/A21-pair/raw/20261005T123059076340Z bc7817b9a4c6e32b39696c71743483826d560ae96bcf21fe78f169e36981f459 + NEW DEV06/A21-pair/result/20261005T123105270611Z 239f96affda0e6dd33b275d77056bfd46e09090d9e88fe91a143bf33bcc7ec7b + NEW DEV06/A21-pair/combined-corrected/20261005T122920525074Z fd25a92a6c540da93d415449b39c2007af7afda67b99cdbb8e91335eff72fb8e + NEW DEV06/A21-pair/foreground/20261005T123111933921Z 92767d1261ea3dff3600c9bbf0cab4594e048f2c69f40366f1f14ec1864ec5f0 + NEW DEV06/A21-pair/response/20261005T123123420522Z c3d7c2f45140b4d8642aed6441463787b609230fa408f6183a62ae2133ff88b5 + NEW DEV06/A21-pair/seeds/20261005T123129936476Z d17adf37948ed7e11d045255642227d68ef2fdcd26f3c667d1419371b9cb00e2 + NEW REPAIR01/A21-pair/combined/20261005T123631204684Z b46c226eb833749fe08d7ddc468acb443ff0cd889e829d57493386eef38b6a00 + NEW REPAIR03/A21-pair-combined/20261005T124804799692Z 2514fc480419782f47a9f96a82c84981752e2aec17688e081ed4ad518a1d3aad + +Disposition and precise acceptance +---------------------------------- + +The original source, guide, frozen report and installed-declaration relationships +were read; the selected17 native PNGs were opened and byte-pinned. The current +NRA/refactor-audit instructions and applicable IDEN-6/BOUND-2 patterns informed +identity and effective-declaration checking; there is no structural refactor, +global AST/R0/R1 proof claim or new scanner. No tests, scorer, external data +decoder, scientific pipeline or renderer were run. Corrected read-only path/key +lookups do not represent scientific retries. No references were opened as +pixels, no live author contacted, no outputs or UNKNOWN dispositions changed. + +The audit acceptance is met at its actual tier: a saved local partition +regression, a stage-specific unresolved pair, coherent positive controls, and +explicit missing cross-run/prefill/marker/basin controls. The pooled causal +allocation remains unresolved. #798 changes this receipt only; parent owns +#797 publication and Planck the #748/#522 native receiving integration.