SWEEP-V2: the v2 cartoon sweep, and what the generator did with it

Work package SWEEP-V2 of the SGH programme (operations/research/sgh-program-20260908/PLAN.md), 8 September 2026. Two spot A100s (fleet slots b and c) ran the 232-job sweep-v2 package to completion, then the whole thing was evaluated on CPU. Slots a and d were not touched. Nothing was committed to git.

Engineering evidence only. Every number below is an image-only measurement — shape, spacing, texture, colour, or one pathology encoder. Nothing here diagnoses, grades or stages anything, no number says an image is diagnostically correct, no pathologist has seen any of it, and Yash has not reviewed a sheet. Every "better" or "worse" names its instrument and its n.

Headline

232 of 232 planned generations landed: 232 records, 232 PNGs, every one non-empty and hashing to its record's output_sha256, 0 skipped manifests, 0 failed jobs, no preemption. The two fix-* jobs reproduced sweep-v1's zero-byte pass-2 image exactly, byte for byte.

On the question the package existed to answer — do gland-calibre cartoon lumens survive the two-pass repaint better than A1's pit-sized rosettes? — the answer is no. Pass 1 follows the v2 cartoon slightly better in absolute lumen IoU (0.574 vs 0.544) but identically as a fraction of its own ceiling (89% vs 90%), and by pass 2 the v2 layout has collapsed further than v1's: lumen IoU 0.017 vs 0.084 at start_index 6, 0.044 vs 0.056 at start_index 3, and ring_with_lumen_fraction 0.076 vs 0.221 at si6. The v2 cartoons are a better drawing of a gland and a worse conditioning signal for this generator.

What did improve, consistently and on every category: detail. Mean absolute Laplacian rises 13.18 → 17.03 at pass 1, 17.07 → 18.80 at si6, 19.05 → 19.90 at si3 (real held-out 22.77). And median_ring_area_um2 at pass 2 is now 423-428 um2 against a real 403 (v1: 435-441). The morphometric envelope distance, though, is worse pooled at every depth (si3 pooled D 6.28 vs 5.64): by category, normal improves (6.79 vs 7.10), gastritis regresses (6.14 vs 5.77), IM and mixed are level.

The mechanism is visible in the images and is the risk A1V2_RESULT.md named in its own section 7 item 6: the 5.8-um chromatin-granule lattice that makes ring_topology seal the gland wall is a rendering trick, and the generator reads it as nuclei. The v2 pass-1 output is a regular, evenly spaced field of oval nuclei rather than a wall around a lumen, and a regular field of nuclei has no pale enclosed rings — hence ring_with_lumen_fraction medians of exactly 0.000 at si6 for normal and gastritis, and 0.000 in six of the twelve (category x cut) cells.

Two engineering faults were found in the evaluation harness while doing this, one of which invalidates numbers already published:

• Every sweep_* probe figure in EVAL_RESULT.md was computed on EVAL-BUILD's dry-run stand-in images, not on the real sweep. probes/run_probes.py caches tile embeddings per set name (probes/embeddings/<set>__1024.npz) and skips any set whose file already exists; evaluate_sweep.py never passes --force-embed. Section 9 has the evidence and the corrected v1 numbers.
• The same cache made the sweep-v2 probe stage report sweep-v1's numbers on its first two runs.

1. Timeline

All times UTC, 8 September 2026. Slot map: package slot0 -> fleet b (sgh-histo-qwen-a100-recovery-b, us-central1-b), slot1 -> fleet c (sgh-a100-slot-c, us-central1-c).

Setup, boot to first generation: 16.7 minutes (SWEEP-RUN's was 45). Generation 80.6 / 79.2 minutes. Retrieval ~10 minutes each, in parallel.

fleet.sh status after the run:

SLOT  VM                               ZONE            STATUS
a     sgh-a100-slot-a                  us-central1-a   TERMINATED
b     sgh-histo-qwen-a100-recovery-b   us-central1-b   TERMINATED
c     sgh-a100-slot-c                  us-central1-c   TERMINATED
d     sgh-a100-slot-d                  us-central1-f   RUNNING      <- F2, untouched

What the 16.7-minute setup bought

SWEEP-RUN's fix 4 (render the cartoon sets in parallel chunks before launch, so the A100 is never idle behind a single-threaded CPU renderer) was applied from the start, plus two additions:

• code/render_chunk.py (new) reconstructs each set's own job list — the arguments each render_set.build() call would receive, in the set's order — and renders only an assigned slice, writing a per-chunk MANIFEST.chunk.json. code/manifest_dir.py (new) verifies each chunk against its manifest, merges the parts into a staging directory and publishes it with os.replace, so worker.sh's "already rendered" glob can never see a half-built set. Verified before use: rendering three stems through render_chunk.py on this Mac reproduced 6/6 and 6/6 files byte-identical to cartoons/a1v2/MANIFEST.sha256.json and cartoons/d1v2/MANIFEST.sha256.json, including the rot180 write path.
• fleet/prerender.sh (new) runs 6 chunks of a1v2 and 2 of d1v2 concurrently (8 processes on 12 vCPU) and copies the v1 run's token-library / token-stats across. Measured: ~62 s per 4096x2048 field under 8-way concurrency; 65 builds producing 70 cartoons in 10.5 minutes wall clock, against 50-65 minutes serial at SWEEP-RUN's measured 45-60 s per cartoon. The A100 was powered for 10 of those minutes rather than an hour.

The manifests use token_reference only, so the worker's own dependency scan reported libraries=[] token_stats=0 and neither library was needed. The copy cost 2 seconds and was insurance against a worker that decided to rebuild them (~50 minutes of A100 each).

The a1 cartoon for the fix job was copied, not re-rendered

packages/sweep-v1/retrieved/records.jsonl records the sweep-v1 pass-1 job's reference_sha256 = c4c934f7…; that cartoon was rendered on fleet slot a, which is busy. Slot b's own v1 run root carries cartoons/a1/intestinal_metaplasia_s14_oblique_cartoon.png hashing to exactly the same c4c934f7…, so prerender.sh copied it verbatim rather than re-rendering. That is what makes the sha comparison in section 4 a real determinism test rather than a re-derivation — and it is itself evidence that slot a and slot b render this cartoon identically, which section 5 extends to slots b and c across all 210 a1v2/d1v2 files.

2. Cost and powered minutes

From fleet/state/<slot>/powered.log, one session per slot, no preemption.

At USD 2.12/h for a spot a2-highgpu-1g, plus roughly USD 0.41 of internet egress for the 3.4 GB retrieved (3.4 GB of PNGs + 41 MB of label maps).

GPU utilisation: 2.51 GPU-hours of generation against 3.60 powered hours = 69.8%, against SWEEP-RUN's 56%. The remaining 30% is 1.2 min boot, 1.5 min stage, 10.5 min pre-render, ~10 min retrieval, per slot.

Throughput, from the 232 records' own elapsed_seconds:

Total 9036 s = 2.51 GPU-hours, mean 38.9 s per 4096x2048 canvas. This reproduces SWEEP-RUN's figures (31.84 / 45.29 / 52.28 s medians) to within 0.2 s at every depth, on the same two VMs four hours later.

3. Coverage: expected vs present, per manifest

packages/sweep-v2/retrieved/merge_records.py (adapted from sweep-v1's) merges the two slots, rewrites each output to slot<i>/out/<path>, adds slot and fleet_slot, and — the addition sweep-v1 needed and did not have — re-hashes every PNG and compares it with the record's output_sha256. sweep-v1's coverage check tested existence only, which is how a zero-byte file was reported as present (EVAL_RESULT.md section 1.1).

evidence/manifest-summary.json: slot b {"ran":7,"skipped":0,"failed":0}, slot c {"ran":5,"skipped":0,"failed":0}. 232 of 232 outputs exist, are non-empty and hash to their record's output_sha256. 0 duplicate records. All 232 canvases are 4096x2048. 232 of 232 records have token_source equal to their extra.expected_token_source (token_reference throughout). By experiment: a1v2 180, d1v2 50, a1 2. By category: IM 97, normal 45, gastritis 45, mixed 45.

No preemption occurred, so the resume path was not exercised here; FLEET_RESULT.md and SWEEP_RESULT.md both verified it record-by-record.

4. The fix job: sweep-v1's zero-byte image is recovered, exactly

packages/sweep-v2/fix_check.py, output in packages/sweep-v2/eval/fix-check.json.

Pass 1 reproduced sweep-v1's bytes exactly on a different VM in a different zone five hours later — the control that says the generator is deterministic on this rig — and the regenerated pass 2 hashes to precisely the output_sha256 sweep-v1's record claimed for the file that came back empty. The image is recovered, at packages/sweep-v2/retrieved/slot0/out/a1/pass2/intestinal_metaplasia/intestinal_metaplasia_s14_oblique.png. Anyone re-running the sweep-v1 evaluation can copy it into place and the IM pass2-si6 cell becomes n = 5 rather than n = 4.

5. The evaluation, and the harness change it needed

P=research/sgh-synthetic-histopathology/.venv/bin/python
$P packages/sweep-v2/retrieved/merge_records.py
$P packages/sweep-v1/evaluate_sweep.py all \
     --records packages/sweep-v2/retrieved/records.jsonl \
     --out packages/sweep-v2/eval --jobs 12 --probes \
     --exp-alias a1v2=a1,d1v2=d1 --exp-drop a1

stage_a1, stage_d1, arm_label, base_cartoon and the sheets all key on extra.exp being literally "a1" / "d1", so an unaliased run reports "no data" for everything. --exp-alias NEW=OLD and --exp-drop were added to packages/sweep-v1/evaluate_sweep.py — additive, default no-op, 4 hunks. --exp-alias rewrites extra.exp on load and keeps the original in extra.exp_raw and in a new exp_raw column of features.csv; nothing on disk is renamed. --exp-drop removes records by their original exp before aliasing, which is how the two genuine a1 fix records are kept out of the aliased a1 rows (they are checked separately in section 4). Verified on a 5-record fixture: 3 records with exp=a1v2 came out as set=sweep_a1, exp=a1, exp_raw=a1v2, and 2 real a1 records were dropped. Re-running sweep-v1's evaluation without the flags is unchanged except that features.csv gains the exp_raw column.

Everything else is the harness as EVAL-RUN left it, --ncc-refs at its default targeted. Wall clock: features 100 s, a1 <1 s, d1 9 s, arms 13 s, screens 240 s (embed 119 s + 222 NCC pairs), sheets, select, report — about 11 minutes, plus the extra measurements below.

The two measurements EVAL-RUN added by hand

EVAL-RUN's a1-layout-iou and within-arm-diversity scripts were not shipped, so both were rewritten against the harness's own functions (evaluate_sweep._d1_one, analysis_masks, label_mask, iou, the same sydney_qa_v1.json) as packages/sweep-v2/layout_iou.py and packages/sweep-v2/within_arm_diversity.py. The rewrite reproduces EVAL-RUN's own file exactly where they overlap — e.g. within_arm_diversity.py run on sweep-v1's features.csv gives real held-out IM 0.0347 and a1 pass2-si3 IM across 0.4906, both identical to packages/sweep-v1/eval/within-arm-diversity.json. sweep-v1's arms were re-measured with it so both sides of every diversity comparison come off one instrument.

The VM-vs-Mac cartoon caveat is now closed

EVAL_RESULT.md section (c) had to measure every layout IoU against the local Mac cartoon label maps, because the VM-rendered cartoons were never retrieved, and flagged that as its caveat. This package retrieved them (20 MB per slot, retrieved/aux/slot{0,1}/cartoons/).

So the two machines still disagree only about the stochastic scatter of stromal nuclei (the numpy/scipy/Pillow skew SWEEP_RESULT.md section 6 characterised), the lumen label is bit-identical, and EVAL-RUN's Mac-based IoU numbers were safe. Both are reported below anyway (a1-layout-iou-vm.* and a1-layout-iou-mac.*); they agree.

The IoU ceiling is reported from the mac run only: the retrieved aux tree carries label maps, not the 734 MB of cartoon RGB, so a VM-labels ceiling would have to pair a Mac RGB with a VM label map. layout_iou.py --ceiling off refuses that mixture rather than silently producing it.

6. (a) a1v2 per category, beside a1 and the real held-out fields

Medians per category pooled over the three cuts, n = 15 per arm (v1 IM pass2-si6 is n = 14 because of the zero-byte file). env pct is the percentage of that category's real held-out fields at or below this image's envelope distance, so 50 is the real median and 100 is outside the whole real spread. D is the C1 MinCovDet Mahalanobis. lap is C2's mean absolute Laplacian. cart IoU is the lumen IoU against the output's own cartoon label map (VM bytes). within SSIM is the mean pairwise SSIM inside that (arm, category) — real held-out fields sit at 0.034-0.047. The last three columns are the copy screens: phikon-v2 max cosine (fail line 0.95), the fraction of candidate tiles whose nearest real reference is their own donor field, and the max pixel NCC (threshold 0.92). Generated by packages/sweep-v2/report_tables_sweep_v2.py; full file packages/sweep-v2/eval/tables.md.

Pooled over all four categories (n = 60 per arm; the cartoon-IoU column is repeated for the VM and the Mac label maps, which agree):

7. Did gland-sized lumens survive the repaint better than v1's?

No. The direct instrument is lumen IoU against the output's own cartoon label map, read against the ceiling that map can reach at all (the cartoon PNG measured against itself through the same masks):

The v2 cartoons are a better target: their self-IoU ceiling is higher (0.644 vs 0.604 for the a1 grid; 0.580 vs 0.407 for the d1 grid), which is exactly what "gland-calibre lumens instead of pit-sized rosettes" should do to a lumen metric. And pass 1 tracks them just as faithfully — 89% of ceiling against v1's 90%, an absolute IoU of 0.574 against 0.544. That part of A1-V2's hypothesis holds.

But the collapse at pass 2 is worse, not better: 2.6% of ceiling at si6 against v1's 14%, and 6.8% at si3 against 9%. The same is true of the topology metric: ring_with_lumen_fraction pooled falls 0.393 → 0.221 → 0.540 across the v1 ladder and 0.195 → 0.076 → 0.503 across the v2 one, with medians of exactly 0.000 at si6 for normal and gastritis. Bigger lumens did not make the repaint keep them; they made pass 1's rings too big (median ring area 962 um2 against a real 403 and v1's 685) and left pass 2 with nothing.

By cut, pooled over categories (n = 20 per cell):

The v2 oblique cut is the best-followed cell anywhere in either sweep at pass 1 (IoU 0.666), which is consistent with A1-V2's oblique_frame="mid" change — that frame carries no background at all. It buys nothing by pass 2: v1's oblique advantage at si6 (0.316 ring) is gone in v2 (0.066).

8. Did realism change?

Detail: yes, better, everywhere. C2 called the Laplacian "the largest single lever on the tell", and it moves toward real at every depth and in every category:

v2's mixed pass2-si3 at 21.69 is the closest any generated arm in either sweep has come to the real 22.77.

Envelope: no, worse. Pooled MinCovDet distance D is 9.99 vs 8.87 at pass 1, 8.12 vs 7.35 at si6, and 6.28 vs 5.64 at si3 (real held-out 3.51). The per-category si3 medians are normal 6.79 vs 7.10 (better), IM 4.76 vs 4.67 (level), mixed 5.69 vs 5.78 (level), gastritis 6.14 vs 5.77 (worse) — so v2 is not uniformly worse; the pooled median rises because it sits in the middle of the four cells, and gastritis moving up 0.37 while normal moves down 0.31 shifts the middle of the 60-value distribution upward. The honest summary is: normal improves, gastritis regresses, IM and mixed are level, and the pooled figure follows gastritis. Envelope percentile stays saturated: every median except gastritis si3 (83) sits at 100, as it did for v1.

Fraction of pass-2 outputs inside both the real ring range and the real envelope band:

All 11 of v2's are pass2-si3: gastritis 5, IM 6. No normal or mixed output is inside both bands (v1 had one normal). Pareto selection (eval/candidates.csv) picks 32 candidates, 8 per category, all on the front, and a1v2 pass2-si3 supplies 26 of them, pass2-si6 5 and d1v2 donor0 1.

Diversity within an arm is slightly worse. Mean pairwise SSIM inside each (arm, category), on the identical instrument for both sweeps (real held-out fields sit at 0.034-0.047; the old two-pass baseline at 0.477-0.527):

The v2 pass-1 outputs are markedly more like one another (+0.08 to +0.11 SSIM) than v1's, which is what a more uniform conditioning field produces. By pass 2 the difference is within 0.02.

9. Copy screens, and a probe-cache fault that invalidates published numbers

Copy screens: nothing fails. phikon-v2 embedding screen over all 145 pass-2 outputs, 1160 candidate tiles against 118 real reference fields, fail line 0.95: max cosine 0.9196, 0 flagged tiles, passed: true, minimum copy margin 0.0304 (on a1v2 pass2-si3). Pixel NCC search (64 px window, stride 256, threshold 0.92, targeted references): 222 pairs, max NCC 0.8666, 0 flagged windows, 0 failures. Both margins are tighter than sweep-v1's (0.9049 / 0.0451 / 0.8635) but neither crosses its line. Own-source tile fraction is high and slightly higher than v1's in three categories (0.917-0.942 for normal/mixed si3), which is donor dependence, not copying — the same distinction EVAL_RESULT.md section (i) drew.

The probe embedding cache is keyed on the set name

probes/run_probes.py line 228 skips embedding any set whose probes/embeddings/<set>__1024.npz already exists, and evaluate_sweep.py never passes --force-embed. Both sweeps register their experiments under the same names (sweep_a1, sweep_d1, …). Consequences, both verified:

1. This package's probe stage twice reported sweep-v1's numbers (13.3 s for a run that should take ~340 s; sweep_a1 "600 tiles", frac_nearest_in_own_donor_field 0.7716666… identical to sweep-v1's to 16 decimals). Deleting the two stale .npz and re-embedding by hand gave sweep_a1 1440 tiles from 180 fields and different numbers.
2. Worse: the cached files were never built from the real sweep at all. Reading the provenance inside every sweep_*__1024.npz in probes/embeddings/:

Those are EVAL-BUILD's dry-run fixtures. EVAL_BUILD_RESULT.md section 6.1 describes standin-rich/ as "265 records over the same 30 PNGs" of the old two-pass fullset run, re-labelled across a1/d1/a2/a3/d2/e2/b1, and says in bold that "the pixels are unrelated to the provenance stamped on them ... every table it produces is meaningless as science". load_sets' own docstring — "refresh provenance from the current manifest, so a provenance fix never forces a re-embed" — is what makes the mismatch invisible downstream: the tile labels come from the real manifest, the vectors from the fixture. Every probe A/B/C/D figure EVAL_RESULT.md quotes for a sweep_* set was therefore computed on 30 old-fullset PNGs wearing sweep labels, including its "probe B puts 77.2% of a1 tiles nearest their own donor field", its "probe A separates every one of the nine sweep sets from real tissue at 100% of fields called generated" and its probe-D percentiles. The non-sweep sets (real_heldout, gen_pass2, native_*, lock_*) come from C2's own caches and are unaffected — they reproduce exactly in every run here.

All nine stand-in caches were copied to packages/sweep-v2/eval/cache/probe-embeddings-standin-backup/ and removed, sweep-v1's sets were re-embedded from the real retrieved PNGs, and its probe stage was re-run into packages/sweep-v2/eval/v1-probes-recheck/ (343 s, exit 0). probes/embeddings/ now holds real-image caches for the eight sweep-v1 sets (sweep_a1 1632 tiles from 145 fields, etc.); sweep__1024.npz, which only the standin/ fixture ever used, is in the backup and is no longer in probes/embeddings/. sweep-v2's own caches are kept separately in packages/sweep-v2/eval/cache/probe-embeddings-v2/, because two sweeps cannot occupy one set name at once — which is the defect. Corrected numbers:

Read carefully: probe A trains its discriminator on whatever sets are in the pool, and the v1 pool had nine sweep sets against v2's two, so the a1-vs-a1v2 A comparison is indicative, not controlled. Probe B is a nearest-neighbour count and does not have that problem. On probe B, donor dependence is essentially unchanged between the two cartoon sets (0.684 → 0.671 for the a1 grid, 0.470 → 0.453 for the d1 grid) — the v2 cartoons did not loosen the generator's grip on its donor field. d1v2 at 0.453 is the lowest of any single-donor arm measured.

10. The d1v2 layout factorial, beside d1

25 outputs per pass (5 cartoons x 4 donors, plus 5 rot180 on donor 0), the harness's own d1 stage.

IoU by donor arm — flat across arms means the donor is not setting the layout:

The v1 conclusion is reproduced exactly and sharpened: at pass 1 the cartoon sets the layout and the donor sets almost nothing — IoU 0.536, 92% of ceiling, flat to 0.036 across four donors, and a rot180 output prefers the rotated label map by 30.6x (v1: 10.8x). By pass 2 that has inverted: same-donor/different-cartoon SSIM 0.901 against same-cartoon/different-donor 0.511, IoU down to 6% of ceiling. The harness's pooled verdict is again "YES, 3 of 5 checks", and again that pooling is what makes it wrong; split by pass it is unambiguous, and the v2 numbers are a stronger version of the v1 ones in both directions.

The donor NCC (0.342) stays well below the native positive control on the same donors (native_generate 0.409, native_img2img 0.642), so pass 2 is still not a spatial re-registration of the donor field — it inherits the donor's content and appearance, not its coordinates.

11. Looking at the images

packages/sweep-v2/eval/sheets-v1v2/v1v2-<category>.png — per category, rows = seed 11 x three cuts, columns = real held-out | v1 cartoon | v1 pass2-si3 | v2 cartoon | v2 pass 1 | v2 pass2-si6 | v2 pass2-si3. Every cell is the identical 768 x 768 px window (192 x 192 um at 0.25 um/px) at x = 1664, y = 640 of its 4096 x 2048 field, shown at 384 px; sheets-v1v2-1to1/ is the across row at 1:1 pixels. Real held-out cells are different tissue from different slides, not the same field. I looked at all four sheets and both 1:1 strips. Honestly:

• The v2 cartoon is a worse-looking drawing than the v1 cartoon at this scale. The 5.8-um granule lattice reads as a regular field of dark dots and dashes across the whole epithelial compartment; at 384 px it looks like a halftone screen with one or two big white lumens punched through it. The v1 cartoon's scatter of purple ovals with small white slots looks more like tissue. That is not a criticism of A1-V2's reasoning — the granules exist to make the ring metric read the wall — but it is what the generator is being shown.
• The v2 pass-1 output has inherited the lattice as nuclei. At 1:1 the normal pass-1 crop is a sheet of pale polygonal cells with round, uniform, evenly spaced dark nuclei and almost no chromatin texture: a plausible cell rendering with an implausibly regular arrangement. There is no basal palisade and no gland outline. This is the direct visual counterpart of ring_with_lumen_fraction = 0.000 at si6 for normal and gastritis: a uniform cellular field has no pale enclosed rings to find. v1's pass 1, by contrast, is a blur of the cartoon and looks like nothing histological, but it keeps the cartoon's holes.
• The v2 pass-2 outputs are the best-looking cells on these sheets. At 1:1 the IM si6 and si3 crops show epithelial groups with dark elongated nuclei that have visible internal chromatin, pale mucin vacuoles sitting inside the wall rather than fused with the lumen, pink fibrillar collagen and open pale spaces. This is a side-by-side impression of seven columns on one cartoon geometry, not a ranking of the whole sweep. The gastritis si6/si3 crops show recognisable columnar bands with a basal nuclear row — clearer than v1's si3, which reads as epithelium but with less organised nuclear rows. The IM goblet vacuoles in particular are a v2 feature: they survive into pass 2 in a way v1's radial mucin wedges did not.
• Normal is where the lattice hurts most. The v2 normal si6 and si3 crops are sheets of pale polygonal cells with regularly spaced round nuclei and thin pink septa. They look like plausible oxyntic cytology in a field with no glands in it. The real held-out normal column has obvious gland profiles with open lumina; neither v1 nor v2 produces those.
• Compared with the real column, every generated column is still visibly generated: more uniform nuclear size, less variation in stromal density, far fewer red cells, and none of the large-scale architectural organisation (a mucosa with a surface, a pit zone and a gland zone) that the real crops show even inside a 192-um window.

12. Is normal or gastritis better than v1 on any instrument?

Normal: yes, on four of the eight measures below. At the best arm (pass2-si3, n = 15 each):

Normal was "the weakest category in the sweep on every count" in EVAL_RESULT.md; on the v2 cartoons it is better on the envelope, on detail, on ring density and on ring-in-range, and it still puts nothing inside the envelope band. It remains the category with no visible glandular lumens.

Gastritis: only on detail. D 6.14 vs 5.77 (worse), ring mean 0.353 vs 0.394 and median 0.200 vs 0.294 (worse), ring-in-range 33.3% vs 46.7% (worse), in-band 100% vs 100% (level — the gastritis band is the wide one, C1 limitation 2), Laplacian 19.07 vs 18.60 (better), median ring area 405 vs 403 (level). The pass-2-makes-gastritis-worse effect is still there and is if anything stronger: pass 1 0.203 -> si6 0.028.

IM and mixed are level on the envelope (4.76 vs 4.67; 5.69 vs 5.78) and better on detail (20.13 vs 19.32; 21.69 vs 20.82). IM si3 is v2's best both-bands cell (6 of 15 = 40%); the best cell in either sweep is still v1's gastritis si3 at 7 of 15 = 47%.

13. What this package found, in order of consequence

1. The granule trick does not transfer. A1-V2 bought an in-band ring metric on the cartoon by putting dark 5.8-um granules in the epithelial cytoplasm so that ring_topology's 6-um closing would seal the gland wall. The generator reads those granules as nuclei, produces a regular cellular field, and the ring metric on the output collapses further than it did on the pit-sized rosettes. A conditioning signal has to be legible to the generator, not to the metric — a cartoon-side score is not a proxy for an output-side one. This is the single clearest result here.
2. The cartoon controls layout only at pass 1, on both cartoon sets. v2 sharpens the v1 finding (92% of ceiling at pass 1, 30.6x rot180 preference) and reproduces the pass-2 inversion exactly. The arm with the best envelope and detail (pass2-si3) is still the arm that has thrown the most geometry away.
3. Detail is where the v2 cartoons pay. +3.85 Laplacian at pass 1, +1.73 at si6, +0.85 at si3, in every category. Which of a1v2's changes drives that was not ablated here — the candidates are the darker fibre strokes, the stronger basement rim, the apical pale cap and the granules themselves — so "keep the detail knobs, drop the granule lattice" is a hypothesis this package supports but does not test.
4. Published probe numbers for the sweep are wrong (section 9). This is a harness/environment defect, not a modelling one, and the fix is one flag or a content-hashed cache key.
5. The v1 zero-byte image is recovered exactly (section 4), and the merge script now re-hashes every PNG so the same fault cannot be reported as coverage again.

14. What this package did not do

• No pathologist review, no clinical claim, no reviewer pass by Yash. Nothing here is "solved".
• The v2 cartoons were not re-tuned. The obvious follow-up — a v3 that keeps the gland calibre and the detail knobs but drops or lightens the granule lattice, accepting a worse cartoon-side ring score — was not run; A1V2_RESULT.md section 4 records that granule_lighten=1.0 costs the cartoon metric, and this package is the evidence that the cartoon metric is the wrong thing to protect.
• No si9 or pass-3 arm was generated for a1v2 (the brief specified pass 1 / si6 / si3 only), so the v2 ladder has three rungs against v1's five.
• The probe A comparison between v1 and v2 is not controlled: the discriminator pool differed (nine sweep sets vs two). Probe B is unaffected.
• The IoU ceiling for the VM label maps was not measured — the VM cartoon RGB (734 MB per set) was not retrieved, only the 20 MB of label maps. The Mac ceiling is used, and the two label-map sets are shown to have bit-identical lumen masks.
• --ncc-refs stayed at targeted: 222 of the ~17 000 exhaustive pairs. The copy screen is a screen, not a proof.
• Nothing was committed to git. probes/ JSON and PNG deliverables were verified byte-identical to C2's at the end (all 21 files); probes/embeddings/ was deliberately changed, as section 9 records, and the previous contents are backed up under packages/sweep-v2/eval/cache/probe-embeddings-standin-backup/.
• The probe caches were not made safe against the next package: the key is still the set name, so whoever measures a sweep-v3 under sweep_a1 will silently inherit sweep-v1's vectors unless they delete the file or run_probes.py starts keying on image content.

15. Files

Download public Markdown export