Z-slice segment-counting line (negative result)

[trisha-ant]

Stacked on #1. This PR contains the full z-slice segment-counting line of inquiry (harness + three variants, previously split across #2–#4, now consolidated here). All results in this PR are on claude-opus-4-6; the Opus 4.7 re-test of zslice_multi is in #6. Retrospective on the full arc below.

---

Retrospective: z-slice segment-counting (PRs #3–#5)

Hypothesis

Max-intensity projections collapse depth, fusing overlapping body segments into one bright band and destroying the discriminative feature for fold stages — how many times the body folds back on itself. A biologist resolves this by scrolling z-slices and counting body-segment cross-sections at a single plane: ≤1 → 1.5fold, 2 → 2fold, ≥3 → pretzel. The hypothesis was that a VLM subagent, given a non-projected z-slice, could perform this count and use it to override the projection-based classifier on fold stages.

What was tried

Three variants, each refining the previous after a negative result:

PR	Variant	Approach
#3	zslice	Single midplane slice (z=Z/2). Symmetric override: count → stage, replaces primary on all fold stages.
#4	zslice_asym	Same single midplane. Upgrade-only override (1.5fold→2fold→pretzel, never backward) to cap the pretzel regression from #3.
#5	zslice_multi	Five z-planes at z={0.25, 0.40, 0.50, 0.60, 0.75}. Count each concurrently, take the max. Symmetric override to test the hypothesis cleanly.

All three run perceive_scientific first and only invoke the subagent when the primary lands on a fold stage, so the subagent never moves predictions into or out of the fold region.

Results

n=233 hard-stage frames (1.5fold + 2fold + pretzel), 4 embryos. Baseline is scientific re-run on the same subset.

Stage	scientific	zslice	zslice_asym	zslice_multi
1.5fold (41)	65.9%	51.2%	19.5%	2.4%
2fold (59)	61.0%	59.3%	79.7%	44.1%
pretzel (133)	88.0%	6.8%	85.0%	54.1%
Exact	77.3%	27.9%	72.1%	42.5%
Adjacent	99.6%	89.3%	100%	95.3%

Override attribution (isolating the subagent's contribution from primary-classifier variance):

Variant	Override fired	Helpful	Harmful	Net
zslice_asym	7 / 88 triggers	0	7	−3.0%
zslice_multi	111 / 227 triggers	26	85	−25.3%

Why it did not work

The subagent cannot count. The core assumption — that a VLM can reliably distinguish 2 vs 3 fluorescent-nuclei blobs in an optical section — is false at this image resolution and contrast.

1. Single midplane returns ≤2 almost always. Across 88 zslice_asym triggers, the count distribution was {1:40, 2:46, 3:2} — 98% ≤2 regardless of true stage. A pretzel embryo's folds sit at different z-depths, so one plane rarely intersects 3+ segments at separable locations; what does intersect tends to merge into one blob. The symmetric zslice variant therefore systematically downgraded correct pretzel calls to 2fold/1.5fold (88%→7%).

2. Asymmetric gating works as damage control but reveals zero signal. zslice_asym restored pretzel to 85% by refusing to downgrade, but the override only fired 7 times in 88 triggers and was wrong all 7. The large per-stage swings vs scientific (1.5fold −46pp, 2fold +19pp) are run-to-run variance in the primary classifier, not the override.

3. Multi-plane max amplifies noise instead of recovering signal. Taking the max over 5 noisy estimators with a ~20% per-slice false-3 rate yields a ~67% any-slice-says-3 rate. The decisive measurement: max≥3 fires at 56% on GT=2fold vs 55% on GT=pretzel — statistically identical. No depth reliably hits 3 (per-depth mean count on pretzel: 1.83–2.19 across the five planes).

Root cause: in fluorescent-nuclei light-sheet data, body segments at fold stages are dense clusters of bright dots without clear membrane boundaries. The VLM perceives a single z-slice as 1–2 diffuse bright regions; it cannot resolve the gap structure that a human counts after contrast adjustment and scrolling. The count task needs either a different imaging modality (membrane marker) or classical image processing (threshold + connected components) rather than a VLM.

(We later re-tested zslice_multi on claude-opus-4-7 to see if its improved vision rescues segment-counting — it does not; max≥3 still has no 2fold/pretzel separation. See #6 for that result alongside the rest of the 4.7 work.)

What's salvageable

• Remove confidence from PerceptionOutput and all downstream uses #1 (confidence removal) — independent cleanup, should merge.
• Add midplane z-slice to TestCase and thread through runner #2 (z-slice harness plumbing) — midplane_b64 / zslices_b64 optional kwargs are general infrastructure. Keep, even though the consumer didn't pan out; future variants may use raw slices for other purposes.
• Add zslice variant: scientific + midplane segment-count subagent #3–Z-slice segment-counting line (negative result) #5 — documented negative results. Keep the variant code and result JSONs for the record.

---

What's in this PR

Harness (benchmark/testset.py, run.py):

• _create_slice_image(volume, z, max_dim) — render one XY z-slice as base64 JPEG, same crop as the projection
• TestCase gains midplane_b64: str (z=Z//2) and zslices_b64: list[str] (z = {0.25, 0.40, 0.50, 0.60, 0.75})
• run.py: _accepted_optional_kwargs() introspects each variant's signature and passes optional z-slice kwargs only if declared — zero churn to existing variants

Variants (perception/):

• zslice.py — scientific + midplane segment-count subagent, symmetric override
• zslice_asym.py — upgrade-only override on 1.5fold/2fold
• zslice_multi.py — max segment-count across 5 z-planes, symmetric override

Docs: program.md × 2 — document optional z-slice kwargs.

Verification

• 18 variants registered; _accepted_optional_kwargs returns {'midplane_b64'} for zslice/zslice_asym, {'zslices_b64'} for zslice_multi, set() for all others
• All three variants run end-to-end on 233 hard-stage frames; result JSONs + chart committed
• Count distributions and override attribution extracted from reasoning field in saved predictions