Where to improve next

1 August 2026 · AI3D-339 follow-on · deep-research synthesis for the point-cloud team

This page retracts its own first headline. The sweep concluded that delineator recall was the dominant defect: 88 posts found where regulation spacing implies 286–572. That assumed an unrailed corridor. It isn't one — ~91% of the two-sided corridor carries guardrail, and German RPS doesn't place Leitpfosten along railed stretches. Corrected expectation: ~26–51 free-standing posts; we detect 88. The gap doesn't exist. What survives is narrower and still true: recall is unmeasured, and we now know the regulation-based proxy cannot measure it — so the label-free instruments below become the priority.

How the wrong headline survived 36 agents. The adversarial verifiers were effective within their scope — several killed techniques with live measurements. But every agent reasoned inside the verticalsigns repo. None had the domain fact that guardrails displace delineators, and none consulted the sibling repo that already measures guardrails on the same corpus. The structure tested “is this technique real?” and never “is the problem real?” — so a confident, well-evidenced, wrong headline passed 36 agents and two independent critiques, and was overturned in one sentence by someone who knows the road.

The recall question

delineators detected

143 segments · out_final sweep

26–51

expected, rail-corrected

RPS spacing over the unrailed 9%

91%

of corridor is guardrailed

1.82 of max 2.0 rail-m per corridor-m

0 / 88

posts within 1.5 m of a rail

median distance 23.5 m — all free-standing

The denominator was wrong, not the detector

Expected post count under two models, against 88 actually detected

Detections read from out_final/segment_*/verticalsigns.json; rail coverage from the guardrails repo's out_full run on the same corpus. Both expectation bands are estimates, not surveys — but they differ by an order of magnitude, which is the whole point: a denominator this sensitive to an unstated assumption cannot support a recall claim either way.

Posts appear where a side is free of rail

Mean delineators per segment, by how many sides carry guardrail

The segment-level correlation is misleading here (Spearman ρ = +0.32, the wrong sign) because rail is nearly universal — 142 of 143 segments carry some. The contrast only appears per side. Segment 131 is the archetype: 99 m of rail on the left, none on the right, 18 posts.

Blind visual audit — 4 of 4 confirm

Four vision agents were given only the perspective renders for four segments, told neither the hypothesis nor which arm their segment was in. Every segment with a free side showed a row of free-standing posts; neither both-railed segment did.

Segment	Detector	Judge: guardrail	Judge: posts
131	left 99 m, right 0 m · 18 posts	absent on the post-bearing road (high)	present on both verges, ≥8, plausibly 10–14 (high)
134	left 99 m, right 0 m · 6 posts	median side present, outer side bare (both high)	5–8 in one regularly spaced row (medium)
019	left 50 m, right 50 m · 0 posts	present on both outer verges (high)	1 ambiguous candidate, none confirmed
011	left 50 m, right 50 m · 0 posts	reported absent — wrong, see below	none (high)

Segment 134's judge ruled out the confound that matters most, unprompted: the posts “stand inboard of and slightly taller than the guardrail beam — separate posts, not reflectors on a barrier” — the same conclusion the 1.5 m distance test reached numerically. Where counts differ the detector is higher, which is expected: the renders crop near-field verges and truncate at tile boundaries, so a judge sees a subset. No judge reported posts the detector missed.

Calibration, and a render bug. Segment 011's judge called the rails absent where the detector reports two 0.6 m w_beam runs. The detector is right — across all 367 rails in the corpus, mean height spans 0.452–0.709 m with zero at paint height. But segment 019 proves rails are visible when the framing cooperates: what differs is camera geometry. Low-oblique views along the road resolve a beam clearly; near-grazing top-down views flatten it into the shoulder, and those were all segment 011 had left — because in all four segments only about 2–4 of 9–12 renders carry usable evidence, with byte-identical blank frames in two of them. Treat a judge's “absent” on low horizontal structure as “not visible in these framings.”

Where those 88 posts come from

One cell per segment, in order 000 → 142

0 posts (114 segments) 1–2 3–6 18–25

The concentration is real and extreme: Gini 0.910, Fano factor 11.5 where Poisson would be 1.0, and 13 segments hold 80% of all posts. A uniform rate of 0.615 posts per segment would leave 77 segments empty; 114 are. But the structure is explained by guardrail coverage, not by detector failure — the empty cells are railed stretches where posts are not installed.

Two corrections to the working notes. The critique cited “54 posts from segments 131/132/134/135” and “~135 zero segments”; the measured values are 52 and 114. More seriously, it read this concentration as a recall failure. It is mostly a map of where the guardrails aren't.

What the sweep returned

Five search angles, 40 unique sources, 15 deep-read, every extracted technique put through an adversarial verifier that was told to refute by default. Of 31 techniques, 21 were killed — several by measurements the verifier ran on battlebox during review, not by argument. Nothing was rated above low payoff.

Technique verdicts by search angle

After adversarial verification · 31 techniques from 14 sources

survives (low payoff) dead unverified / extraction failed

The pattern is uniform: the literature keeps proposing things this repo has already built (the taper family), already ships (ring_fill_ratio, the p98 intensity gate), or cannot run on this data. Two extractions failed outright — both in the label-free angle, which is why the weak-supervision ideas below come from the critique pass instead.

The 13 adjudicated false positives, and which lever attacks them

From data/qc_verdicts.json · visual-QC verdicts on the 143-segment sweep

record multiplicity (static vs dynamic) plate normal vs road tangent neither — vegetation confusion

Two cheap additive features plausibly reach 9 of 13. Neither exists yet, and neither needs a label, a retrain, or a new data channel.

Recommended next steps

Delineator recall audit highest magnitude
Project detections onto the road centreline to get station, then autocorrelate per side. A 25 or 50 m peak confirms the class is being found systematically; the gaps at integer multiples enumerate the missed posts by location — a targeted worklist and a set of predicted-position weak labels, with no annotation anywhere. Then examine the three mechanical suspects: seed_min_vertical_span_m = 0.80 against a ~0.7 m usable post above the 0.30 m floor; the self-referential p98 brightness gate (one large guide panel raises segment p98 and suppresses the post class); delineator_min_points = 300.
Record-multiplicity feature + split-half repeatability ~5 lines
n_records_present, record entropy, max per-record point share — purely additive, and candidate_record_id is already computed in grid.py and carried into process_segment. Then run the detector on odd vs even records, match by XY, report agreement: that number becomes the project's standing label-free evaluation metric, and the consensus set is the auto-label seed for the rules → auto-labels → human-correction ladder.
Threshold-margin census as a standing instrument
For every accepted detection, distance to each conjunct's threshold. Tells you which of ~40 hand-tuned constants are load-bearing and which are inert — and becomes the cheap pre-filter that kills future technique proposals before they cost a research round. Phase 6 did this once and it produced the sharpest result on that branch.
Plate-normal vs road-tangent scalar hours
_plate_thickness already eigendecomposes the upper-half covariance and throws the eigenvector away. One additive |cos| scalar between plate normal and corridor tangent attacks the guardrail top, both road-paint FPs and the verge strip: a real plate faces traffic, guardrails and paint run parallel to it.
Merge-splitting: deprioritize, and run the zero-code probes first payoff arithmetic
It remains the tree path's #1 defect, but the adjudicated merged population is about six objects on a path that ships off by default. Before writing a splitter: a config-only eps/hull-margin sweep on the known-merge segments, connected components at ~0.3 m on the six dumped masks, and a footprint-bimodality census. If those say “splittable”, the paper-backed design is multi-stem flag → Dijkstra voxel-graph assignment (Hui 2022).

Experiment portfolio

Fourteen probes, sorted by cost inside four goals. Several are deliberate long shots — the brief asked for experiments that may or may not pay off, and the cheapest ones buy information either way.

Cost of each probe

Log scale, 20 minutes → 1 day · solid = expected to pay off, ring = speculative

The scanline-order probe has the best information-per-minute in the report: if acquisition order survived the exporter, scanlines give free neighbourhood topology, along-scan normals and occlusion structure, and the “no sensor geometry” blocker cited by six separate findings evaporates. If the array is shuffled, that whole family closes forever for the cost of one plot.

All 14 probes, with what each one buys

Probe	Cost	What it buys or kills
Scanline recovery — walk one run3 file in stored row order, plot consecutive-point distance periodicity	~20 min	Unlocks or permanently closes the entire sensor-topology family (range image, along-scan normals, ray-cast occlusion reasoning).
λ̂₁ closed form — compute linearity from the existing `omnivariance` and `change_of_curvature` columns	zero new code	Normalized eigenvalues have 2 DOF, so λ̂₁ is already implicit in the dumped CSV (verified to 1e-15). Tests whether explicit linearity separates known pole FPs from accepts.
Dawid–Skene EM over the judge panels already collected	~2 h	Per-judge reliability and a soft consensus label. Directly tests whether “no shape feature separates conifers” — the verdict that killed four techniques — rests on reliable judges or one confident outlier. At n≈13 that verdict is no evidence, not a refutation.
Off-ground intensity bimodality — histogram candidate-point intensity against the p98 actually in use	~2 h	Quantifies how content-dependent the delineator brightness gate is; tests whether a mode split is stable where a percentile is not.
Nadir trajectory reconstruction — per record, take \|scan_angle\| ≤ 2°, test whether the XY locus is a 1D curve	1–2 h	Sensor path → per-point range → true incidence angle. Reopens radiometrics on a defensible basis and enables ray-cast reasoning for the fusion and occlusion failure modes.
Grid-phase stability — re-run seeding at four half-cell XY offsets	cheap	Flags detections that exist only because of where the occupancy grid happened to land.
Point-thinning stability — keep every 2nd/4th point	cheap	Separates robust geometric cues from density artifacts — directly relevant to the conifer density band.
Multi-radius texture — recompute curvature/omnivariance at 2–3 neighbourhood radii	cheap	Whether scale robustness recovers any of the fence/bush confusions a single radius misses.
Road-referenced intensity normalisation — per-file median over the run4 road mask	read-only pass	The one surviving radiometric fragment: a label-free per-file gain reference. Reject it if it tightens one proxy population by collapsing another.
Leave-one-record-out feature jitter	~half day	A free per-detection confidence score. Any detection that survives via a single record is a ghost.
Station-periodicity autocorrelation	~half day	Confirms systematic detection and enumerates missed posts by location — the recall worklist.
Windowed continuity counter on the 13 curated conifer candidates	~1 day	The one taper-family variant not yet falsified: longest run of adjacent decile-pairs passing a size-monotonicity test, evaluated only on masks that already exist.
Per-record multi-view agreement — recompute cluster features per record, require consistency	~days	The correct reading of “evaluate per azimuth”. A line-of-sight-fused pair is a cone from one record and two blobs from another; a real cone is a cone from every record. Object-centric sector binning was measured saturated — 8 of 8 sectors for delineators, signs and trees alike.
Copy-paste recall probe — translate verified masks from `point_masks.npz` into ~20 segments at known coordinates	~1 day	Manufactured ground truth with zero annotation: exact recall vs range, vs record count, vs point count. The decimation variant (75/50/25% of points) prices `delineator_min_points = 300` directly.
Corpus audit — does any corpus have real RGB in leaf-off? Do A1 and A4_5 overlap geographically?	~hours	Season and colour availability are perfectly confounded across corpora, so no colour result is currently separable from a season result. Geographic overlap would give repeat-epoch pairs — another free-label source. A2 and A3 are unused.

What's dead, and why

Two of these deserve promotion to repo-level constraints so they stop being re-argued per technique.

promote to constraint Multiplicative radiometric corrections

Cosine-incidence, R² range falloff, proxy-target calibration — all dead, and measured during verification. Road intensity is flat against |scan_angle| even though that binning gives a 3–4× range lever, so either the product is pre-normalised or the field is inert. More fundamentally the intensity channel is log/dB-quantised (successive representable values in a constant 1.00199 ratio, 2168 levels, 6:1 span), which makes every multiplicative operation the radiometric literature specifies undefined on it.

promote to constraint All colour work on A1

Dead upstream, before any index argument starts. Phase 6 §9.3 measured that of the 18 accepted clusters the colour rule can reach, the volumetric clause (plate_thickness_m ≥ 0.175) passes zero, and 25 threshold combinations changed no verdicts. No colour index at any threshold can alter any A1 output until that clause moves. Separately, GLI is not a second cue: it shares ExG's numerator exactly, and measures Spearman ρ = 1.000 against it.

The taper family — one technique, already built

Slice-area trunk run, adaptive-radius local taper, knee-based crown profile: four names for the thing implemented as _decile_radii / _conicity / _crown in the conifer worktree, and already measured. Caveat: that negative rests on n ≈ 11–13 objects adjudicated by blind agent judges. At that sample size it is a no-evidence result, not a refutation — which is what the Dawid–Skene probe is for.

Trunk-seeded tree delineation

Assumes urban street trees scanned at close range with clean visible stems. This is a highway corridor: embankment crests hide lower trunks, and the conifers that were actually found sit 24 and 31 m back behind verges and barriers.

Anything that changes cluster membership in place

Slice-and-link candidate building, supervoxel re-segmentation, gap-threshold region growing. All of them move every frozen FEATURE_NAMES input of the locked RF, breaking the additive-only and byte-identical regression gate. Viable only as experimental parallel paths.

Three techniques that are descriptions of shipped code

The ring-isolation gate is ring_fill_ratio; the per-segment intensity percentile is hi_intensity_all_points_percentile = 98; the billboard-size ceiling is panel_len_major_max_m. The literature sweep rediscovered existing code three times — which is itself a useful signal about where the pipeline already sits relative to published practice.

Beware the vacuous census: any tree-path-only change trivially zero-diffs the 143 A4_5 segments because tree_detection.enabled defaults false. That null is not evidence of regression safety — the gate it claims to pass was never exercised.

Sources worth keeping

Source	What it gives
Hui et al. 2022, Forests · `10.3390/f13081245`	Multi-stem under-segmentation flag + Dijkstra voxel split; F1 0.81 roadside. Fails on ground-touching branches — our own skirt failure mode.
Cabo et al. 2014, ISPRS JPRS · `10.1016/j.isprsjprs.2013.10.008`	Slice-and-link pole detection, the classical baseline.
Li & Oude Elberink 2013, ISPRS Annals II-5/W2	Percentile-windowed continuity; N-of-M cue voting.
El-Halawany & Lichti 2013, GIScience & RS · `10.1080/15481603.2013.866815`	Road-pole detection; PCA linearity plus RANSAC line support.
Tao et al. 2015, ISPRS JPRS · `10.1016/j.isprsjprs.2015.10.007`	Shortest-path crown segmentation from ecological theory.
Li et al. 2016, ISPRS JPRS · `10.1016/j.isprsjprs.2016.07.009`	Dual-growing individual tree extraction from MLS.
Kashani et al. 2015, Sensors · `10.3390/s151128099`	Radiometric processing taxonomy — the reference for why our channel supports none of it.
Wen et al. 2016, IEEE T-ITS · `10.1109/TITS.2015.2418214`	Traffic-sign inspection from MLS.
Remote Sensing 2023 · `10.3390/rs15133254`	Vegetation-index thresholds for coloured point clouds (SCND) — superseded by phase 6's three-population fit.
JSTARS 2019 · `10.1109/JSTARS.2019.2929546`	Rapid roadside tree inventory; height-slice profiles.
Schwab & Kolbe 2026 · `arXiv:2603.11252`	Radiometric fingerprinting. Needs an LoD3 semantic model and a trajectory, so dead here — but the road-surface gain-reference fragment survives.
Brunner & Houtmeyers 2022, Eur J Forest Res · `10.1007/s10342-022-01481-5`	Conifer crown segmentation in mixed Scots pine–Norway spruce stands.

Method & caveats

Five parallel Codex web-search agents (one per angle) → URL dedup, capped at four sources per angle → 15 deep-readers fetching and extracting techniques mapped onto concrete pipeline stages → an adversarial verifier per source, instructed to assume the technique does not transfer → two independent completeness critics, one search-grounded and one reasoning from first principles. 36 agents, ~1.8M subagent tokens.

Two deep-readers emitted schema stubs, both in the label-free angle — so that angle returned essentially nothing from the literature, and its content here comes from the critique pass instead.
Verdicts described as measured were backed by real runs on battlebox during review, not by argument alone.
All delineator counts on this page were re-derived from out_final/segment_*/verticalsigns.json after the first draft's figures failed to reconcile with data/qc_verdicts.json, which is a QC sample of 23 confirmed detections and 13 confirmed false positives — not a census.
The expected-count band is an estimate from regulation spacing, not a survey. It assumes posts on both sides for the whole corridor and so overstates the gap somewhat; the audit in step 1 is what would replace it with a measurement.

Full working notes: docs/research/deep-research-20260801.md in the repo.