Vertical-sign detection — pass 8

1 · Headline numbers

141 + 44 admissions − 1 removal = 184, an exact match to the projection. The build is cp -r p7_final_* p8_final_* then admissions, then removals — removals last, the pass-7 invariant preserved. Re-running both apply steps against the already-overlaid tree gives applied 0 / NOOP and a recount of 184: idempotent. Every clusters.csv and verticalsigns.json under p8_final_a1 md5-matches its p7_final_a1 counterpart, zero diffs — A1 received no overlay at all this pass.

2 · The human session — 11 blind cases, 2 flips, controls 6/6

The package was blind (key held in p8_relook_key.csv), built on the proven pass-7 pattern: plain-path .bat, RGB-only PLY, cages, CloudCompare one-click. Standing authority rules applied unchanged — human HIGH is the truth apex, human MED against a standing LLM false-positive verdict is authoritative, LOW is a note, unsure changes nothing.

• Controls lifetime 6/6 (retro 4/4 from pass 7 + 2/2 here). Sol's caveat stands and is repeated in his seat: six selected agreement cases estimate neither sensitivity, specificity, nor candidate-selection bias. It is agreement evidence, not independent calibration.
• The LLM stack's vegetation bias is now 6 human overturns of veg/FP calls (4 in pass 7, 2 here) — 7 counting the rail-post s025 overturn in §3. This is a measured, repeated defect of the judge stack, not an anecdote.
• The field-stake family grew: s012c211 and s026c43 are both "stake in the field". Two of the four holdout "vegetation FPs" were never vegetation at all — the stack was wrong about the family, not just the verdict.
• A4_5 gantries are resolved as far as the three known disputes go: one gate kept, one pulled as vegetation, one unsure and flagged. Sol is right that this closes three cases, not the gate class.

3 · Rail half-posts — 101 → 45 → 43

~91% of A4_5 is guardrailed, and freestanding shape gates can never catch an Aufsatzleitpfosten — a stub bolted to the beam. Pass 7 researched a rail-relative search and never built it. Pass 8 built it, on Miro's directive to use the guardrail repo's own detector rather than our internal rail registry:

Environment note worth keeping: the guardrails pyproject.toml pins private Nexus packages and battlebox has no Nexus credentials, so uv sync was never attempted — the detector imports and runs unmodified under the verticalsigns venv, which already satisfies every version floor. The only degraded path is the XML exporter (STAND-IN XML exporter used); we consume guardrails.json only. No fallback rail extraction was needed.

The measurement that made it work

polyline_top_z_m is built from the median member height, so on many rails the beam's own crest sits up to ~0.5 m above it. Anchoring the search band on the fitted top fused every cluster into one 45-metre, 190 000-point blob per rail. The fix is to measure the crest directly — per dense sample, the p95 elevation of returns within 0.45 m laterally and −0.35…+0.55 m of the fitted top, gap-filled and rolling-median-smoothed over ±1.0 m of station, on the reasoning that a 0.15 m-wide half-post cannot lift its own floor. After re-anchoring the widest real-band cluster is 0.22 m along-rail. The fusion is gone.

Calibration was the two pass-7 rail_half_post admissions. One was rediscovered at 0.01 m (and on the neighbouring segment's rail, which is why the search is cross-segment) — a 0.11 m column standing clear of the beam with its retroreflector saturating at 65535 at +0.30…+0.50 m above the crest. That single candidate defined the gate; every threshold admits it with margin. The other was not rediscovered, because there is no rail under it: nearest detected rail vertex anywhere in A4_5 is 11.91 m away, and the lateral return histogram at the site shows a smooth embankment ramp where the seg125 site shows a sharp W-beam peak. The pass-7 type label on that device is impeached — it is a ground-mounted delineator on the verge; its admission is untouched, only its type is wrong (retype queued as pass-9 bookkeeping).

Null separations — the strongest label-free evidence in the pass

All four bands share identical geometry, clustering, features and gate; only the band's position moves, and every band is terrain-corrected to sit the same height above local ground. mirror and lat3 samples within 1.5 m of any detected rail are dropped, so those bands provably have no rail beneath them.

A saturated return plus multi-record support takes real:mirror from 4.8× to 35× and empties both rail-free nulls. Honest caveat, stated in the report before anyone asked: along30 is degenerate as a null and carries no evidential weight in its base form — a band displaced 30 m along the heading on a 91%-railed corridor lands back on rail, and its base rate (0.638) duly equals the real rate (0.626). Only mirror and lat3 are real nulls. And nulls bound the background rate, not precision: 35× enrichment says the band finds rail-attached structure, not that the structure is a delineator.

Judging: fleet, then Miro on every PNG

101 candidates, spanning 83 of 143 segments, strikingly homogeneous (median height above rail top 0.49 m, median width across 0.11 m, median maxI 51 816, median distance to the nearest already-accepted device 100 m — these are not re-detections). The 45 in the prime stratum (n_sat ≥ 1 ∧ nrec ≥ 2) were rendered as crop pairs and judged by Opus and Grok independently; Fable tiebroke 13 disputes by viewing, and all 13 went to real — Grok's vegetation credulity biting for the third pass running. Miro then reviewed all 45 PNGs himself before anything was admitted; the draft admissions were committed HELD pending exactly that.

rhp_s117_r0_c0 — the silhouette the judges converged on: a compact head on a thin neck, rooted at the beam crest, with the fuzzy vegetation mound to the left staying below and behind the rail. The along-rail panel underneath resolves the W-beam and its bare support posts for scale. maxI 62 935, n_sat 2, nrec 3 — Fable tiebreak, real/high.

The context panel every judge saw alongside the close-up — rhp_s019_r0_c0, 40 m of corridor from above. This is the view that keeps "regularly spaced verticals in a field" from being mistaken for a delineator row: the candidate (circled) sits on the rail line between the carriageway and the tree band, not out in the vegetation.

What this search still cannot see, stated up front in the report: the band floor at crest + 0.15 m is forced by beam fusion, so the NCHRP valley-mounted delineator class — reflectors set into the W-beam trough — is structurally outside the search and remains unmeasured. Rail-reference recall is also a hard ceiling: a half-post on an undetected rail is invisible (segment_126 has no right-side rail instance at all). And the 55-candidate non-prime tail is unjudged — which is precisely why it is pass-9 item 1.

4 · The edge-line gate — three rounds to get one directive right

Miro's directive: "Signs between the detected edge lines should be rejected. This is a car." Motivating case: holdout b0 s011 c88, the smeared single-record vehicle from pass 7. Then, after the first attempt: "Do not use the road mask. Use the lines from run7 … from the XML. Much more reliable. And much faster."

The XML's Axis of the Edge features carry a LaneID and a role (A_OuterEdge_3, B_InnerEdge_1) with spline control points, so a carriageway is looked up, not inferred: the band between one lane's outer and inner edges. The median is the space between the two lanes' inner edges and is therefore outside every band by construction — which is what lets the gate keep the full-carriageway promise without endangering median devices. The <Geoshift> is already the point clouds' frame, so no coordinate shifting is needed.

Census, round 3

Runtime is the "much faster" half of the directive: the XML path reads no point cloud at all — segment extent from tile metadata, edges from a cached per-corridor XML parse. A1 went from minutes to 2.9 s. Armed fraction rose from 22/33 to 32/33 on A1 (the one holdout is b0 053 at coverage 0.14) and 35/46 to 36/46 on A4_5, where only 6 of 31 XML files carry complete outer+inner lane pairs.

The decision rule stayed unchanged throughout: REJECT (onroad_transient) requires all of a gated type (gantry_or_gate exempt by type — its centroid is mid-span), on_road == 1 with a 0.3 m inward margin (delineators stand on the paint), h_max < 4.5 m, and transient corroboration (nrec ≤ 1 and n_records_local ≥ 2). Position without corroboration is a FLAG only. 408 tests green (17 new in test_run7_xml.py, 27 in test_edgeline.py).

Residuals, stated rather than smoothed over. A4_5's verge control fails 4 of 1139 in the unsafe direction (0.35%) — the binding null still passes, but that is the number to watch if the gate is ever widened. XML red flags: accepted judged-real devices reading as ON a carriageway — 2 on A1 (b1 seg008) and 8 on A4_5 (segs 025, 046, 066, mostly pole_other). None is rejected, because the corroboration rule and the height/type exemptions hold them, but the lane model disagrees with reality there and those segments are disarmed rather than trusted. And the census evaluated 76/139 devices against canonical 75/141 — a JSON-handling difference that owes a reconciliation (Sol's finding 6, pass-9 item 8).

5 · crown_circles — the regression that killed 6% of A1, fixed bit-identically

Pass 7's burn surfaced a MemoryError in radius.crown_circles that killed b0 segments 015 and 016 even on the 85 GB box. Pass 8 fixed it (ca39229: a bounded neighbour-chunk budget instead of one materialised radius-neighbour graph) and re-ran the two dead segments.

Under the identical 45 GB virtual-memory cap these two segments blew through in pass 7, peak RSS was ~2.5% of the cap. The fix is verified bit-identical to sklearn DBSCAN across 360 randomised configurations, zero mismatches, so the two segments are directly comparable to their 28 siblings; the suite went 364 → 408 tests. p7_holdout_verify.py now reports VERIFY OK: 34/34 segments valid on both sides.

6 · The saturated-head rule: refused, nothing built, reserve intact

Pass 7's holdout produced an obvious candidate rule — require a saturated/compact bright head for solo non-rail delineators — which would have killed all 4 leaked FPs at the cost of one real find, turning a PRIMARY failure into a pass. The strategist wrote two refusal terms before the human session. Both triggered.

7 · Panel: CONTINUE 3–1, with Sol dissenting

Fable's seat was sealed first (d4a0df6) before any other seat read the dossier, as in pass 7.

Sol's seat is worth reading in full rather than summarising away. The sharp version:

"Candidate adjudication cannot prove class completeness because it labels only surfaced candidates; a stratified human inventory or equivalent independent ground truth is needed to measure misses."

"Burning reserve merely to finish a checklist would test little about the unresolved A1 recall denominator."

And his self-check is honest in the other direction too — he names the strongest case against his own verdict and prices it into the 78%:

"Risk of premature stop: an A1 rail-relative probe could uncover a cheap repeat of the A4_5 gain … If that probe yields a material, independently auditable A1 population, the premise that only more data can move the main result would be weakened and the panel should reopen."

Sol's dossier-overreach findings — all seven accepted onto the record

1. "The pass-7 directed queue is fully executed" is too absolute. The strategist's burn-LAST and half-split amendments were not executed. The rule branch closed without a burn — correct, but the dossier should say that rather than claim every directed action ran.
2. The holdout retro-pass validates the pre-registered FP:gain criterion, not corridor recall. It "does not measure unseen missed devices". A1 recall remains instrument-weak at coverage 0.20.
3. 6/6 controls is agreement evidence, not independent calibration. Six selected agreement cases estimate neither human sensitivity, specificity, nor candidate-selection bias.
4. The 43/43 human ratification is not auditable from the rail report Sol was given. True and fair: it lives in the verdicts_pass8 CSVs, the admissions provenance strings and Miro's message — auditable, but poorly cross-referenced. It also covers the prime stratum only, not the unjudged tail or the valley-mounted class.
5. "The A4_5 rail-band search is consumed" is wider than the actual audit — and the dossier's "35× worse null separation" for the 55-tail was a misattribution: 35× is the prime stratum's real-to-mirror enrichment. The source calls the complement "near-null-rate" and gives no comparative factor. That one was simply wrong in the dossier.
6. Census device counts (76 / 139) are not reconciled to canonical (75 / 141). JSON-only handling differs. The zero-reject claim holds for census inputs; the reconciliation is owed.
7. Three-case gantry adjudication closes the known disputes, not the gate class.

Recorded alongside: Opus's pre-committed falsifier — STOP-3 next pass if the pass-9 queue yields fewer than 5 A1 gains, an uninformative-or-empty instrument, and a clean burn — and Sol's four reopen conditions. Both dissent directions are on the record before the evidence, which is the point: pass 9 is built so the next verdict is decisive either way.

8 · Pass-9 queue (panel-directed)

1. Judge the 55-candidate non-prime rail tail. The control arm. Converts +43 into a measured precision and closes Sol's overreach-5 as mandatory hygiene. Same protocol, fleet + tiebreak; expect near-null yield.
2. Half-split record-partition Chao2 repair on A1. Replace single-drive occasions (a degraded protocol, per-drive p≈0.05, coverage 0.20) with random disjoint half-partitions of the record set, Chao2/jackknife across many replicates. Explicitly not the banned Chapman split-half. Pure battlebox compute, judge-free, zero holdout contact.
3. A1 rail-relative probe. A1 is lightly railed so it is cheap — and Sol names it the one move that could reopen a material A1 delta.
4. Valley/trough-mounted rail-delineator band on A4_5 — the class structurally excluded by the pass-8 band floor.
5. Rescue sieve over unjudged full-run rejects (short_bare_post, unclassified) — a new physical signature applied to never-judged detector-internal rejects, not the exhausted veg-rescreen of already-judged items. 7 lifetime veg-overturns say gold remains.
6. TREES SCOPE DECISION → Miro. Opus and Sol converge that stop-1 plausibly already holds for trees (the vegetation RF has no real supervision). Formal descope or supervised re-entry is his call; runs in parallel, blocks nothing.
7. Reserve burn LAST and ONLY IF ≥5 plausible AFTER-only A1 gains accumulate. Full stack, per-rule toggle columns computed in the same single read, pre-registered before any read.
8. Bookkeeping: seg126 retype; census-count reconciliation; s066 stays flagged; the 12 half-vs-full unclears carry their note.

Baselines: p8_final_a1 (75) / p8_final_a45 (184) on battlebox · holdout CONSUMED and retro-passed · reserve b0 091–136 intact · instruments in out_eval/pass8/ · zero human training labels · reports: pass 7 · pass 6 · journey