RampNet 1.0 closeout: merge queue, what to close, what moves to 2.0

This is the checklist for calling RampNet 1.0 finished so the repo's attention can move to RampNet 2.0 (#86). It is a triage of every open PR and issue as of 2026-09-15, bucketed by what actually has to happen. The bucket for each issue was checked against its thread's last comment, not its title.

The test for "finished": every 1.0 result is merged, findable from one page, and carries its caveat beside the number. Nothing here requires a GPU except the optional seeds in F.

A. Merge queue

Seven PRs are open; five have sat since 2026-09-03 with finished work in them, some of it paid.

• #161 — seed variance scored: s_gap 0.0114, ambiguous band, the noise is RampNet's. Closes the last open item on #51 and #135.
• #160 — launcher docs: the REPO=… RAMPNET_ENV=… invocation that actually runs.
• #147 — time and money accounting (#143). compute_log.jsonl is not on main until this merges; compute spend has been going into issue comments in the meantime.
• #146 — claude-opus-5 on nine splits (#139). $70.41 of scored detections.
• #137 — Vistas at resolution parity (#126). The issue has been treated as closed since 2026-08-18; the PR never merged.
• #136 — the two collapsed Vistas runbook snippets.
• #157 — Claude Fable 5/5.1 legs (#156). Decision needed: a roster refresh is not a 1.0 result. Recommend closing the PR and #156 and refreshing the roster when 2.0 has a model to compare.

B. Real work still open for 1.0 — bounded

• #148 — the #140 seam wrap moved three committed YOLO baseline cells and benchmark_eval/ holds the pre-fix values. Re-scores committed detections; no GPU, no cluster. The numbers the paper would quote are currently wrong in one place.
• #127 — carry the review notes and Bend's train-overlap flag into the published rampnet-benchmark. records-only rebuild plus a card push; the residual of #21.
• #72 — the YOLO instability caveat, retitled per its last comment: all five configs collapsed at the epoch-3 warmup peak, not YOLO11-pano specifically. One paragraph in scripts/model_comparison/yolo_baseline/README.md or docs/model_comparison.md.
• #131 — replicate #46 Phase 1 on a second machine. silent_activation.json has one provenance and §0c of curb_ramp_data_sourcing.md rests on it. Cheap (makelab2, committed caches); do it or state in the doc that it was produced once.
• #145 — split the runbooks out of model_comparison.md and curb_ramp_data_sourcing.md. Optional for closeout, but the paper is written from these files and neither can be read whole.
• #74 — Budapest second rater. Blocked on a human. Either schedule a rater or close with the caveat that Budapest is single-rater, LOW confidence, and carries the benchmark's only ranking inversion.
• Seam residual from #132: the published Stage 1 dataset double-labels 8,361 seam ramps and drops ~71% of seam ramps. The detection-deficit claim was retracted; the dataset defect stands. Decide whether it is fixed in the HF dataset now (a rampnet-dataset revision) or carried as a known defect into #150.

C. Close as done — the finding is recorded, the issue just never got closed

Each of these has its result in a merged doc or PR. Closing comment should point at the doc.

• #51 — after #161. Final read: gap of replicate means 0.0115 ± 0.0065, published 0.039 was a favourable-vs-unfavourable draw; manual_gold YOLO 0.912 > RampNet 0.905 at matched thresholds.
• #135 — after #161. Run B decided not-run 2026-09-03; A1.2 bounds every unpaired single-seed comparison at 0.011.
• #126 — after #137. Transfers but does not compete; naive union dead; cascade ceiling ~19 ramps, FP cost unmeasured.
• #139 — after #146. The annapolis edge reversed on the other eight splits.
• #143 — after #147.
• #144 — the recommendation not to run claude-opus-5 on manual_gold alone is the answer; recorded in the issue body.
• #58 — manual_gold is a live split, full roster, RampNet 0.908 vs gemini-pro 0.568. Its own last comment recommends closing.
• #44 — clovis landed and the diversity ask was overshot (11 splits). Its last comment recommends closing.
• #49 — clovis is in the published benchmark (9 splits shipped in #21).
• #39 — OWLv2, Grounding DINO, Molmo all on every split. Its last comment recommends closing.
• #46 — Phases 0+1 merged (#94, #99); residuals spun out to #130/#131; the second-rater pass has no rater.
• #43 / #130 — the seam scoring defect is fixed in #140 (matcher wraps; 10 manual_gold pairs adjudicated and merged, seam_verdicts__jon.json). Close both; the dataset-side residual is the B item above.
• #132 — same. The "RampNet is blind at the seam" result was retracted (skimage exclude_border); the scorer fix and GT adjudication are merged.
• #59 — answered in curb_ramp_data_sourcing.md (#93): 500k is not reachable on assessed data; composition beats volume. The follow-on work is #96, which is 2.0.
• #20 — the harness exists (scripts/model_comparison/, rampnet/roster.py); everything it asked for was built under other issues.

D. Won't-do for 1.0 — close with a one-line reason

• #90, #70 — YOLO-pano LR tuning / stabilized rerun. Superseded: the tiles arm is the reported baseline and it is stable (s_A 0.0023).
• #95 — attributing the RampNet-vs-YOLO gap across backbone / representation / budget. The gap to attribute is ~0.01 F1, inside one s_B; there is nothing to decompose.
• #52 — YOLOE. Another zero-shot entrant will not change a ranking RampNet leads by 0.12–0.34.
• #45 — VLM prompt study. Same reason.
• #47 — nadir/hood mask in the reprojection rig. A VLM-cost optimisation for a leg we are not extending.

E. RampNet 2.0 — move, don't close

Live proposals with data behind them. Recommend a RampNet 2.0 milestone (the repo has none) and moving these onto it untouched.

• #86 (roadmap), #48, #83, #107, #113, #111, #112, #101, #102, #158, #159, #149, #151
• #35 — hybrid/cascade; #126's cascade-gate run says the ceiling is ~19 ramps and an arbiter-on-crop is the load-bearing piece.
• #25 — input resolution; forecast-backed by the depth analysis.
• #38 — per-ramp deployment recall / sampling density.
• #82 — augmentation as a transfer lever.
• #85, #96, #103 — the 500k data-sourcing track (aerial as source, precision gate, street-level review instrument; #105 merged, Denver sheet awaiting review).

Separate tracks, not 1.0 and not 2.0: #150 (FAIROS release, grant deliverable) and #114 (crop-window scoring for sidewalk-auto-labeler; lives here only because the gold boxes do).

F. The page that makes "closed out" true

• docs/rampnet1_findings.md — one line per headline result, with the doc/PR that holds it and its caveat. At minimum:
  • RampNet vs supervised YOLO on the same data: ~0.01 F1 at matched operating points, not 0.039 (seed_variance_51_135.md, operating_point_parity_51.md)
  • Seed variance bounds every unpaired single-seed comparison at 0.011 F1 (A1.2)
  • RampNet vs 8 VLMs / open detectors: leads by 0.12–0.34 on all 11 splits (model_comparison.md)
  • manual_gold: YOLO 0.912 > RampNet 0.905 at matched thresholds; RampNet 0.908 vs best VLM 0.568
  • Operating point 0.55 → 0.30 uniform (operating_point.md)
  • Stage 2 recipe: 1 epoch is the plateau start, no resolvable peak, Run B not run (stage2_epoch_curve_84.md, stage2_cosine_rung_135.md)
  • Vistas transfers but does not compete; naive union dead (vistas_parity_126.md or wherever #137 lands it)
  • The seam: scorer defect fixed, detection-deficit claim retracted, dataset defect open
  • Laurens: the deficit is the rig, not the town (#151)
  • Data scaling: composition beats volume; 500k not reachable on assessed data (curb_ramp_data_sourcing.md)
  • Miss taxonomy: σ/calibration, not vocabulary; sourcing prize 0.013 recall pts (#46)
  • Budapest: single-rater, LOW confidence, only ranking inversion
• Optional, free: RampNet seeds 4–6 on klone ckpt-all to tighten s_B (n=3 → n=6 moves the mean-gap read from ~1.8σ to ~2.4σ if the mean holds). Not blocking; the staged dataset on /gscratch/scrubbed has a purge clock at ~2026-09-29, so it is now or never. Cancel if not started within a week.

🤖 Generated with Claude Code — Opus 5 (1M context), claude-opus-5[1m]

Start with the checklist in this issue and read the referenced findings documents, especially docs/rampnet1_findings.md, curb_ramp_data_sourcing.md, and model_comparison.md. Trace each 1.0 item to its referenced issue or PR; done means the required merges, closures, decisions, and caveats are recorded on one findable findings page, with unresolved dataset and rater decisions explicitly handled.