Far-field anomaly Phases 0+1 (#46): the pixel floor falls, and 'silent' is almost never absent

[jonfroehlich]

Phases 0 and 1 of the four-phase far-field study designed on #46 (2026-07-31): before the 34/36-visible verdicts are allowed to mean anything, check the sample they were made on and the pixel-starvation framing against the model's own behaviour (Phase 0), then read the model's heatmap inside every silent miss's match window to ask whether "silent" ever meant "saw nothing" (Phase 1). Phase 0 is committed data only — no GPU, no network, no imagery. Phase 1 ran the published checkpoint (projectsidewalk/rampnet-model, the same weights every committed cache came from) on a local RTX 3070, single-pass fp32 matching op_cache.

Findings

1. Survivorship (H1) is real, mild, and now quantified. The rated 37 are the ≥30-source-px slice of the 46 unwitnessed far-field silent misses, and the floor is not one floor — it is 7.5 model px on the 16384-px splits and 30 on morgantown, so the deck's composition follows stored width (27 GSV / 10 Mapillary; gainesville alone contributes 14). Net bias: AUC(rated px vs unrated far-silent px) = 0.600, rated median at the 62nd percentile. The 9 below-floor exclusions are the extreme tail (median 117 m, 3 above-horizon clamps). So the 94% visible rate generalizes to the far-silent core (~18–50 m) and says nothing about the deep tail.

2. The model's own hits refute a hard pixel floor. The decisive readout needs no reviewer:

band	GT	recall	silent	rated visible
18–25 m	395	0.777	25	12
25–40 m	226	0.549	32	13
40–150 m	72	0.292	22	9

For each visible miss, the model's recall over all far-field GT within ±20% of that miss's apparent size is median 0.57 (q1 0.31, q3 0.74). A hard pixel floor would put these near zero. Far-field failure is graded sensitivity, not a cliff — a silent far miss is the unlucky tail of a process that succeeds on most same-sized ramps.

3. Guards hold. Zero above-horizon clamps among the rated 37; the deck is majority GSV, where flat-ground distance is trustworthy (Spearman 0.95).

4. Phase 1: silent was peak bookkeeping, not blindness. Decomposing the in-window heatmap mass for all 128 pooled silent misses: absent (act < 0.01) 10 — the only true "saw nothing" cases; faint local (0.01–0.05) 39 — a real on-site response below the floor (mass at the ramp in 30/39, nothing else within ~3.8 R), the sub_threshold continuum; tail (≥ 0.05) 79 — a confident adjacent mode's tail (nearest cached floor peak ~1.4 R away at median score 0.685; argmax in the window's outer quarter in 75/79), the σ/representation family whichever of neighbour-TP / FP / self-just-outside-radius it is. The anomaly population itself — the 34 far-field rated-visible — splits 3 / 12 / 19: the model responds at or next to ~91% of the far ramps a human called resolvable. And of the 45 near-field silent misses under the 0.013 sourcing estimate, only 6 are heatmap-absent — the rest are the calibration and σ families §0b already prices, so 0.013 stays unrevised in either direction until Phases 2–3.

The takeaway (now stated explicitly at the end of §0c): distance is a stressor on failure families the taxonomy already prices, not a new category of failure — 62% σ/representation, 30% sub-threshold attenuation, 8% genuine absence. The implied lever is decoder- and representation-side (target σ, peak spacing, threshold calibration, Phase 2's scale question for the residual), not far-field training vocabulary; multi-view remains the one remedy that sidesteps all three mechanisms at once.

What this changes in the docs

• New §0c in docs/curb_ramp_data_sourcing.md with the tables above and the boundaries of the claim: §0a's far/near measurement stands, its "fixable by" binary falls (the 18 m boundary marks difficulty, not reachability); the §0b sourcing bracket's exclusion of all 83 far-field silent misses is no longer safe, but 0.013 is deliberately not revised until Phases 2–3 run (scale counterfactual, decoy control); multi-view's remedy logic is untouched, though the "MV ceiling" column shares the binary assumption.
• §0a gets an in-place qualifier pointing to §0c.
• Two stale §0b claims fixed: the reviewer pass is done (one rater), and the point estimate exists (0.013, docs/replication.md).
• Phase 3 design consequence recorded: the decoy deck should stratify by distance band and oversample 40–150 m, where visible went 9-of-9 down to 10.5 model px — the spot where pointed-verification bias would show most.

Mechanics

Phase 0: scripts/analysis/farfield_forensics.py + 21 tests (tests/test_farfield_forensics.py), result JSON at analysis_out/farfield_forensics.json. Phase 1: scripts/analysis/silent_activation.py + 14 tests (tests/test_silent_activation.py), result JSON at analysis_out/silent_activation.json (both whitelisted in .gitignore like the other result files). The integration tests pin the population arithmetic (83 = 37 rated + 9 below-floor + 37 witnessed) and cross-check the gallery manifest's widths against the imagery manifests. One instrument catch for the record: the first azimuth-null implementation could read a site's own faint bump back as "chance" — a unit test caught it before any numbers existed, and the null now self-excludes a 2 R zone around the site. Full suite: 533 passed.

🤖 Generated with Claude Code (claude-fable-5)

[jonfroehlich]

Scope grew from Phase 0 to Phases 0+1 in this branch — Phase 1 (activation forensics over all 128 silent misses, local GPU) landed as 59d5a9e + 64763d0. Headline: only 10/128 silent misses are heatmap-absent; 62% are a confident adjacent mode's tail (~1.4R, median score 0.685); the far-field rated-visible population splits 3 absent / 12 faint-local / 19 tail. Full numbers in the #46 comment and docs §0c. Suite 533 passed.

🤖 Generated with Claude Code (claude-fable-5)