Files

8.2 KiB

Gates: what each one measures, what it catches, what it is blind to

Every gate here was built because a specific failure slipped through the one before it. Run them in this order and report which gate each pick passed. Verify your verifier first: score the plate against itself (must be ~perfect) and against the plate shifted 3 px and scaled 1.01 (must fail) before you trust a threshold. scripts/smoke_test.py does exactly that on a synthetic town.

0. Plate gate (before any renderer call)

Overlay the map's checkable facts on the plate seen through the same camera. For a square: elongation, the closing range, the saw-tooth of gables, the ways in, the tracks, open ground clear of buildings. Any miss found here costs one re-run of massing. Found after re-surfacing it costs everything downstream, because the renderer photographs the plate's mistakes as faithfully as its truths. Also eyeball the untextured plate for: merged house blocks (split them), footprint slivers (they become ruins), a mannequin cut by the frame edge near the camera (it becomes a stone post), and built features that run off the frame with no end (the renderer extends them forever; model where a quay or wall STOPS).

1. Geometry drift (frame vs plate), gate.py

metric what measured bands
grad_corr Pearson correlation of gradient magnitude at 896x504 textured facade at 20 m: 0.5-0.7; town from 3 km: 0.4-0.55; smooth landscape: 0.15-0.3 (low by nature, judge layout by eye)
edge_rec share of plate edges that have a frame edge within 4 px reported
ecc_cc, scale, rot_deg, tx_px, ty_px affine ECC frame -> plate a faithful take: scale 0.995-1.01, shift under ~10 px; a reframed take: scale 0.92-0.98 or 1.04-1.10
landmark_grad_corr grad_corr inside the landmark solids' mask castle at 3 km kept: 0.4-0.7; redrawn as a palace: 0.16

Calibrate per view. The plate vs itself scores 1.0 / scale 1.0; +3 px and x1.01 must fail on every instrument.

Blind to: water becoming land (a surface change, not a shift), an invented building in a low-texture far band, a feature relocated to a face the camera cannot see, extra people. Those need the gates below and the eye.

2. Skyline and invented structures (wide views), gate.py

The sky/ground boundary per column, fitted plate -> frame by scale + shift (sky_s, sky_tx, sky_ty, sky_res in px). Then invented_structures: runs of columns where the frame's skyline rises more than --bump px (default 18) above the plate's, outside the dilated landmark columns. This is the gate for the one failure the block method still has: in a city-wide frame each house is a few pixels, the plate barely constrains the renderer, and it "improves" the skyline with halls, towers and churches. Reject any frame with a run; prefer a closer camera or stronger plate conditioning over more retakes.

Related measured failures on mountains: the renderer shrinks a massif to fit its summit in frame when the plate's wall runs out of the top edge (gate per column: sky in the frame where the plate has none, under 10 % of columns), and it paints snow on a low flank through a telephoto lens (fix: one edit-with-itself "replace every white patch with rock, scree, grass; change nothing else"). Color classifiers are haze-confounded beyond ~6 km.

3. Figure scale brackets, gate.py

Every proxy person the plate can see (from <V>_people.json, visibility from the oid buffer) is drawn on the frame as a bracket: foot line, true-height head line, and a +-25 % band (orange). Registered to the frame through the ECC warp when the warp is credible (ecc_cc >= 0.55, scale within 8 %). Pass = the dressed figure stands on its bracket with its head inside the band. qc/<id>_figzoom.jpg crops the 16 largest.

What the numbers looked like: street-level figures with proxies and the eye-line clause land at +5-13 % (heads on the horizon line); without the clause the nearest walkers came out +25-28 %; without proxies at all, aerial figures were 2.4 m at the far edge and 4-5 m at the near edge. Indoors with a face reference attached, close figures bias +8-22 % even with proxies. At ~20 px (aerial) scale holds but figures are not 1:1 with the proxies (a few move or drop).

Blind to: people the renderer ADDED that no proxy owns (a wide view added six tiny travelers at true size for where they stood; disclosed, not rejected). Count heads against the proxy count by eye.

4. Corner and edge sweep, gate.py

Four corners and the top/bottom bands of every frame, contrast-boosted (CLAHE + unsharp), on one sheet per position. Looking for: text, runes, numbers, watermarks, logos, captions, borders, a sun disc in frame. Read the sheet at 1:1.

5. Sign and emblem lock across angles (eye check)

Every sign is an emblem only, never lettering (a renderer "writes" pseudo-script on anything flat). Crop each sign at 4x from every camera and compare: same emblem, same board, same position on the building. A coin plaque with ring marks in one take and clean in another: pick the clean one and say so.

6. Occlusion check (eye, per camera)

For every named feature in the frame ask: which face is it on, and can this camera see that face? A balcony that is really on the far side but appears on the near face passed every element-by-element check and was still impossible. The fix is upstream: prompts per camera from a visibility list, with occluded features disowned as a positive instruction ("the balcony is on the far side and is NOT visible from here").

7. Storm intensity and snow depth, matrix cells

  • Storm cells: a snow-loaded roof hides the texture and lifts the eaves, so the full-frame gate fails by design. Ladder: full frame -> building band (ground to ridge) -> wall band (the same thresholds measured only on the walls, roofs masked out). Report which rung passed. Controls: keeper vs itself 0.0; a thin-snow take passes; a real whole-frame shift fails on every rung.
  • Night whiteout: brightness-normalized contrast metrics score a dark, soft frame as "heavy" when the eye does not. The instrument is a silhouette gate on the building band against the already-gated plate (sigma-2 flake smoothing), with the side-by-side sheet as the decisive check. A visible veil of lit, snow-filled air is the requirement, not low contrast.
  • Deep snow (medium/heavy): gate structure ABOVE the snow line (walls higher than depth + lee drift, eaves, ridges, parapets) and exempt the buried ground band. Measure depth and walkways against the site's Snow (light) cell of the same camera and light, not against the sunny plate (sun and shadow bias the snow line by 0.3-0.5 m). Count NEW dark lines only for walkways; the shared line network scores the same in every cell. Walkways on class C/D routes are a FAIL (references/matrices.md). At overview scale (1-2 px/m) depth reads but walkways and drifts do not; expect them only at street scale. Small field sheds turn into oversized people under a "tiny dark upright marks = townspeople" clause; eye-check open fields at 2x.
  • Seasonal edit of a winter keeper moves water: snowy low ground becomes a pond, frozen channels drain into grass. The geometry gate is blind to it. Review the water at 1-2x against the plate's water class outlined on the frame, or build summer keepers from a summer plate.

8. Same-tier recognizability (contact sheet)

Thumbnail every batch of exteriors on one wealth rung before calling it done. Seven map-true houses on one rung came out as seven identical timber barns; each passed its gate. Fix: a character pass, a one-change edit per pick that adds the house's tell at the frontage (one emblem sign, its trade visible: casks and trestles, boots on a rail, a smelter stack and slag heap, herb-drying racks), gated with ECC against its own source frame plus the corner sweep.

Reviewer sub-agents

Automatic gates catch shifts, zooms and scale. Independent reviewer passes, one per view, sweeping 1:1 quadrants, catch what they cannot: extra people, see-through ghosts, lost window columns, sky slivers, arch-shaped doors, a river that appeared in a view facing away from it. They take 15-50 minutes each. Brief them with the bracket rule and the density rule relationally (tight numeric ranges over-flag), and have them produce a fix list with boxes.