Scene Loop Trial

BAKE9 refitted the lights on the real paint and, before baking, tested the predicted picture against the gates. The left wall passed; the front failed (warmth off by 0.087 against ±0.05, pools 25% too strong), so it stopped without spending a bake. Its known-answer control found that the light colour had been applied twice in every rebuild, squaring it. With that fixed the steel is neutral, but the amber pools went too: one colour per light group cannot be both a neutral wash and a warm pool. BAKE10 splits the lamps by role.

MATTE5 fixed the back-wall seam in the engine (a sampling fault, not the painting). Its sharper repaint failed on the AI tool's size limit, which the skill had not asked anyone to measure up front.

The new gate saved a bake; a control found a bug three rounds had carried; my earlier explanation (fitting on grey) was only part of the cause.

BAKE7 gave every deck enough map pixels to hold a pool of light (20+ per pool, up from 0.2 per grating triangle), but it baked with the wrong light rigs, so its scores could not count. BAKE8 re-baked both options with the right rigs, checked by hash, and with the speckle removed. Both still failed: similarity 0.55 and 0.54 against a floor of 0.63, and the steel too warm by 0.12–0.21.

The skill then asks for a cause, not a third try. BAKE8's control render measured the steel in the ray-traced picture itself, before the engine touches it: it is already 0.28 warm against the target's 0.09–0.17. So no amount of work on the engine side can fix it. Two candidates remain, and DISPLAY1 is measuring both now: the colour curve (Blender's and the engine's tone curves differ by up to 0.13 on a warm grey card), and the steel's paint and grime colour (the targets show clean grey steel).

Stopped a dead-end method after two rounds and narrowed the cause to the scene's lights, paint or colour curve, not the engine.

Before picking a new method, the skill asks for a measurement rather than a guess. Through one shared tone curve, the two middle pictures are almost identical: the curve closes only 6–12% of the warmth gap. Splitting the steel's light by group on the real paint shows where the rest comes from. The front work lights alone give 66–74% of it, and the colour BAKE4 solved for them is a deep orange that washes over every member, where the target keeps the steel grey and the warmth inside each pool.

Named the cause (the fitted light colours, chosen while the steel was grey) and priced three ways forward, so the method choice is yours with numbers instead of another round.

Before the skill, a round's verdict was "still not real". With the sheet, BAKE2 was measured against the concept: brightness 0.030 → 0.066 (concept 0.083), warmth +0.10 → +0.26 (+0.28), crushed black 17% → 6% (3%). The squint column showed the defect the numbers hid: the concept has separate bright pools, ours an even glow, and the colour column shows the warmth sitting in the steel, not in the lamps.

Turned one vague verdict into three numbers and one named defect, which became BAKE3's brief.

The skill's rule is to judge lighting on flat grey materials, one light group at a time. That showed the sky gives 0.6% of the tower's light and the warm work lights 45.7%: 171 lamps with 150° cones and no falloff, whose pools overlap until no dark gaps are left. Without this step the next round would have been another guess at colours.

Cause found in one round, with a number per light group.

The skill says AI does the painting, from our own renders, and only an image that passes the align check may be used as an exact target. Both options passed (edge recall 0.913, shift at most 0.16%, scale 1.00), so the next round could fit to them pixel by pixel instead of comparing against a concept with a different layout.

An exact target for lighting in one step, and a picture-based choice for you.

BAKE3 hand-tuned two light rigs; both failed, so it stopped, as the skill requires. BAKE4 changed method: render each of 29 light groups alone, then solve for the brightness and colour of every group that best reproduces the chosen option. Checked in a window it was not fitted on, the error fell from 1.96 to 0.54.

The rule prevented a third and fourth round of guessing, and the new method matched the target on grey.

Because lighting had been matched on grey, the drop to 0.016 with real materials could only come from the materials: an effective reflectance of about 0.09, where painted grey steel is about 0.45–0.55. BAKE5 traced it to the paint colours and weathering, fixed the paint (0.146 → 0.495, non-metallic) and brought the brightness onto target.

Still wrong and now being fixed (BAKE6): the pools that the ray-traced render has are lost in the baked version (similarity 0.51 against 0.63 before), the lamp heads do not glow, and brown grime streaks remain.

Found the real blocker. But only after two lighting rounds: see the gaps below.

An automatic score had judged the first painted backdrop worse than the bare pad; by eye it was clearly better. The skill forbids using a score that disagrees with human judgement. The score was rebuilt and tested on 14 pairs I had judged and on synthetic damage (blur, darkening, shifts) before use: 14/14 and every stress test passed.

Prevented a broken number from stopping good work.

The skill only lets an AI image be projected if it keeps the render's layout. In MATTE4, 25 of the 54 AI tiles had shifted or rescaled (horizons, coastlines, buildings), so they were held back and those regions fell back to the plain render. That kept wrong geometry off the walls. It also showed the method's ceiling: the remaining patchwork of painted and unpainted tiles reads as blocks and smears, clearly worse than round 1's single painting.

The check worked as designed; the method did not. Rounds 2–4 never beat round 1, and the skill had no rule to go back (see below).