Light and optics

Rules

1. The era's screen is the key light, and the brightest thing in the cornea. Its colour reaches the eye only as light: reflection, spill on the skin, catchlight, sheen.

2. The reflection is the era's real home page, captured from the codebase (npm run eye:screens writes public/eye/screens/<era>.webp), never synthetic content when a real page exists. Re-run it whenever a home page changes, then recapture the posters.

3. Mirror by ray tracing. The reflected ray from the 7.8 mm cornea cap is traced to the screen's actual rectangle (centre, normal, right, up, size), weighted by Fresnel (f0 0.035). No fake geometry, no environment map.

4. Each era's page is sized and framed for its key interface, within two checks:

   • the era's key interface reads through its own medium, on desktop and phone;
   • it still reads as an eye: a dark ring of iris all round the page, the pupil showing through, the page never filling the iris edge to edge.

   The size is the screen's stylised size ×0.62 × reflect.scale. reflect.frame holds one region of the page still, cut at the device's own aspect; the whole page is otherwise drawn into the 16:9 screen and stretched onto the device (a portrait phone squeezed it to a third). Per era: ERA_LIGHT[era].reflect in look.ts. Tune it with the lab's "Reflection" section, or scripts/eye-realism-baseline.mjs reflect for side-by-side sheets.

   • Why: 60 % of the iris was a design rule, never a realism one. A real cornea mirrors a 14" screen at 55 cm as ~2 × 1.5 mm, about 3 % of the iris, so the size serves the concept.
   • Its floor is the original reason for it: a whole-iris reflection read as a lens, not an eye.

5. A white page is compressed, not stopped down twice. Reinhard and contrast keep its layout: "what survives a highlight is contrast, not light". The reflection is sharp at the dome's centre and softer toward the limbus (mip bias grows with distance, tear slope, dryness).

6. Mip bias follows the shot's scale: U.mip = max(0, log2(span/1.09))·1.15. Why: a dithered terminal otherwise turns into a snowstorm of specks on a phone.

7. Screens are larger than life and placed where that era's device sat (CRT centred and low at night, phone above at arm's length, a window for 2026). Fill lights are physical: overhead fluorescents (1981), a desk lamp disc (1991), a tabletop bounce (2013).

8. The tear film lives. A blink lays it fresh; it creeps up for ~1 s; it thins (~2%/s, ~10%/s over weak spots) and breaks up after about 8 s of staring; dry spots blur and dim the reflection. It never wobbles like water.

9. The meniscus is a thread, brighter on the lower lid than the upper, wider at the tear lake (R12): lower ~0.36 mm, upper ~0.2 (0.3 at rest × 1.2 / × 0.65), tapering out at the outer corner; its normal from the bead's own curve; brighter where the screen faces into it (the lower bead with the screen below); still glints along it, never moving.

10. The page is the light, not the reference photo (user, 2026-10-28). Every surface the screen lights takes the era's page as its key: the average colour and brightness, plus the blurred picture in front of it, mirrored (throwUv = (0.5 − x/72, 0.5 + y/44), mip 8, tint picture / avgLum clamped 0.45–2.2: skin ×0.22, white ×0.25 where it faces the screen, iris ×0.3). The iris's crescent is the page focused through the cornea, inverted (mip 6, ×0.3–2.6). Why: the user's photos were lit by a phone or a lamp; the loader's eye is lit by each era's interface, so light is never tuned to a photo.

11. Reflex tears: 4+ era requests inside 1 s add tearing (+0.25, decaying 0.15/s), counted on the request, not the cut. Why: rate-limited cuts never reach four.

The era lens: one recording style per era

• Each era gets one cue at most, with a kill switch (lens.on). Every lens number damps at rate 3; an artefact fades out before the next switches in.
• Grain steps at 12 Hz (never faster than 18), never per frame ("a 60 Hz shimmer tires the eye"). It's ×0.45 on phones, and 0 for eras whose medium quantises (the dither is the grain).
• Hashes are Hoskins hash12, never sin-based. Why: sin hashes show stripes at pixel coordinates in the thousands.
• Artefacts that failed at full frame: CCD smear ghosts (2000, removed) and flare rings on the brow (2009, kept at 0.45). Test every artefact at full frame, not cropped.
• Vignette: only the excess over 0.3 counts, and an era whose medium is full-bleed colour sets it to 0 when it reads as a frame (2013).

Post chain order (post/pipeline.ts)

1. Scene pass with MRT: output, emissive (8-bit), viewZ.
2. Macro depth of field: half-res, 9-tap separable gather. A tap counts only if its own blur reaches the pixel, so a soft face never smears over the sharp iris; things a little in front of the plane (lashes, margins) get a fine full-res 4-tap cross under ~2.5 px instead (R14). Not on T1.
3. Bloom on emissive only (0.25 strength, quarter res). WebGPU only: three's MRT merge fails on WebGL2.
4. Era lens, linear stage (weave, webcam, interlace, lateral colour, unsharp, halation, comet, CCD, flare, white balance, split tone, saturation with isolate).
5. Exposure and the room's falloff (a dark room round the face).
6. Filmic ACES fit, the only tonemap (the canvas runs flat); then FXAA on T2/T3 (R14), before any grain.
7. Era lens, display stage in γ2.2 (gain, lift, gamma, grain).
8. The era medium, reading the finished frame as an 8-bit display-encoded mipmapped texture with depth in alpha (eye-era-media).

Focus

• Focus stays on the iris plane (11.6 mm). It racks slowly (rate 3) to lashes (15.5 mm) or skin under the pointer.
• No focus breathing: the frame never scales as focus racks. Why: it read as the face zooming.

Posters

• 20 AVIF stills (<era>-\{wide,tall\}.avif: 1600×900, and 780×1688 for phones at 2×).
• Placed by the eye's own framing (eyeShot), not object-fit, so the crossfade to the live eye does not jump.
• Removed 650 ms after the live eye is up, so the crossfade never dips through the room.
• Capture: an offscreen stage at the exact CSS size, reduced motion, non-interactive, media on. Lab "Capture all eras", or window.__eyeCapture(eras?, variants?). A dev-only route writes AVIF q52.
• Recapture after any change to an era's look, its medium or its screen capture, and only once the user has signed the look off.
• The site's share card is cut from the wide posters (ten vertical era bands of one eye, scripts/og-card.tsx → public/og/share.jpg). Run the eye-share-card loop after every recapture, or link previews keep the old eye.

Realism pass (R1 cornea, R11, R12, R14): making an era's light real

Phase specs and the per-era light table: references/realism-light.md. Shared cards: ../eye-loader-system/references/era-realism.md.

Making an era's light more real? Walk it in order:
1. Is its ERA_LIGHT built? (all ten since batch 2) → for a new era, land it from ERA_POV_PLAN.md first
2. Direction → shadows project along the key:
     from above → down across the top of the iris, under the brow
     from below → onto the lid above the lashes, above the lower lid; the hood's underside lit
3. Size → penumbra = gap × realMm.h / distMm, never the stylised size
4. Spectrum → should red tissue darken (green or blue light)? → raise `spectral` as data and A/B it;
     1970 only with sign-off
5. Fills → does the card name a catchlight? yes → reflect it at ≤ 35 % of the page; no → light only
6. The page in the cornea → its size and framing are the era's `reflect` (rule 4), chosen by eye;
     optics only bend and blur it

1. Never a fixed shadow band. Every lid, lash and crease shadow follows the era's key (C3). Why: today's band under the upper lid is wrong for the light from below in 1970, 2000 and 2009.
2. Near reflections are always on and always dark: the lids and a lash comb in every mirror ray. They are the cheapest realism in the deck.
3. FXAA and any idle refine run before the lens grain and the medium. Why: grain and dither must stay fixed in screen space.

All per-era light positions, lens values and optical constants are in references/optics-numbers.md.