11 KiB
11 KiB
Life Agent SDF Notes
Reference: artifacts/ig_0b80263584b1a81e016a051475aeac8191acb0ced03b9a3e7e.png.
Supershape shell reference: D:/Downloads/supershape-2026-05-14-0437-09.json.
Working model:
- The readable form is a memory nautilus, not a generic seed.
- An ellipsoid silhouette with spiral material masks reads as wallpaper and fails the reference.
- A disc/cup welded onto the shell is worse: it enshrines the failed cut-face model.
- The source-backed model comes from
artifacts/Nautilus.glsl: solve the nearest logarithmic spiral boundary in cylindrical shell space, then derive wall thickness, septa, aperture/cut exposure, ribs, stripes, lip, pearls, cracks, and chamber light from that solved domain. - Do not replace the solve with a loose
sin(phase)whorl. That was the false shortcut. - Do not recreate the Shadertoy Nautilus. Its durable lesson is spatial-domain mapping, not the Life silhouette.
- For Life's C mantle, do not cut a closed Nautilus section circle with an angular wedge. A wedge intersection creates planar cut walls and umbrella fins. The C profile must be a distance to an arc in the solved section circle: inside the kept angular span, distance to the circle; outside it, distance to the nearest arc endpoint/contact region.
- Do not carry over the Nautilus aperture/opening stabilizer as an independent torus-like fallback in the Life C-arc topology. In the source shader it patches a closed-shell discontinuity; in Life it becomes a giant false shell that overwhelms the actual spiral branch solve.
- Life's C profile radius should not default to the full Nautilus section-circle radius. To make the mantle start near the ember equator and build on itself, derive the local C arc radius from previous-turn contact: with current whorl center radius R, previous-turn ratio k=exp(-BTAU), and endpoint angle halfSpan, solve R + acos(halfSpan) = kR + ka for profile radius a. Keep this as domain math, not a bridge pasted on after the fact.
- Correction to the Life C-contact math: exact endpoint contact with the previous turn's section circle is possible only if cos(endpointAngle)^2 >= 1-k^2, where k=exp(-B*TAU). If the kept C span is too narrow for the growth rate, gaps are mathematically unavoidable. The shader playground should surface this by deriving/clamping the feasible half-span and solving the real section-circle contact equation, not by adding a fake bridge.
- Further correction: the Life mantle section does not have to grow at the same exponential rate as the spiral centerline. Keep separate exponents for centerline growth (
B) and C-profile growth (PROFILE_B). The previous-turn profile ratio iskp=exp(-PROFILE_B*TAU), which belongs in the endpoint contact solve; assumingkp==koverconstrains the form and can force persistent gaps. - Bug in the first separate-profile attempt: adding
PROFILE_Bto the contact equation is meaningless unless the actual C profile radius field uses it. Define a profile radius field such asPROFILE_END_RADIUS * exp(PROFILE_B*u), solve current endpoint contact against the previous turn's actual profile radius, and consider adjacent branch candidates beyond floor/ceil because the nearest closed-circle branch may not be the nearest C-arc branch. - Better Life profile parameterization: do not define profile radius as
PROFILE_END_RADIUS * exp(PROFILE_B*u), because setting the terminal amplitude to zero makes the whole field zero andPROFILE_Binert. Use a seed/ember radius plus age-based growth, e.g.PROFILE_START_RADIUS * exp(PROFILE_B * (u + TURNS*TAU)), with terminal flare as an additive option. Contact solve should fall back to the base profile when exact contact is infeasible rather than widening into a pillbug. - Contact must be solved surface-to-surface, not centerline-to-centerline. If the current C endpoint is placed on the previous turn's profile centerline, the wall radii interpenetrate and any seam placed there will expose the lie. Solve endpoint distance against
previousProfile + previousWall + currentWall, refine current wall thickness once, and keep the endpoint geometry in the same angular frame as the solve. Terminal flare should be released from contact pull soPROFILE_END_RADIUScan actually widen the mouth instead of being cancelled by the previous-turn constraint. - Further Life C-contact correction: never force contact by widening the C span toward a nearly closed circle. That only turns the section into an overlapping pillbug. The previous turn in the contact equation must be the actual solved previous profile radius, not
profileBaseRadius(u - TAU). Evaluate the finite same-phase growth lineage and solve each turn against the solved turn before it. - Missing Life C-contact degree of freedom: a fixed C arc centered on the spiral whorl cannot generally touch the previous turn for the desired half-open span.
PROFILE_Bcannot fix an impossible triangle; below the spiral growth it stalls, above it grows but gaps. The coherent grammar is radius for fullness plus a solved section-space profile-center offset for contact. Solve|vec2(D + oCur - oPrev,0) + radius*endpoint| = previousRadius + previousWall + currentWall; do not fake contact by changing span. - Life C-profile feasibility: once contact is solved by x-offset, radius is constrained by the endpoint's perpendicular component. If
radius * abs(sin(endpointAngle)) > previousRadius + previousWall + currentWall, no x-offset can create contact. Clamp or derive profile radius inside that envelope before solving offset; otherwise highPROFILE_Bexplodes and low settings appear inert. Do not use a terminal release blend, because it breaks the seam continuity the end radius is supposed to preserve. - Life C-profile recurrence must not anchor at the first visible loop. A variable-generation recurrence (
floor((u + TURNS*TAU)/TAU)) creates a branch cut every revolution: the innermost loop grows naturally until it hits the clamped origin, then each later loop inherits the same discontinuity. Use fixed prehistory for everyuand walk forward the same number of turns so there is no special visible origin. - Fixed prehistory is also wrong for Life C-profile contact: it removes the branch cut by converging every visible turn toward the same mature solved profile, producing a tire/pastry. The better model is closed-form self-similarity: express profile fullness relative to spiral radius and solve offset directly with
k=exp(-B*TAU), e.g.|vec2((1-k)*R*(1+beta),0)+radius*endpoint| = previousRadius + previousWall + currentWall, thenoffset=beta*R. No recurrence, no clamped origin, no inherited mature profile. - Retreat from the overfit contact solves: the readable fixed-center C profile was the right visual base even when it intersected. The coherent non-overlap rule is accretion ownership: newer shell may meet older shell but may not occupy its volume. Restore the readable fixed-center contact radius and trim each branch by the previous turn's signed volume (
current minus previous) so the seam is an actual boolean boundary instead of another fragile radius/offset compensator. - The readable overlapping model specifically included contact pull: widen the C span to the minimum endpoint-contact-feasible span, then pull the fixed-center profile radius toward the previous-turn contact solve. If that is omitted, parameter changes can produce a non-overlapping gap model instead of the useful raw overlap that the accretion trim is supposed to cut.
- The spec authority is a protective revolved seed/cardioid shell around a visible ember. The Nautilus spiral math should only coordinate seam, bead trail, aperture logic, and surface breathing.
- Lip, pearls, ribs, cracks, and chamber light are fields on the same shell/spiral coordinates, not separate child objects.
- Concept target: front view should show a heavy teal shell mass with a round outer whorl, a gold inner aperture bowl on the left, a visible ember near the inner spiral, pale raised ribs arcing from the center to the outer shell, and memory beads along the spiral lip. Oblique views may reveal thickness, but must not expose detached ornaments or a hollow ring.
- Working checkpoint as of
20260514-163152: build Life in ordered layers. First solve the log spiral curve until it reads alone: ember at the core, about two turns, and a wide lip termination on the mouth side. Then inflate that curve into shell mass, carve the aperture from the same mouth field, reveal the ember, and only then add ribs/pearls as masks on the solved spiral domain. - Cut line: if a future pass tries to tune ornament before the curve reads, stop. If the shell becomes an egg with spiral decals, delete it. If the spiral alone becomes a hollow C-shell, add support from the same domain instead of welding on separate organs.
- Aquarium placement correction: the Life spiral plane belongs in world
YZ, with shell profile thickness along worldX. In the preview set,life-rightis the lip-sideY+camera, so uselife-leftas the reference comparison view when matching the concept art's visible spiral face. Do not fake the mouth by trimming the shell against the previous turn's volume; that cuts along the spiral plane, which is not where the biological aperture lives. - Life material/detail correction: pearl ribs must be raised circular-section distance fields on the shell surface, not painted color masks. Inner shell material should be nacre/pearl with warped Voronoi gold fractures; the same fracture field can tint the teal enamel outside. Failed hypothesis: closing the C span near the terminal turn creates a false lip-side blade and should not be kept.
- Life containment correction: removing the terminal
ubound before a correct C-span taper exists lets the final branch run into giant nacre fins. Keep the terminal bound as a temporary containment invariant until the mouth profile is solved by the C-domain itself. - Rollback checkpoint: after the 2026-05-15 failed mouth/profile detour, the source shader was restored to the live-reload shader from
artifacts/dev-reload/slots/20260515-145843-490c3cac/Render/Shaders/D3D12LifeAgent.hlsl. Treat that as the current coherent baseline before attempting more Life mouth work.
Latest active checkpoint:
artifacts/agent-previews/Life/20260514-193811/life-front.pngartifacts/agent-previews/Life/20260514-193811/life-mouth-oblique.pngartifacts/agent-previews/Life/20260514-193811/life-three-quarter.png
Current judgment:
- The active source is the last accepted pre-detour C-sweep shader, not the later scaled-inverse/profile-rewrite experiments.
- Do not reintroduce the rounded terminal-overrun cap that replaced
d = max(d, u - LIFE_TERMINAL_U). It was the only diff between the restored source and the actual last known pre-nightmare build, and it did not earn its keep. - The current form is not final. The next productive move is a narrow, reference-checked mouth/lip pass from this baseline, with the actual
Life_C_Sweep_ShaderToy.glslplayground in sight.