| Photon-primitive roster | Analytic caustics by sweeping a light path | per-row median 1.000 | measured |
| Three-plane mixture | Removes the grazing singularity | worst pixel 259 → 1 | measured |
| Direction-sweep trio | Sphere, disk and cone mixed | worst pixel 6291 → 1 | measured |
| Cross-family mixture | The three planes and the direction trio under one light-area measure, six closed-form densities | lowest per-pass variance at every emitter size; energy agrees with 12 and 13 to 0.1 % | measured |
| Light-size series | One glass ball, five emitter sizes from 2° to 77°, every estimator timed | both written crossovers falsified inside the range | measured |
| Surface caustic photon map | The photons the beam walk already traces, deposited at the first diffuse hit behind glass and gathered by the camera walk; the floor under every glass ball lights up | energy within 2 % of a first-principles trace on a receiver 62 % hidden by the ball; 21 pinned films byte-identical with it off | measured |
| Direction-preserving surfaces | Not only the estimator's probes: every integrator in the mode menu now sees through a passthrough wall, so the camera stands outside the fog box on all of them | preset 63 renders the Reliquary from outside on modes 2 through 8 — means 0.1988 and 3.47 to 8.89, against exactly 0.000000 through the same wall left opaque on every one of them whose control finished; 21 pinned films byte-identical | measured |
| Continuous-orientation planes | Closed-form envelope over the sweep angle | median 0.999 | measured |
| Photon hourglass | Caustic through a dielectric sphere | median 0.976 | measured |
| Hourglass and beam mixed | Spectral, relay-aware pairing | 1.0000 against the beam at the same radius; 2.1× lower error; consistent, not unbiased — its energy reads +0.9 % at r = 0.15 against its own r = 0.0375 rung | measured |
| Vertex hourglass, entry-seeded | Seeds on the glass, not the light | energy 1.007 / 1.020 | measured |
| Exit refraction cone | Two cones apex-to-apex at the vertex | 1.0001 against the hourglass | measured |
| Reflection cone | First analytic reflection caustic here | 0.998 against the beam | measured |
| Stochastic photon beam | The kernel reference the rest is judged against | 1.048 against next-event | measured |
| Surface next-event reference | One Lambertian connection at a hit | 1.0005 against the GPU | measured |
| Variance study | Why each estimator fails, and where | 15 rules, each cited | measured |
| Byte-identical regression | A feature off must change nothing | 8 films, two resolutions | measured |
| Behaviour triage document | Symptom, expected, bug — with corrections | 6 corrections in one day | measured |
| The in-engine gate stack | Four logs behind one config key: an analytic airlight gate, an exact-Le camera gate, a relL2 against a GPU film, and two brute-force oracles at five named probe pixels that cross-check each other at one bounce | Gate C caught the 1/cos θL over-brightness on preset 34 that the analytic gate structurally could not see — the source says so at the line it fixed | measured |
| Mirror relays | Caustics that arrive by reflection | +0.16% of frame energy | measured |
| Photon ring-alt | The family's first glossy vertex: the reflection cone taken off the delta, five free dimensions instead of three, the BRDF evaluated at every crossing | 1.0005 ± 0.0088 against the camera walk isolated to the same class at roughness 0.16; 2.2–3.1× lower variance per pass and 6.6–8.3× on equal time, with no crossover in range | measured |
| The one-dimensional group | The user’s photon ring and photon arc, and their mixture with the tracer beam. Fix one of (θ,φ,t) and sweep two and you get the kernel-free cone, disk and sphere; fix two and sweep one and you get these — the lattice closes, and the beam’s familiar 1/sin θ turns out to be the same expression at unit speed | worst pixel 545.7 → 5.0, lower than the beam’s own, and 2.7× the best single member per pass at r = 0.15, 2.5× at r = 0.05, at three draws per pass against that member's one; bias ladder over five kernel radii, every rung with its own bar; consistent, not unbiased — 0.33 % high at r = 0.30 where the ring alone is 33 % high | measured |
| The float32 cliff it found | A lobe-sampled cosine pins one coefficient to 1 and the cone quadric's leading term cancels in single precision | 81.5× / 108.8× / 8.7× invented energy at roughness 0.05 / 0.09 / 0.16 before the shared solver was moved to double | measured |
| The refraction chain | The swept sheet pushed through a chain of glass balls instead of one — exact where the chain is collinear, and refused by name where it is not | 60.4×/116× lower variance per pass than the stochastic beam and 35.5×/48.5× on equal time; energy on the caustic mask 1.0023 ± 0.0061; the off-axis pair renders exactly zero with one refusal line | measured |
| Post-ball planar mirror | Light → ball → fog → mirror → eye: a reflection is an isometry, so the camera ray is reflected and the same quadric is solved against it | 0.9979 ± 0.0017 against the same scene shot from the eye's mirror image; 14.5–28.3× the beam per pass on that view | measured |
| BSDF-general surface shading | The estimator's camera walk shades every non-delta surface, not only Lambertian ones — Metal, Anisotropic, CoatedDiffuse and the Airy film, with the delta rule taken from the megakernel's own source | Metal 1.0133 / 1.0096 / 1.0111 against the GPU at roughness 0.10 / 0.15 / 0.30, inside the band two Lambertian controls set at 1.0008 and 1.0141 in the same frame; reciprocity and white-furnace tables computed by the rendering binary itself; 21 pinned films identical with it off and on | measured |
| The gate family | One lamp, one gobo plane, four apertures: a slot, an annulus, a grille and a dispersive slit | the annulus matches the slot's aperture area to 1.4 % (2.98 cone² against 2.94) and still drives the shaft's core-to-limb ratio from 0.94 down to 0.40; the open variant's film mean falls 5.9× while its shaft-to-surround contrast rises 4.8× → 10.1× | measured |
| GPU megakernel tracer | One dispatch for the whole transport, in real time | seen; fallback ladder in source | seen |
| Real-time bidirectional path tracing | BDPT inside the megakernel — eye and light subpaths, interior connections, all six (s,t) cells, MIS method B — with a clickable strategy pyramid | seen; each cell isolable from the pyramid overlay | seen |
| ReSTIR direct lighting | Resampled importance sampling into per-pixel reservoirs, temporal reuse against the previous frame's reservoirs, then spatial reuse from neighbours — the full construction, not a filter, replacing the tracer's direct term | 907 lines of GLSL across two compute passes; seen | seen |
| Interactive Metropolis | Kelemen primary-sample-space MLT with lazy per-coordinate mutation, in the CPU integrator bank | seen; interactive, not real-time | seen |
| Self-recompiling kernel | Specialises to the live render state | seen; measured instruction ceiling | seen |
| Material models | 14 types, Lambertian through hair, subsurface and dispersive glass | all 14 reachable from a preset; 56 covers the last four | seen |
| Spectral dispersion | Per-object wavelength-dependent index | seen in the prism preset | seen |
| Participating media | Homogeneous and noise-modulated | seen; inert in the CPU integrator | seen |
| Solid-angle sphere sampling | No wasted samples near the silhouette | seen; variance-only comparison | seen |
| The prism is two solids | Preset 30's dispersive prism is a 45° box to the path tracer and a 45° mesh to the rasteriser — and the tracer's half-extents were pre-fattened to the rotated bounding box before the GPU learned to rotate the ray, so the fattening is applied on top of the rotation it was compensating for | 0.5 → 0.7071 in X and Z: √2 wider, twice the volume. Read from source, not measured | seen |
| Node-graph material bake | Ten channels of a graph baked per entity into three RGBA8 layers the megakernel samples — and a twelfth slot booked twice: the dispersive gate reads the same alpha as the dispersion spread it gates | the layer comment says anisotropy angle, the write stores dispersionSpread / 10, and three of the eleven seeded BSDF inputs have no mask bit at all | seen |
| VCM “Fix Bias” | The checkbox reaches VCMConfig::fixBias and the field is read nowhere in the integrator — its own declaration says “always enabled; kept for UI compat”. The greyed MIS-mode selector beside it names this control as the live one | zero references to fixBias in VCMIntegrator.cpp; the catalogue grades the row seen and describes it as toggling the weights | seen |
| Time-of-flight, three systems | Modulation carried along the traced path, four-phase continuous wave off the depth buffer, and a geometric-path post-process with five light shapes | implemented, unphotographed; the path-length gate is a fourth, separate thing | never viewed |
| Off-axis holography | With a live error readout | implemented, unphotographed | never viewed |
| Structured light | Pattern projection; what gets decoded is phase, never pattern index | implemented, unphotographed; no pattern decoder exists | never viewed |
| Two emission profiles | Cone and ring emitters | no preset uses them | never viewed |
| Photon Primitive Validator | Pick a camera pixel and a photon on a light, budget the two legs against one total, sweep any vertex, and ask a zero pairing why it is zero | eleven rows in the inventory and not one of them has been on a screen | never viewed |
| Time-of-flight noise models | An ambient term and a “shot noise” term on both synthetic time-of-flight ladders | read from source: ambient is added to all four shots and cancels exactly in both differences the recovery is built from, and the noise term is uniform, signal-independent and frame-frozen. One capture would settle it, and the imaging page carries the recipe | never viewed |
| Bilateral-grid preview | A live-parametrised diagram of the bilateral grid drawn from the filter's own sliders, with the support ellipse at the cursor | read from source: it splats and draws; the blur and the slice its own footer names are not performed | never viewed |
| Complement mode | A primary estimator plus a backup for the transport it does not own, partitioned by path class and never mixed | built; the backup walk reads 1.0008 on the diffuse half and 0.9949 on the glass against the GPU, preset 0 | measured |
| Correlated primitives | One batch of primitives, every camera ray | 12 films byte-identical off; converges to the uncorrelated energy | measured |
| Gaussian-process surfaces | Microfacet and medium from one object | both blocking gates passed | design only |
| Camera-side primitives | Sweep the camera path instead of the light, and splat the result onto the film | PP strategy 17; 0.992 against the camera walk on preset 48, both legs shadow-tested since 2026-09-05 | measured |
| Node-based material and medium authoring | 57 node types, live 2-D and orbiting 3-D previews, evaluated per entity by the GPU tracer. Its medium half is a different machine from its surface half, and neither reaches any CPU integrator | fourteen of the nineteen node-editor rows in the catalogue are graded never viewed, the canvas among them; the five that are not describe what the graph reaches rather than the editor being looked at. Reachability is not viewing, and this row said otherwise until 2026-09-06 | never viewed |
| Structured-light projector | Four delta geometries with four falloff laws, seven patterns, a beam angle derived from a quad you size; a real next-event light, and on two of the four shapes a real emission disk, so it appears in mirrors | implemented, unphotographed — this row read seen until 2026-09-06 on the strength of the ToF work, which uses the projector as a carrier and never as a lamp | never viewed |
| The projector's top-down setup view | A draggable plan of the whole rig — scene, camera frustum, image and focal planes, the effective quad, the emitter, and the time-of-flight wavefront as an arc for a point source and a straight line for a panel | implemented, unphotographed; the difference it draws is the one the imaging page currently asserts in prose | never viewed |
| Photograph-to-light | A captured pose paints a photo into the tracer’s albedo array, where it then receives correct global illumination | implemented, unphotographed | never viewed |
| 157 stackable image operators | Over 21 non-photorealistic modes — Ricci flow and persistent homology share a combo with halftone | grouped, mostly unphotographed | seen |
| Integrator comparison report | Renders each configuration, writes PNGs and a browsable HTML report, round-trips the run as JSON | the only working image export in the program — and nobody has generated a report and looked at it. This row read measured until 2026-09-06; the catalogue grades it never viewed and the catalogue is right, because writing a file is not a measurement | never viewed |
| Unit test suite | 61 assertions over the shape classifier and the evaluator; exit code 1 on failure | already a CI gate | measured |
| Comp-mode class counters | Hunt double counts and holes that the mean, the median and a seam shift all miss | wired to the guard that can actually fire | measured |
| Three holographic reconstructions | Off-axis synthesis, GPU FFT sideband crop, and a WLS-PCG Poisson phase unwrapper | self-scored in wraps against ground truth | never viewed |
| Scene-setup advisor | Ten rows naming each way your interference fringes fail — nine two-state, one warn-only | nothing else here diagnoses its own setup | never viewed |
| Non-standard-state alarm | The renderer notices it is in a diagnostic configuration and offers a one-click way out | seen | seen |
| Interactive ray tree | Walk a single path’s branching structure in the viewport | implemented, unphotographed | never viewed |
| Cone sphere and volume emitter sampling | Solid-angle sampling with its legacy alternative kept beside it for comparison | a variance-only A/B | seen |
| Five physics solvers | Coupled to a path-traced medium — though only the Euler solver feeds the grid | seen | seen |
| Classical vision bank | GPU Hough with compute non-maximum suppression, epipolar geometry, stereo, four-mode reconstruction | implemented, unphotographed | never viewed |
| A finite aperture as a transport object | Opening the pinhole adds a light-tracer connection strategy to the bidirectional MIS count, so a camera control changes which paths exist | the strategy counter reads the aperture radius directly; the panel warns in red when it is zero | seen |
| Shaped-aperture bokeh | Circle, polygon, star, cross and ellipse, rejection-sampled in the shader, with blades, tilt, inner radius, arm width and aspect | seen; GPU only — every CPU integrator rebuilds a bare pinhole | seen |
| Two aperture implementations | A complete C++ lens model beside the megakernel's own, with a different shape set | every call site passes a zero lens sample, so the C++ aperture sampler has never produced a pixel and its shape enum has no writer outside its own directory | unreachable |
| The projection self-check | The rasterised gather establishes that the camera it was handed really is the inverse of the projection it is about to use | 15 probes at 0.05 px; on failure it refuses and names the camera, and the tracing gather runs instead | measured |
| Goldstein branch-cut unwrapping | A second phase-unwrapping algorithm | the function body is empty and its selector has no writer | design only |
