spike: mirror R1's measured 7.651 ms derived-capacity frame (326a06b, 3ba0bf9)
R1: the interpreter's per-invocation stack priced as a DEPTH ladder on the 1070 (93% of a walk)
wiki: R1's grid x step frontier lands as a same-session baseline on the GTX 1070, and the step budget drops off the lever list (6ef57d0)
wiki: R1's named instrument ran at 1080p on the GTX 1070, and the arm resolves there at the march's own cap (4b68606, ff8fb26, ea1daa4)
The front page still said the demo had no GPU upload and no frame timing and that R1 waited on hardware; the
hardware has now reported. r1_march_alone's camera path on buildbox's GTX 1070: cap 1 15.17 ms (inside 16.67),
cap 8 70.76, cap 64 91.15 - 5.5x over budget on the instrument the gate names, with the read-back near
cap-independent and the old 551.5 ms 46-pass row shown to be measuring the harness. R1_ARM_CAP then prices the
bound-skip arm at the cap the march actually uses: armed 70.519 ms against the +inf control's 73.945, paired
median 3.420 ms (+4.6%, 112.0 sigma, 14/15), and the 768-ray null survives a third instrument at -0.001 ms. So
R1 moves from unmeasured to measured-and-unmet, the arm's share (3.0-6.0%) is the part that travels between GPUs
while its milliseconds are the machine's, and the missing pitch-demo piece is a window, not an upload path.
wiki: the 768-ray null survives its re-pricing - the bound-skip arm is grid-dependent, both ends measured (5757b11)
wiki: the bound-skip arm priced at 1080p - the 768-ray null was the statistic, and the paired term survives an order swap (48649f6)
docs: the device arm priced against an unlicensed-bound control - the saving is inside the noise floor on the MX250
docs: the tally pair is drift-guarded in docs/evidence.json, and the front page's stale 'no GPU' clause goes
docs: the device counts its own bound-skip refusals - 22,237 on the demo scene's 3,072 rays against the CPU's 42,240
docs: the bound table's totality fix, and the arm's layer priced at 42,240 node entries on the demo scene
docs: the bound-skip arm lands on the device, with the clamp finding and the no-visit-count caveat
wiki: the bound-skip carrier is the pool tail; slot 2 and the header are both closed (repo 0732018)
wiki: the Phase 1 exit lead sentence stops denying the device path its own rows measure (repo e4f7449)
wiki: which bound layer the device port should carry (repo 4f578b1)
wiki: the remaining R1 lever stated as work, with its binding-ceiling constraint (repo 196c2aa)
wiki: the march/read-back row keeps its ordering and loses its shared percentage (repo 0f678f4)
wiki: the paravirtual row keeps its ratio and loses its budget claim (repo d9d5344)
wiki: the march/read-back split is now a measured second vendor, not a forecast (repo a0d8ebb)
docs: R1's frame split - the read-back is 0.7-1.4% of the frame, and on-device shading is withdrawn as a budget lever (repo 894a6f1)
Mirror 96b0519: device-generated rays measured - the 66 MB was real and was not R1's floor, and the four-lever list loses its first entry
Mirror bf47df6 and e9ccbb0: S7's field roots are asserted constants on every gates runner, and the row no longer names a weaker check than the one that runs
Mirror a616836 and 9eb6a5d: S7's consensus half, and the Phase 1 headline that catches up with it
docs: the pitch demo's scene is content now - .ron -> assetc -> GPU (0 disagreements on 3,072 rays, two vendors) and the shipped binary boots the .bin with 17 of 18 published lines identical, plant 605->590 surfaces (repo 334fb7c, 739f7c4, f8b0938)
docs: AS4's asset side measured at the edit cadence - upload_into re-uses the shape's buffers, 1,843,200 depths identical to a fresh upload, refusal and stale-plant arms in-binary (repo b7e842a)
docs: R2's ten minutes ran on the device path - 36,000 frames, 3,072 hits asserted equal every one, and the read-back as the growing term (repo e057ce7)
R1's frame split at 1080p: three vendors agree on the surfaces, none on the clock, and the 640x360 saturation is retracted at R1's own grid
The span's gated integer reproduces on three runners; its durations do not
R2's ten-minute span is measured: 36,000 frames, one charge asserted on every one, and the plant that could not fire
docs: AS4's clause measured - one shape upload, bound per frame, and the rays are the real per-frame cost
docs: R2's edit loop, on the CPU reference and on the device - what a device march can assert is a count