Skip to content

Physics Soft Compression Work Budget

easylogic edited this page Oct 10, 2026 · 1 revision

Physics soft compression work-budget discovery — planned

Updated2026-10-10. Existing826plan6094502750.

Next independent SOFT-01 question — planned bounded work-budget experiment

After actual1065native6094492409, use existing826discovery scope while actualproductadoption is serial. This is NOT planned public1003/GPU1004/Viewer1005/host1006implementation. Original383f and projectionorder1c37 NO-GO remain unchanged. No duplicate issue needed:826 explicitly owns bounded model comparison. Actual selected private1c37 source/evidence is retained; this independent standalone numerical research has no missing publicAPI or screen dependency. Do not mislabel it adopted productcode.

Consumer and one variable

A developer needs a bounded4-second compression/recovery preview of a .4m/.2kg soft block,125nodes/384positive tetrahedra/604edges, same pinned vertices. Old volume-first8substeps x8iterations preserves volume but maximumedgeextension .302261>.25. Question: at SAME material/geometry/plane/pins/damping and8substeps, do16projectioniterations meet ALL unchanged base gates within8.33ms? This is a finite work-budget hypothesis, not stiffness retuning or calibrated material. Do not search arbitrary iterationcounts if16fails.

Only compare unchanged volume-first8x8 against explicitly named volume-first8x16. Copy original immutable fixture/topology/volume/compliance/projector/metrics/binary modules from1c37 byexactGitblob into NEW examples/physics-soft-budget-reference/reference/** with source/hashledger, retaining their relative structure. New candidate world/profile/trial/controls/saveddecoder stay separate. Oldpaths/privategraphs/packages/root/lock/GPU/Viewer remain unchanged. ActualselectedcurrentRoot/cleanisolatedbranch must be recorded at code start.

Exact inputs, outputs and errors to freeze at start

F64 SI/right-handedXYZ/Yup. .4m/.2kg grid, original125node mass distribution and pins, edgecompliance1e-3m/N and volumecompliance1e-10m5/N. Exact original XPBDlambda resets perinnerstep, deterministic volume->edge familyorder, plane/pin aftereachiteration, dampingexp(-4*innerDt), no gravity/bodymotion/sourceedit. dt1/120,8substeps,480steps/4s. Plane0..1s .4-.08t,1..2s .32,2..3s .32+.08(t-2),3..4s .4. Preserve originaltemplate/sequential floating arithmetic; only candidate iterations8->16. No hardstrainclamp, hidden extra substeps, compliance/material/gate change.

Copied375position/375velocity values, step/time and separate counts. Baseline totaledgevisits18,554,880/volume11,796,480; candidate predictedexactly37,109,760/23,592,960 for480acceptedsteps. Predictions must be measured. Wrongdt/cap/nonfinite/bound/nonpositive-volume candidate refuse before fullstatecommit; bitwise copied acceptedstate retained. Private research errors/lifecycle must not be called publicPhysicsErrorABI. Candidate hook remains TEST-only/OFF in timed runs. No hidden repair.

Unchanged numerical and actual evidence gates

ALL481states/finite375counts/positiveeverytet/pinposition+velocityexact, componentpositions<=2m/velocity<=100m/s, peaklocalvolumeerror<=.20, peakedgeextension<=.25, upperplane penetration<=.002m, finalmaxrestdisplacement<=.005m/finalrelativevolumeerror<=.02. Wholeoriginaltargets preserved. Fulltwo-run rawLE position/velocity/signedzero/counts repeated; every-step independent determinant/distance/corner/plane diagnostics, firstfailedsteps retained. Independent projector finite-difference/algebra/old8x8 originalraw reproduction/refusal/copies controls before measuredcampaign. No modelGO from recoveredfinalstate alone.

Onecompleteuntimedwarmup then3serial480-steptrials EACHmodel insameNode22session; fullworld prediction/projection/internalbounds/acceptedreturnedcopies timed, externaldiagnostics/IO/hash/construction outside. Eachp95<=8.33ms; actualall samples/median/min-max/counts. Rootalone campaigns; helpers disjointstaticimplementation only. Stop conditional qualification if ANY base/numeric/reference/cost gate fails; report bounded NO-GO rather than a new material search.

Only if candidateALLbase+costpasses, run same16recipe dt1/240/960 steps and require allbasegates/cost and maxmatchednodepositiondifference<=.005m at481common times. Then actualChromium source numerical-repeat/copiedstates/refusals/cost/bundleraudit/console0/ownedcleanup. This is a numericaltable reference, NOT3DViewer or FEM/stress certification. Saved losslessraw and independentstandalonedecoder/sourceledger must replay. No arbitrarypublicsoftprofileimplementation before modelGO/newfrozenpubliccontract/exactrequiredadoption.

Stage PLANNED; targets UNMEASURED. Precise copiedfile set/profile/error/controls/selectedlocalparent and cleanmeasurement freeze before code remains required. Native1065handoff/Integration/currentpendingeditorvalidation/publication staydistinct. No productpush/PR/deploy/pack/publish/release; no Planner routine message.

Clone this wiki locally