CHALL TUNNEL
a live wind tunnel is being built here — by a machine, in the open
the live tunnel needs webgpu — chrome on an apple-silicon mac . the build continues
watch it being built → admin.chall.net/tunnel-missiontunnel.chall.net
TEST DECLINED
below conversion speed the osprey hangs from her rotors . this instrument measures wings . above the corridor she is tested like any other airframe . below it, an honest tunnel names what it cannot measure
the corridor, cited: wing lift builds 40–80 kn · wing fully effective, rotor cyclic locked out ~100–120 kn · the tunnel cannot be honest below the conversion corridor
SPEC — at this scale air clings like syrup and lets go early — re ~10³ in here, ~10⁷ up there . its flow holds on to far steeper angles
SPEC — two strakes ahead of the wing roll the air into standing vortices . lift from organized spin — invisible to a 2d tunnel
SPEC — the fat, deeply cambered wing . at this reynolds the camber does the holding
SPEC — the f-16's section quit ~2° ago at this angle . camber is doing the holding
SPEC — this is the tunnel's crude stall . see the textbook one → wing
SPEC — the leading edge wears a felt cover on the ground so it doesn't cut the crew . a wing this sharp stalls like a switch, not a hinge
SPEC — sharp edges fix the separation point . the reynolds gap barely matters to a corner
SPEC — four propellers pour fast air over the wing before she moves . lift the engines make, before the sky does . no propeller turns in this tunnel — she is tested gliding
SPEC — the rotor discs are drawn in hairline and kept out of the solve — a spinning disc is mostly air
SPEC — a 2d section at toy reynolds — direction is real, road numbers don't exist here
·
receiving specimen…
SPEC — this is the job . fairings, boat tails, chopped mirrors — every drag count a truck ever lost was lost like this: trace, test, revise, re-test . the tail matters more than the nose
· · · ·
move a point · add at a diamond · select + x deletes · smooth fairs the run — four verbs, the whole tool . z undoes, 50 deep
WHAT THIS IS
this tunnel is honest about being small . the solver is real — d2q9 lattice-boltzmann, computed live in front of you — and it is 2d, incompressible, and running at re ~10³–10⁴, three to four orders below flight . nothing on this page is flight data; everything is measured in this tunnel or cited to its source . the gaps are printed, badged EST, and taught.
THE SOLVER
d2q9 bgk with a smagorinsky eddy-viscosity term (Cs 0.1), fp16 storage with fp32 math, streaming in place through a single buffer — one copy of the lattice lives on your gpu, 1.6 million cells at 1792×896 — the H1 rung; slower gpus step down W1 1536×768 · W2 1280×640 · W3 1024×512, and the readout names the pick — dozens of substeps every frame at 60 fps . walls are halfway bounce-back with a moving-wall term, so the model rotates through the air rather than teleporting; cells the body uncovers are refilled at equilibrium . the outlet copies its neighbor and re-anchors density to 1 — without that anchor the whole domain slowly pressurizes, and we say so because we measured it.
one law binds every backend: the shared-constants oracle . all physics constants live in one exported object inside the cpu reference solver; gpu shaders interpolate those exact values into their source . no number is ever retyped into a shader string — the cpu oracle and the gpu instrument cannot drift apart by a typo.
τ and re, disclosed: τ₀ = 0.53 (lattice ν = 0.01), τ floored at 0.5005; the readout prints this run's chord reynolds live, in the 10³–10⁴ band . one hedge the solver itself earns: the eddy-viscosity term raises effective viscosity wherever strain is high, so the effective re near the body is lower than the nominal number . the stall you see is laminar-bubble physics moderated by that term — not flight-re turbulent stall . the gap line on every article teaches exactly this.
VALIDATION
the cylinder first . a frozen fixture (400×160, d = 24, 40,000 steps) sheds vortices at a strouhal number the literature has measured for a century: the cpu oracle reads 0.2139, the gpu instrument reads 0.2138 — agreement to 0.05% . the unbounded-cylinder literature band near this re is St ≈ 0.19–0.21 (shedding onset re ≈ 47, williamson 1996); our fixture's 15% blockage sits it at the top of the band, and the number that matters is the gpu-versus-oracle agreement, not the absolute.
the pair . the f-16's thin section separates at 9.0°; the a-10's thick cambered section holds to 13.5° — a 4.5° window, measured across three seeds with zero spread, geometry never tuned . the requirement was 2°.
the batteries . L0 is the deterministic cpu-oracle battery — seven lanes (strouhal, anchor, settled state, poiseuille, moving-wall sweep, mass, layout), ~70 s in node, exits nonzero on any red . this build's full board stands at 466/466 lanes across fifteen decks, run foreground before every ship . the gpu-versus-oracle diff lane runs both solvers in lockstep on a pinned 128×64 geometry and holds L∞ < 10⁻³ over 500 steps — 11/11.
the sphere, for the volume . a sphere at re ~300 sheds hairpin vortices at a strouhal number two independent groups measured within a percent of each other — johnson & patel (jfm 378, 1999) read 0.137, tomboulides & orszag (jfm 416, 2000) read 0.136 . this program's 3d cpu oracle, on the ratified fixture, reads 0.1354; the d3q19 gpu solver on the same fixture reads 0.1377 — inside the ±15% band two different grids honestly allow, and the lockstep diff between them holds L∞ < 10⁻³ over 200 steps (measured 0.000043) . the negative control matters as much: at re ~100 the same fixture goes steady and the seeded perturbation decays by four orders — an instrument that always sheds is not measuring anything.
and the part we care most about: you can run this yourself, in your own browser . add ?qa=st to the url and this page runs the frozen 40,000-step fixture on your gpu and prints the strouhal it measured . add ?qa=diff and it boots the pinned diff geometry for inspection . add ?qa=volst and it runs the 3d sphere fixture the same way . the validation is not a pdf — it is a mode of the instrument.
THE VALIDATION APPENDIX
v1–v6, measured . every number below is measured on named hardware or cited with its source — the two never blend . bench where not stated: apple m5 max 128 gb, chrome 151, headed, real gpu . each row ends with what it does not prove.
v1 · the zero — cylinder strouhal. the frozen fixture (400×160, D=24, 40,000 steps): cpu oracle St 0.2139, the page's fp16s webgpu solver 0.2138 (0.02% apart), zero clamps . literature for an unblocked cylinder at this re: ~0.20 — our +7% is decomposed, not hidden: 15% blockage elevates St (cited), and at re 240 the fixture sits above the re≈190 mode-a transition where 2d over-predicts St vs 3d experiment by construction . the fixture is a regression anchor, recorded-never-corrected . reproducibility is public: ?qa=st re-runs the fixture on the visitor's own gpu and prints the St it measured — with its own wall time — on the page; the replay re-bake ran it through all five rung paths — 5/5 within 0.00% of the default row . not proven: grid convergence — one point at one resolution is a zero, not a study.
v2 · the 3d zero — sphere pair. d3q19 sphere at re 289: cpu oracle St 0.1354, the page's gpu bench 0.1377 — both inside the cited band (johnson & patel 0.137, tomboulides & orszag 0.136; program band 0.12–0.16) . negative control: sphere at re ~100 stays steady — the detector can say no . diameter convention stated: D_eff = D+1 under halfway bounce-back; Cd 1.234 inside the schiller-naumann ±15% band under D_eff, raw-D 1.393 retained in the record . not proven: any quantitative 3d claim — chord ~44 cells, re ~10³; the badge (`VOLUME · QUALITATIVE · chord 44 cells · no numbers leave this view`) is the only number the volume ever emits.
v3 · implementation parity — lockstep diffs. 2d: the gpu solver runs 500 steps in lockstep with the independent cpu oracle on the pinned 128×64 geometry; gate L∞(u) < 1e-3, green in every gate run since first air (tick 3, 11/11) through the latest 466/466; the measured L∞ prints in every run . 3d: 200 steps vs the 3d oracle, L∞(u) = 4.28e-5, zero clamps, measured 2026-08-08 . not proven: physical truth — parity shows the gpu implements the same discretization as the oracle, not that either matches nature.
v4 · thermal honesty — soak-20. the shipped page, h1 rung, 82 samples over 20.5 continuous minutes: minute-10 mean τ 0.1219 ms/substep vs minute-20 mean 0.1247 → ratio 1.023 (law ≤1.1), 0 tier demotions, fps floor 119.0 — ~120 fps sustained for the full soak . software adapters hard-refused by the harness . not proven: cross-device performance — one machine, mains, lid open; on weaker silicon the governor demotes and prints it.
v5 · the slow record — long exposure l1. 33,554,432 true cells (8192×4096, d2q9 fp16s), 81,920 steps — one full crossing — in 245.4 s wall = 11.2 glups raw; repeat run 253.0 s (3.1% spread); portrait sha256 ed3a15bc…; 1600 MiB gpu buffer ledger; zero context loss — the live tunnel resumed at 120 fps . census discipline: ×2.0 the largest documented live client-side grid (roquefort 16.8M), claimed as single-run only; the live real-time cells row is not claimed — our live hero is 1.6M cells and says so . not proven: validation — this is a capability record, offline-in-browser, printed `computed in this browser · not real-time` on its face.
v6 · the pedagogy gate — the pair window. the claim under the wing lesson — thin section separates while the thick one holds — passed by numbers before it shipped: oracle window [9.0, 13.5]° at power 60 (width 4.5°, requirement ≥2°; f-16 across 3 seeds, spread 0.00°; re-attach hysteresis 8.5°/9.5°; geometry never tuned; re 720) . the live h1 tunnel measured its own window [8.0, ~12–13.5]°, and α_pair 11.25° sits inside both instruments' windows . not proven: aircraft stall — 2d sections, no sweep, no strake, five decades below flight re; the ORDERING is the claim, and the page's own reynolds line prints the gap.
standing gaps, named (the panel's ask): no grid-convergence study yet (St at D=24/48/96, richardson slope), no multi-re validation matrix (St(Re) at 100/150/190 vs williamson, sub-transition), no taylor–green order-of-accuracy row, no blockage-doubling bound, no measured fp16s-vs-fp32 delta table (today's cross-precision evidence is v1/v3 oracle parity) . they are the next cheapest respect available; nothing above pretends to cover them.
THE LIMITS
2d. these are sections, not airplanes — no wingtip, no sweep relief, no strake vortex unless it is drawn in the section . real wings get three-dimensional help this tunnel cannot give . in 2d the streamwise vorticity is identically zero: tip vortices and pillar swirl are absent from the equations, not under-resolved.
the volume. objects you load in yourself get a second solver — d3q19, the same laws in three dimensions: the same shared-constants oracle, fp16 storage streaming in place through one buffer, halfway bounce-back, the same ρ-anchored outlet . it runs 256×128×128 with a chord about 44 cells, and it is a different instrument, not a better one . 44 cells resolves a tailgate recirculation as structure and a pillar core as a smear; it closes the gap the silhouette card confesses, and it earns no digits doing it . so it prints none — the badge on that view says qualitative, states its own chord, and that badge is the only number the volume ever emits . the aircraft never enter it: honest 3d models of them do not exist here, and lofting one from two flat views would fabricate exactly the geometry that makes those aircraft themselves.
incompressible. no mach effects, no shocks . supersonic lives on paper here, cited — never in the solve.
tunnel re. 10³–10⁴, orders below flight . separation here is honest physics at THIS re, and the per-plane gap line prints what the full-scale number does differently.
solid inlets. engines are solid bodies; nothing flows through the nacelle . propellers do not exist here — the c-130 is tested gliding, her slipstream lives in spec voice only, and her prop disc is drawn as a dashed ghost the solver never sees.
the declined test. the v-22 ships as two truths . above the conversion floor she is a fixed-wing turboprop and the article measures her like any other airframe; below it she hangs from her rotors and a wing tunnel has nothing true to say — so the instrument prints a TEST DECLINED card instead of a number . she also carries no wing article: bell's section is proprietary, and her wing chip reads dim `wing · —` with the card as its door . refusing those two measurements is the strongest proof the rest are real.
the near-plate admission. thin sections are thin: the f-16's 64a204 spans 4.8 cells at its thickest and rides the minimum-feature floor over most of its chord — in this tunnel it behaves closer to a cambered plate than to its section . thick sections resolve honestly (the a-10's 6716 spans 19.2 cells, the c-130's 64a318 stand-in 21.6) . the f-104's biconvex 3.36% is the extreme case and the honest one: the one jet whose wing nearly is a plate in the sky — here the near-plate rendering is closest to true . we print which regime each article is in rather than pretending they are the same.
SOURCES
every silhouette is traced from a public-domain us-government document; each plane's dossier records the document, figure, mirror url, access date, and sha256 . every ratio prints its basis, every sweep its datum, every aspect ratio its span convention . EST marks a literature estimate; ◆ marks sources that genuinely conflict, with the conflict documented . the v-22 is the one exception: no public-domain side view survived the 500-point trace law, so her profile is constructed in-house from published dimensions and us-government photographs — est-grade, and badged est wherever it prints . her span prints its basis everywhere: nacelle-to-nacelle 14.0 m, never rotors-turning 25.8 m.
the sourcing law runs before the pen, and it has already fired twice . the f-16 trace came off a flight-manual mirror carrying an fms disclosure statement — the master was quarantined and re-based to nasa tp-1538, a public report, the same day . the a-10's t.o. 1a-10a-1 carries distribution c — caught before tracing began, re-based to fm 44-80, distribution a . those demotions are not embarrassments; they are the system working, and this page will keep printing them.
FIRSTS · AS FAR AS WE CAN CITE
more lattice updates per second than any browser fluid sim we can cite — 7.96 billion, sustained, at the demo rung — with 2,097,152 passive tracers coupled to the live solve. if we're wrong, tell us and this line changes.
measured on apple m5 max · chrome 151 · 2026-08-07 · sustained means the thermal plateau, never the first warm minute . the hero floor is 6.45 billion by construction at the shipped rung . the full prior-art census — who we compared against, in whose units, with every url — lives in the build record: admin.chall.net/tunnel-mission, re-run before every print . the census found real neighbors (a 16.8-million-cell webgpu smoke box; a d3q19 browser tunnel that takes stl files), and this line is scoped to survive them: updates per second, live, client-side, cited.
and one slow record, measured 2026-08-08 on the same machine: 33,554,432-cell single exposures — the largest single-run lattice we can cite in a browser page . the long exposure door runs 8192×4096 — twice the largest live browser lattice the census found (that 16.8-million-cell smoke box) — as a deliberate offline exposure in this browser: 4.1 minutes of measured wall clock for one crossing, progress printed the whole way, never sold as real-time . the plate it develops says so on its face: computed in this browser · not real-time . both timed runs and the portrait's sha256 live in the build record.
THE MACHINE
this instrument is building itself . an autonomous loop of agents works under a written law: an owner's decisions file, an append-only ledger, a spend ceiling enforced in code . physics waits for the oracle; taste waits for a human eye — artifacts sit behind gates until the owner looks . every number on this page was measured by the machine that ships it, on named hardware, dated . the owner holds the pen and a kill switch.
PRIVACY
zero runtime network requests — the page you load is the whole product, and it runs with the wifi off . no account, no analytics, no cookies.
the model shop runs in this browser only: scans are parsed, sectioned, and shelved locally; nothing uploads, nothing is logged. a share link carries the outline after the # — the fragment never reaches any server; sending the link is sending the shape.
THE LINES THE LAWYERS GET
aircraft names and designations are used for identification only . no aircraft manufacturer, armed service, or government has endorsed, sponsored, or approved this instrument . silhouettes derive from public-domain united states government publications . nothing here is flight data, and nothing here supports operational, engineering, or certification use.
the wall grows as planes land . one scale, one rail — true size is the composition . reference shapes hang at their declared size.
THE REFERENCE ROW — drawn clean · mount them, measure against them · the row never changes
YOUR SHELF — user specimens · kept in this browser only · a struck row keeps its serial forever
·
lowercase a–z · 0–9 · space · dash · 24 max · plain text, label register only
your scan is read in this browser and thrown away . only the ink outline remains .
up to three . each body is re-measured live, in this session — never a stored constant . the aircraft are not on this sheet and never will be .
scaled to equal height — this compares shape, not size .
2D SECTION · TUNNEL RE ~10³–10⁴ · RELATIVE ONLY · equal height · same tunnel settings · in this tunnel
THE PLACARD REDLINE · 450 KIAS or M 0.75, whichever is lower — near it the manual's own compressibility text: shocks form on the thick wing (korn M_dd ≈ 0.66 at t/c 0.16) · this airframe never goes supersonic; the bow-shock pen stays dry
nose radius + corner deflections EST — from the dossier's machFeatures corner list · every printed angle is the closed-form solve, two decimals · construction ink is the drawing office's blue
drawn from θ-β-M + Billig 1967 · not the live solver · transonic 0.7–1.2 left honest
billig, f. s., "shock-wave shapes around spherical- and cylindrical-nosed bodies," j. spacecraft & rockets 4(6), 1967, 822–823 · doi:10.2514/3.28969 · θ-β-M + prandtl-meyer per anderson, fundamentals of aerodynamics · γ = 1.4
a live wind tunnel is being built here — by a machine, in the open
the live tunnel needs webgpu — chrome on an apple-silicon mac . the build continues
watch it being built → admin.chall.net/tunnel-missiontunnel.chall.net