S2 · s2_h2_exact_tree · the exact H^2 tree: decodability, sheet identity, bits, float failure settles: t = 11/13 for the H^2 tree (D3), L3's control t = 3/4, the H^2/H^3 bit cost ratio; decides no claim cores, unchanged: hyperbolic_core.js 2.0.0 (Q, X, P, Rng), decoder3d.js TreeEmbedding3D (t = 3/4) ~ node v22.23.4 data: synthetic reduced words only; Rng(90000+depth) random words, Rng(90100+depth) 106+- words, Rng(90200+depth) H^3 90-random words, Rng(90001) planning replay 1. Construction, exact root directions, generators 0..5: (1,0) (3/5,4/5) (3/5,-4/5) (-7/25,24/25) (-7/25,-24/25) (-1,0) local directions, P.children(w)[k], k = 0..4: (1,0) (3/5,4/5) (3/5,-4/5) (-7/25,24/25) (-7/25,-24/25); back (-1,0) [PASS] all six directions are unit vectors exactly (3^2 + 4^2 = 5^2, 7^2 + 24^2 = 25^2) cosines between directions: (1,0)~(3/5,4/5) 3/5, (3/5,4/5)~(-7/25,24/25) 3/5, (-7/25,24/25)~(-1,0) 7/25, (-7/25,24/25)~(-7/25,-24/25) -527/625 so the angles are 0, +-53.13, +-106.26 (the "106-degree" pair), 180 degrees; the gap from 106 to back is 73.74 m t cosh sinh cosh+sinh tau = ln m 3 1/2 5/3 4/3 3 1.0986 5 2/3 13/5 12/5 5 1.6094 7 3/4 25/7 24/7 7 1.9459 12 11/13 145/24 143/24 12 2.4849 19 9/10 181/19 180/19 19 2.9444 20 19/21 401/40 399/40 20 2.9957 [PASS] for every m: cosh^2 - sinh^2 = 1 and cosh + sinh = m exactly (so e^tau = m, tau = ln m), and X.coshSinh(t) agrees [PASS] for every m: cosh = (m^2+1)/(2m), sinh = (m^2-1)/(2m) [PASS] every rotation R(u) is orthogonal and takes (1,0) to u; every step R(u)B(t) is Lorentz (M^T J M = J); J M^T J times M is I · 36 step matrices [PASS] D3: t = 11/13 gives cosh = 145/24, sinh = 143/24, cosh + sinh = 12, tau = ln 12 = 2.4849 [PASS] L3's control: t = 3/4 (m = 7) gives cosh = 25/7, sinh = 24/7, tau = ln 7 = 1.9459, the H^3 tree's own boost integer form of the m = 12 steps (index 0..5): common denominators 24, 120, 120, 600, 600, 24 (boost 24 times rotation 1, 5 or 25) 2. Exact greedy decodability: successes / 500 random reduced words per depth, Rng(90000+depth), the same words for every m depth 2 4 6 8 10 12 16 20 24 28 m=3 t=1/2 303 105 23 7 4 1 0 0 0 0 m=5 t=2/3 375 179 80 54 28 16 5 0 0 0 m=7 t=3/4 500 443 376 311 274 226 148 124 84 55 m=12 t=11/13 500 500 500 500 500 500 500 500 500 500 m=19 t=9/10 500 500 500 500 500 500 500 500 500 500 m=20 t=19/21 500 500 500 500 500 500 500 500 500 500 H^3 t=3/4 500 500 500 500 500 500 500 500 500 500 (decoder3d decodeExact, octahedral tree) exact ties among the best-scoring children: 0 in 390000 greedy decisions (none happened) [PASS] sheet identity x0^2 - x1^2 - x2^2 = 1 with x0 > 0, exactly in Q, on every H^2 point · 30000/30000 [PASS] every H^2 point computed two ways agrees: Q frame product (X.matmul, left to right) and BigInt integer steps (right to left, one reduction) · 30000/30000 [PASS] the integer form satisfies N0^2 - N1^2 - N2^2 = D^2 on every point · 30000/30000 [PASS] every successful exact decode ends exactly at (1,0,0) in the leaf's frame · 18721/18721 [PASS] sheet identity on every H^3 point (X.onSheet) · 5000/5000 first wrong level (0 = root letter) among the failing words at depth 28: m=3: 500 failures · 0:242 1:123 2:63 3:35 4:19 5:9 6:4 7:1 8:3 11:1 m=5: 500 failures · 0:113 1:105 2:85 3:60 4:53 5:21 6:18 7:11 8:9 9:10 10:4 11:1 12:4 13:1 15:1 16:2 17:1 21:1 m=7: 445 failures · 0:37 1:46 2:45 3:36 4:31 5:21 6:26 7:18 8:13 9:20 10:18 11:18 12:11 13:15 14:9 15:11 16:9 17:9 18:8 19:10 20:5 21:9 22:4 23:13 24:2 25:1 m=12: no failures m=19: no failures m=20: no failures planning replay, Rng(90001): 100 words at depth 12 then 40 at depth 28 from the same stream: t = 11/13 100/100 and 40/40; t = 3/4 48/100 and 1/40 [PASS] the replay reproduces the planning figures (100/100 and 40/40 at t = 11/13, 48/100 at t = 3/4) [PASS] L3's control: t = 3/4 decodes noticeably worse than t = 11/13 at depth 12 · 226/500 against 500/500 m with 500/500 at every depth 2..28: 12, 19, 20; the smallest such m among {3, 5, 7, 12, 19, 20}: 12 2b. Supplementary: where the decodability threshold lies (m outside the spec's set included; it decides nothing) the same 500 words per depth (section 2): depth 2 4 6 8 10 12 16 20 24 28 m=8 t=7/9 500 483 443 404 375 362 285 253 198 178 m=9 t=4/5 500 500 498 480 489 490 466 452 443 430 m=10 t=9/11 500 500 500 500 500 500 500 500 500 500 m=11 t=5/6 500 500 500 500 500 500 500 500 500 500 [PASS] sheet identity on every supplementary point · 20000/20000 exhaustive, every reduced word of length 1..6 (23436 words): m=7 17528 · m=8 21108 · m=9 23236 · m=10 23436 · m=11 23436 · m=12 23436 half-width of a child's subtree seen from its parent, max over descendants 0..d levels below the child (degrees; the bisector is 26.57 = atan 1/2): levels below 0 1 2 3 4 5 6 7 beyond the bisector (of the nodes) m=3 t=1/2 0.00 38.66 78.23 >=90 >=90 >=90 >=90 >=90 46712 of 97656 m=5 t=2/3 0.00 27.15 47.75 75.75 >=90 >=90 >=90 >=90 25614 of 97656 m=7 t=3/4 0.00 20.47 31.42 40.22 50.30 67.29 >=90 >=90 9034 of 97656 m=8 t=7/9 0.00 18.17 26.51 31.82 36.09 40.21 44.98 51.77 4640 of 97656 m=9 t=4/5 0.00 16.31 22.83 26.25 28.34 29.74 30.75 31.51 690 of 97656 m=10 t=9/11 0.00 14.79 20.01 22.32 23.46 24.06 24.38 24.55 0 of 97656 m=11 t=5/6 0.00 13.52 17.79 19.42 20.09 20.38 20.51 20.57 0 of 97656 m=12 t=11/13 0.00 12.45 15.99 17.18 17.61 17.77 17.82 17.85 0 of 97656 m=19 t=9/10 0.00 7.97 9.31 9.55 9.59 9.60 9.60 9.60 0 of 97656 m=20 t=19/21 0.00 7.58 8.78 8.98 9.02 9.02 9.02 9.02 0 of 97656 exactly below the bisector at 7 levels (2|x2| < x1 for every descendant): 10, 11, 12, 19, 20; beyond it (an exact witness of a failing word exists): 3, 5, 7, 8, 9 margin to the bisector at 7 levels (a lower bound on the half-width, so an upper bound on the margin): m=10 2.01 · m=11 6.00 · m=12 8.72 · m=19 16.96 · m=20 17.54 degrees [PASS] the half-width analysis agrees with sampling: every m whose subtree crosses the bisector fails some sampled or exhaustive word, and every m that stays below decodes all of them 3. Bits: max numerator / max denominator bits over the coordinates, each coordinate in lowest terms 3a. random words by depth (the words of section 2) depth 2 4 6 8 10 12 16 20 24 28 m=3 t=1/2 12/11 23/20 33/29 43/38 52/45 63/53 81/69 99/87 120/98 139/115 m=5 t=2/3 17/14 34/28 51/42 63/54 78/66 90/77 122/103 147/126 177/147 205/168 m=7 t=3/4 19/15 37/30 56/45 70/58 87/70 102/83 136/110 167/136 197/151 232/181 m=12 t=11/13 18/15 37/28 54/41 68/53 85/62 100/71 133/97 161/114 191/131 225/152 m=19 t=9/10 22/16 46/34 70/52 88/60 108/75 124/86 170/117 202/134 241/160 277/182 m=20 t=19/21 24/19 48/37 72/55 91/69 114/85 134/98 178/132 220/156 262/188 300/206 H^3 t=3/4 11/6 22/12 32/17 42/23 52/29 63/34 84/45 104/57 122/68 142/79 supplementary m at depth 28, the same random words, per coordinate: m=8 (t=7/9) 225/168 · m=9 (t=4/5) 189/114 · m=10 (t=9/11) 247/185 · m=11 (t=5/6) 235/167 3b. depth 28, per coordinate (numerator/denominator bits), and pt = the whole point over one common denominator m=3 random (500) x0 136/113 x1 133/114 x2 139/115 pt 529 straight (1) x0 88/45 x1 88/45 x2 0/1 pt 221 106+ (1) x0 126/126 x1 126/131 x2 130/131 pt 518 106+- (500) x0 146/126 x1 150/131 x2 150/131 pt 580 m=5 random (500) x0 205/165 x1 205/168 x2 205/168 pt 782 straight (1) x0 130/66 x1 130/66 x2 0/1 pt 326 106+ (1) x0 193/191 x1 198/196 x2 196/196 pt 788 106+- (500) x0 226/191 x1 229/196 x2 230/196 pt 885 m=7 random (500) x0 231/179 x1 229/177 x2 232/181 pt 876 straight (1) x0 157/79 x1 157/79 x2 0/1 pt 393 106+ (1) x0 206/204 x1 208/206 x2 206/209 pt 837 106+- (500) x0 250/204 x1 251/206 x2 255/209 pt 972 m=12 random (500) x0 225/152 x1 223/150 x2 218/148 pt 823 straight (1) x0 201/102 x1 201/102 x2 0/1 pt 504 106+ (1) x0 225/183 x1 229/188 x2 227/185 pt 877 106+- (500) x0 248/183 x1 251/188 x2 249/185 pt 943 m=19 random (500) x0 274/180 x1 276/182 x2 277/182 pt 1007 straight (1) x0 237/119 x1 237/119 x2 0/1 pt 593 106+ (1) x0 189/119 x1 190/122 x2 191/122 pt 694 106+- (500) x0 287/205 x1 298/219 x2 301/219 pt 1119 m=20 random (500) x0 298/205 x1 300/206 x2 298/204 pt 1105 straight (1) x0 243/123 x1 243/123 x2 0/1 pt 609 106+ (1) x0 320/248 x1 323/253 x2 321/250 pt 1225 106+- (500) x0 331/248 x1 335/253 x2 333/250 pt 1260 H^3 random (500) x0 142/79 x1 141/79 x2 142/79 x3 142/79 pt 637 straight (1) x0 157/79 x1 157/79 x2 0/1 x3 0/1 pt 393 90 fixed (1) x0 133/79 x1 133/79 x2 128/76 x3 123/73 pt 605 90 random (500) x0 133/79 x1 133/79 x2 133/79 x3 133/79 pt 606 [PASS] sheet identity on every depth-28 class point (H^2 all m, and H^3) [PASS] straight words recomputed in closed form: x0 = (m^56 + 1)/(2 m^28) for every m, and (7^56 + 1)/(2 7^28) in H^3 3c. by depth for m = 12 and H^3: max numerator / max denominator bits per class depth H2 random H2 straight H2 106+ H2 106+- H3 random H3 straight H3 90 fixed H3 90 rand 2 18/15 15/9 18/15 18/15 11/6 11/6 10/6 10/6 4 37/28 29/16 35/28 37/28 22/12 22/12 19/12 20/12 6 54/41 44/23 51/41 55/41 32/17 33/17 29/17 30/17 8 68/53 58/30 68/55 73/55 42/23 44/23 38/23 39/23 10 85/62 72/37 84/68 91/68 52/29 56/29 48/29 48/29 12 100/71 87/45 100/81 109/81 63/34 67/34 57/34 58/34 16 133/97 115/59 132/108 145/108 84/45 89/45 76/45 77/45 20 161/114 144/73 164/134 180/134 104/57 112/57 95/57 96/57 24 191/131 173/88 197/161 216/161 122/68 134/68 114/68 115/68 28 225/152 201/102 229/188 251/188 142/79 157/79 133/79 133/79 3d. the bit cost ratio at depth 28, H^2 over H^3 (t = 3/4), class against class: numerator, denominator, pt H^2 m = 12 (t = 11/13), the D3 choice: random 225/152 vs 142/79: 1.58x, 1.92x · pt 823 vs 637: 1.29x straight 201/102 vs 157/79: 1.28x, 1.29x · pt 504 vs 393: 1.28x 106+ / 90 fixed 229/188 vs 133/79: 1.72x, 2.38x · pt 877 vs 605: 1.45x 106+- / 90 random 251/188 vs 133/79: 1.89x, 2.38x · pt 943 vs 606: 1.56x H^2 m = 7 (t = 3/4), the same boost as H^3: the cost of the rotation denominators alone: random 232/181 vs 142/79: 1.63x, 2.29x · pt 876 vs 637: 1.38x straight 157/79 vs 157/79: 1.00x, 1.00x · pt 393 vs 393: 1.00x 106+ / 90 fixed 208/209 vs 133/79: 1.56x, 2.65x · pt 837 vs 605: 1.38x 106+- / 90 random 255/209 vs 133/79: 1.92x, 2.65x · pt 972 vs 606: 1.60x range of the per-coordinate ratios at t = 11/13 over the four classes: 1.28x .. 2.38x worst over all classes at depth 28: H^2 t = 11/13 251/188 bits, H^3 157/79 bits (1.60x, 2.38x) 4. Float64 decoding: successes / 500, the words of section 2 float = the same construction in floats (correctly rounded step entries, float frame products, float J M^T J, float argmax) exact>f = the exact point rounded once to float64, then the float decoder depth 2 4 6 8 10 12 16 20 24 28 m=3 float 303 105 23 7 4 1 0 0 0 0 exact>f 303 105 23 7 4 1 0 0 0 0 m=5 float 375 179 80 54 28 16 0 0 0 0 exact>f 375 179 80 54 28 16 1 0 0 0 m=7 float 500 443 376 311 274 196 0 0 0 0 exact>f 500 443 376 311 274 215 0 0 0 0 m=12 float 500 500 500 500 306 17 1 0 0 0 exact>f 500 500 500 500 368 48 0 0 0 0 m=19 float 500 500 500 395 20 2 0 0 0 0 exact>f 500 500 500 446 31 2 0 0 0 0 m=20 float 500 500 500 375 22 0 0 0 0 0 exact>f 500 500 500 436 30 1 0 0 0 0 H^3 float 500 500 500 500 500 500 57 0 0 0 (decoder3d point/decode) exact>f 500 500 500 500 500 500 83 0 0 0 float (same construction): last depth of an unbroken 500/500 run from depth 2, first depth at 0/500, first depth where float falls below exact (on this depth grid): m=3 last 500/500 at none · first 0/500 at 16 · float below exact from depth none (tau x depth there = -) m=5 last 500/500 at none · first 0/500 at 16 · float below exact from depth 16 (tau x depth there = 25.8) m=7 last 500/500 at 2 · first 0/500 at 16 · float below exact from depth 12 (tau x depth there = 23.4) m=12 last 500/500 at 8 · first 0/500 at 20 · float below exact from depth 10 (tau x depth there = 24.8) m=19 last 500/500 at 6 · first 0/500 at 16 · float below exact from depth 8 (tau x depth there = 23.6) m=20 last 500/500 at 6 · first 0/500 at 12 · float below exact from depth 8 (tau x depth there = 24.0) H^3 last 500/500 at 12 · first 0/500 at 20 · float below exact from depth 16 (tau x depth there = 31.1) float decoding never succeeds on more words than exact decoding at any (m, depth): yes 5. Time per point at depth 28 (machine-dependent) ~ m=3 exact point (Q frame product) 0.624 ms · exact decode 2.441 ms · float point 11.8 us ~ m=5 exact point (Q frame product) 1.078 ms · exact decode 4.470 ms · float point 12.9 us ~ m=7 exact point (Q frame product) 1.110 ms · exact decode 3.355 ms · float point 12.9 us ~ m=12 exact point (Q frame product) 0.836 ms · exact decode 1.057 ms · float point 10.8 us ~ m=19 exact point (Q frame product) 1.168 ms · exact decode 1.507 ms · float point 11.5 us ~ m=20 exact point (Q frame product) 1.293 ms · exact decode 1.821 ms · float point 11.4 us ~ H^3 exact point (decoder3d pointExact) 1.672 ms · exact decode (decodeExact) 0.793 ms 17/17 checks pass SETTLES: - D3: t = 11/13 gives cosh = 145/24, sinh = 143/24, tau = ln 12 exactly (cosh + sinh = 12); the exact H^2 tree at t = 11/13 decodes 500/500/500/500/500/500/500/500/500/500 of 500 random words at depths 2/4/6/8/10/12/16/20/24/28, with 0 exact ties; the sheet identity holds exactly on all 30000 points. - smallest m in {3, 5, 7, 12, 19, 20} with 500/500 at every depth 2..28: 12. Supplementary: m = 10, 11 also decode 500/500 at every depth; the exact half-width of a subtree (7 levels) crosses the bisector tan = 1/2 for m = 3, 5, 7, 8, 9 and stays below it for m = 10, 11, 12, 19, 20, so the threshold lies between m = 9 and m = 10; t = 11/13 keeps at most 8.72 degrees of margin (m = 10: 2.01, m = 11: 6.00). On the same depth-28 random words m = 10, 11, 12 cost 247/185, 235/167, 225/152 bits (max numerator/denominator). - L3's control: t = 3/4 (m = 7) decodes 226/500 at depth 12 and 55/500 at depth 28 (planning replay: 48/100 at depth 12 reproduced); the same t in the octahedral H^3 tree decodes 500/500 and 500/500. - bits at depth 28, t = 11/13 (max numerator/denominator): random 225/152, straight 201/102, 106+ 229/188, 106+- 251/188; H^3: random 142/79, straight 157/79, 90 fixed 133/79, 90 random 133/79. - the H^2/H^3 per-coordinate bit cost ratio at depth 28 is 1.28x .. 2.38x over the four classes (whole point over a common denominator: 1.29x, 1.28x, 1.45x, 1.56x). - float64 (same construction) at t = 11/13 keeps 500/500 up to depth 8 and reaches 0/500 at depth 20; exact rationals are required beyond that. ~ total runtime 69.2 s