date_scored,prediction,kind,point,lo80,hi80,observed,verdict,abs_error,brier,note 2026-07-18,QG-1 frozen MNIST config steps,interval,500,300,1000,300,HIT,200,,boundary hit 2026-07-18,QG-2 MNIST frozen-config test-render PSNR dB,interval,33,28,40,39.2,HIT,6.2,, 2026-07-18,QG-3 A1 anchor gap W1-W3 pts,interval,30,12,45,80.4,MISS,50.4,,hedged toward graph-net literature against own PO-9 mechanism 2026-07-28,QG-4 frozen CIFAR-10 config steps,interval,1000,500,3000,1000,HIT,0,,point exact 2026-07-28,QG-4b P(config A w32L2 passes CIFAR gate),probability,0.15,,,1,—,,0.7225,it passed; I under-rated the cheap config 2026-07-28,QG-5 CIFAR frozen-config median render PSNR dB,interval,27,22,32,40.1,MISS,13.1,,under-rated overparameterisation (8707 params vs 3072 target values) 2026-07-28,QG-7 CIFAR gate-CNN real-pixel accuracy pct,interval,66,60,72,67.45,HIT,1.45,, 2026-07-28,QG-8 sign of CIFAR gate gap (positive = renders lose),probability,0.8,,,1,—,,0.04,positive at the frozen config; flips negative at higher-fidelity configs 2026-07-28,H-S1-1 P1-P0 pts,interval,0.1,-0.4,0.6,-0.39,HIT,0.49,,TOST also declares equivalence at margin 1.0 2026-07-28,H-S1-2 W1-P1 pts,interval,-3.8,-6.5,-1.5,-3.23,HIT,0.57,, 2026-07-28,H-S1-3 W1-W2 pts,interval,2.0,0.0,6.0,-0.68,MISS,2.68,,assumed a nuisance that is null; sign wrong 2026-07-28,H-S1-4a W1-W3 pts,interval,80.4,79,82,80.43,HIT,0.03,,mechanism trusted; hit to the decimal 2026-07-28,H-S1-4b canonicalization recovery f(W4),interval,0.06,0.01,0.20,0.177,HIT,0.117,, 2026-07-28,H-S1-4c canonicalization recovery f(W5),interval,0.10,0.02,0.30,0.628,MISS,0.528,,forced a rewrite of the canonicalization claim per prereg S1 section 8.3 2026-07-28,H-S1-5 (W7-W3)-(W6-W3) pts,interval,15,5,35,-0.52,MISS,15.52,,both interventions are null; the contrast could not exist 2026-07-28,H-S1-6 W10 distance outside [W4 W5] pts,interval,0,-3,3,0,HIT,0,,W10 = 35.54 inside [28.19 64.41] 2026-07-28,QG-6 P(full 50k CIFAR path chosen by R-CIFAR),probability,0.35,,,1,—,,0.4225,full path chosen: 14.2 h projected against a 30 h threshold 2026-07-29,H-C1-1 P0 real-pixel MLP test accuracy pct,interval,53.0,47,59,55.808,HIT,2.808,, 2026-07-29,H-C1-2 P1-P0 accuracy points,interval,-0.4,-2.0,0.6,0.42,HIT,0.82,,render marginally beats real pixels at 40 dB 2026-07-29,H-C1-3 W1-P1 accuracy points,interval,-12.0,-22,-5,-11.94,HIT,0.06,, 2026-07-29,H-C1-4 W3 test accuracy pct,interval,13.0,10.5,17.0,12.644,HIT,0.356,, 2026-07-29,H-C1-5 W1-W3 perception gap accuracy points,interval,27.0,14,41,31.644,HIT,4.644,, 2026-07-29,H-C1-6 W1-W2 accuracy points (optimization noise),interval,-0.7,-2.5,1.0,-0.9,HIT,0.2,,direction W2>=W1 registered and observed 2026-07-29,H-C1-7 recovery fraction f(W4) template-free c_sort,interval,0.17,0.08,0.3,0.1075,HIT,0.0625,, 2026-07-29,H-C1-8 recovery fraction f(W5) c_align to theta0,interval,0.62,0.42,0.78,0.3244,MISS,0.2956,,clears the 0.30 falsifier but halves the grayscale value 2026-07-29,H-C1-9 recovery fraction f(W6) bounded group augmentation,interval,0.04,-0.02,0.12,0.128,MISS,0.088,, 2026-07-29,H-C1-10 recovery fraction f(W7) K-marginalization,interval,0.04,-0.02,0.12,0.1015,HIT,0.0615,, 2026-07-29,H-C1-11 (W7-W3)-(W6-W3) accuracy points,interval,0.0,-2.0,2.0,-0.842,HIT,0.842,, 2026-07-29,H-C1-12 recovery fraction f(W9) frame averaging R=64,interval,0.0,-0.03,0.05,0.0009,HIT,0.0009,, 2026-07-29,H-C1-13 recovery fraction f(W10) exact L=2 invariants,interval,0.38,0.12,0.62,0.534,HIT,0.154,,c=3 channel mechanism of P-C1-B 2026-07-29,H-C1-14 W8 canonicalize-then-augment test accuracy pct,interval,11.0,10.0,14.0,10.6467,HIT,0.3533,, 2026-07-29,H-C1-15 X1 forward W1-reader on W3 features pct,interval,11.5,10.0,15.0,10.43,HIT,1.07,, 2026-07-29,H-C1-16 X1 reverse W3-reader on W1 features pct,interval,12.5,10.0,17.0,12.014,HIT,0.486,, 2026-07-29,H-C1-17 W10 signed distance outside [W4 W5] accuracy points,interval,0.0,-3,3,6.628,MISS,6.628,,the bracket breaks upward: W10 overtakes W5 2026-07-29,P-C1-A probability separated ordering f(W5)>f(W10)>f(W4)>max(f(W6) f(W7) f(W9)),probability,0.65,,,0,—,,0.4225,observed f(W5)=0.324 f(W10)=0.534 f(W4)=0.108; the two exact methods cross over 2026-07-29,P-C1-B probability f(W10)_CIFAR > f(W10)_MNIST=0.269 (c=3 channel mechanism),probability,0.6,,,1,—,,0.16,observed f(W10)=0.534; mechanism call resolved correctly 2026-07-29,P-C1-C probability all three label-shuffle controls within 2 pts of chance,probability,0.75,,,1,—,,0.0625,"shuffles {'W1': 11.02, 'W3': 10.08, 'W5': 10.2}" 2026-07-29,P-S4e-1 planted control max R_theta worst width (pilot-informed),interval,5e-08,1e-09,1e-05,0.0,HIT,0.0,, 2026-07-29,P-S4e-2 kappa median local condition number at w=32 (pilot-informed),interval,0.006,0.003,0.012,0.00551,HIT,0.00049,,eps=1e-4 rung of the ladder 2026-07-29,P-S4e-3 warmstart recovery fraction eps=1e-4 w=2 (pilot-informed),interval,0.85,0.55,1.0,0.78125,HIT,0.06875,, 2026-07-29,P-S4e-4 warmstart recovery fraction eps=1e-4 w=32 (pilot-informed),interval,0.0,0.0,0.15,0.0,HIT,0.0,, 2026-07-29,P-S4e-5 best R_f over independent students w=2 (pilot-informed),interval,0.0003,1e-06,0.003,0.0,MISS,0.0003,, 2026-07-29,P-S4e-6 R_theta at the best-R_f independent student w=2 (pilot-informed),interval,0.35,0.05,0.65,0.0,MISS,0.35,, 2026-07-29,P-S4e-7 independent-student recovery rate R_theta<0.05 at w=2 (pilot-informed),interval,0.02,0.0,0.15,0.015625,HIT,0.004375,, 2026-07-29,P-S4e-8 production median R_theta same-image pairs w=32 (pilot-informed),interval,0.45,0.25,0.7,0.278772,HIT,0.171228,,mnist 2026-07-29,P-S4e-9 production median R_theta minus different-image null (pilot-informed),interval,0.0,-0.1,0.1,-0.001333,HIT,0.001333,, 2026-07-29,P-S4e-C probability the falsification criterion is met at some width,probability,0.15,,,1,—,,0.7225,conjecture 6.5 dies if 1 2026-07-30,H-G1-1 P0 real-pixel MLP test accuracy pct,interval,47.0,41,53,47.502,HIT,0.502,, 2026-07-30,H-G1-2 P1-P0 accuracy points,interval,0.2,-1.0,1.5,0.262,HIT,0.062,, 2026-07-30,H-G1-3 W1 test accuracy pct,interval,38.0,28,46,38.164,HIT,0.164,, 2026-07-30,H-G1-4 W3 test accuracy pct,interval,12.5,10.5,16.0,10.544,HIT,1.956,,at the lower edge 2026-07-30,H-G1-5 W1-W3 perception gap accuracy points,interval,25.5,14,36,27.62,HIT,2.12,, 2026-07-30,H-G1-6 recovery fraction f(W5) c_align (the discriminator),interval,0.45,0.28,0.66,0.3238,HIT,0.1262,,identical to RGB CIFAR's 0.324 at c=3 2026-07-30,H-G1-7 recovery fraction f(W10) exact L=2 invariants,interval,0.45,0.26,0.6,0.4929,HIT,0.0429,,close to RGB CIFAR's 0.534 2026-07-30,H-G1-8 recovery fraction f(W4) template-free c_sort,interval,0.13,0.06,0.24,0.0582,MISS,0.0718,,misses the lower edge by 0.002 2026-07-30,H-G1-9 recovery fraction f(W9) frame averaging,interval,0.0,-0.03,0.05,-0.0003,HIT,0.0003,, 2026-07-30,H-G1-10 median render PSNR dB,interval,59.5,55,63,59.7682,HIT,0.2682,, 2026-07-30,P-G-A probability f(W5)>0.50 (paper section 9 conjecture substantially right),probability,0.35,,,0,—,,0.1225,observed f(W5)=0.324 at c=1 against 0.324 at c=3 — unchanged 2026-07-30,P-G-B probability f(W10) f(W5)=0.324 2026-07-30,P-G-C probability all three label-shuffle controls within 2 pts of chance,probability,0.8,,,1,—,,0.04,"shuffles [10.58, 9.71, 10.22]" 2026-07-31,H-W11-1 recovery fraction f(W11a) (pilot-informed),interval,0.26,0.14,0.42,0.2653,HIT,0.0053,,"vs c_sort 0.177, c_align 0.628" 2026-07-31,H-W11-2 recovery fraction f(W11b) (pilot-informed),interval,0.44,0.3,0.62,0.5261,HIT,0.0861,,vs W10 0.269 2026-07-31,H-W11-3 W11a test accuracy pct (pilot-informed),interval,34.5,25,48,35.262,HIT,0.762,,"1,879,874 params at width 424" 2026-07-31,H-W11-4 W11b test accuracy pct (pilot-informed),interval,49.3,38,64,56.242,HIT,6.942,,"1,845,210 params at width 288" 2026-07-31,H-W11-5 f(W11b)-f(W10) the pooling question (pilot-informed),interval,0.17,0.02,0.34,0.2574,HIT,0.0874,, 2026-07-31,P-W11-A probability f(W11a)f(W10) i.e. learned pooling beats eigenvalue spectra,probability,0.7,,,1,—,,0.09,f(W11b)=0.526 vs f(W10)=0.269 2026-07-31,P-W11-C probability f(W11b)>f(W5) i.e. invariant reader overtakes alignment,probability,0.3,,,0,—,,0.09,f(W11b)=0.526 vs f(W5)=0.628 2026-08-02,H-S5-1 function-query accuracy K=4 pct (pilot-informed),interval,45,28,62,16.374,MISS,28.626,, 2026-08-02,H-S5-2 function-query accuracy K=16 pct (pilot-informed),interval,78,62,90,51.542,MISS,26.458,, 2026-08-02,H-S5-3 function-query accuracy K=64 pct (pilot-informed),interval,93,84,97,95.336,HIT,2.336,, 2026-08-02,H-S5-4 function-query accuracy K=256 pct (pilot-informed),interval,96,91,98,98.232,MISS,2.232,,exceeds the real-pixel MLP (97.97) 2026-08-02,H-S5-5 smallest K whose accuracy exceeds W5 64.41 (pilot-informed),interval,16,8,64,64.0,HIT,48.0,,the accuracy curve is far more sigmoidal than registered 2026-08-02,H-S5-6 FLOPs ratio W5 over function-query at that K (pilot-informed),interval,3.9,1.5,12,3.4162,HIT,0.4838,, 2026-08-02,H-S5-7 function-query P-random minus P-shared-det at K=16 pts (pilot-informed),interval,4.0,0.0,10.0,5.43,HIT,1.43,,fit-quality effect; weight access swings 80.4 pts on the same pair 2026-08-02,H-S5-8 learned minus frozen probes at K=16 pts (pilot-informed),interval,6.0,0.0,15.0,11.25,HIT,5.25,, 2026-08-02,P-S5-A probability function-query at K<=64 beats W5 at strictly lower FLOPs,probability,0.85,,,1,—,,0.0225,K=64 gives 95.34 vs W5 64.41 at 3.42x fewer FLOPs 2026-08-02,P-S5-B probability amortized weight access overtakes function-query at T<=100,probability,0.3,,,0,—,,0.09,"weight per-task decoder 3.743 MFLOP already exceeds function-query's whole 1.594 MFLOP, so amortizing the 1.704 MFLOP preprocessing never closes" 2026-08-02,P-S5-C probability learned probes beat frozen probes at K=16 by >2 pts,probability,0.6,,,1,—,,0.16,learned-frozen = +11.25 pts 2026-08-03,H-S6-1 delta_sym identity permutation B=0 pts,interval,10.0,2.0,30.0,62.904,MISS,52.904,,sign flips alone cost 63 of the 79 points; we registered 10 2026-08-03,H-S6-2 delta_sym identity permutation B=3 pts,interval,15.0,3.0,35.0,64.044,MISS,49.044,,same event as H-S6-1 2026-08-03,H-S6-3 permutation share of delta_sym at B=3 pts,interval,64.0,45.0,76.0,14.744,MISS,49.256,,"the two components' roles are the reverse of the registration: D_inf 63, permutation 15" 2026-08-03,H-S6-4 W11a recovery of delta_sym at B=3,interval,0.60,0.30,0.88,0.631,HIT,0.031,, 2026-08-03,H-S6-5 |W11b scattered minus unscattered| pts,interval,0.3,0.0,1.5,0.586,HIT,0.286,,"validity check: 85.39 scattered vs 84.81 unscattered, i.e. seed noise" 2026-08-03,H-S6-6 raw accuracy on scattered P-random B=3 pct,interval,13.0,11.0,15.5,14.422,HIT,1.422,, 2026-08-03,P-S6-A probability the permutation part dominates (identity-perm delta_sym < half),probability,0.80,,,0,—,,0.64,"observed 64.04 against half of 78.79 = 39.4; D_inf dominates instead" 2026-08-03,P-S6-B probability extra scatter on P-random costs less than 2 pts,probability,0.85,,,1,—,,0.0225,"+0.13 at B=0 and -0.50 at B=3: P-random is already group-saturated" 2026-08-03,H-S7-1 f(W10c) MNIST,interval,0.14,0.02,0.28,0.125,HIT,0.015,, 2026-08-03,H-S7-2 f(W10c) CIFAR-10,interval,0.22,0.05,0.42,0.216,HIT,0.004,, 2026-08-03,H-S7-3 f(W10) minus f(W10c) CIFAR-10,interval,0.31,0.11,0.48,0.318,HIT,0.008,,the quantity the external review asked for 2026-08-03,P-S7-A probability f(W10c) < f(W10) on both datasets,probability,0.75,,,1,—,,0.0625,0.125<0.269 and 0.216<0.534 2026-08-03,P-S7-B probability f(W10c) <= f(W4) on CIFAR-10,probability,0.55,,,0,—,,0.3025,"0.216 against c_sort's 0.108: the nonlinearity contributes on its own, so W10-vs-W4 is not a clean symmetry comparison" 2026-08-03,H-S9-1 f(W12) MNIST P-random,interval,0.55,0.35,0.72,0.917,MISS,0.367,,"registered between W11b and c_align; it beat both" 2026-08-03,H-S9-2 f(W12) minus f(W11a),interval,0.29,0.10,0.45,0.651,MISS,0.361,,same event as H-S9-1 2026-08-03,H-S9-3 f(W12) minus f(W11b),interval,0.02,-0.15,0.18,0.390,MISS,0.370,,same event as H-S9-1 2026-08-03,P-S9-A probability f(W12) exceeds f(W11a),probability,0.85,,,1,—,,0.0225,0.917 vs 0.265 2026-08-03,P-S9-B probability f(W12) exceeds f(W11b),probability,0.45,,,1,—,,0.3025,0.917 vs 0.526 2026-08-03,P-S9-C probability f(W12) exceeds f(W5),probability,0.25,,,1,—,,0.5625,"0.917 vs 0.628; per S9 section 4 the paper's practical claim is withdrawn" 2026-08-04,H-S8-1 W1 accuracy at 300 steps,interval,88.0,82.0,93.0,88.5,HIT,0.5,, 2026-08-04,H-S8-3 f(W5) at 300 steps,interval,0.6,0.45,0.72,0.501791,HIT,0.098209,, 2026-08-04,H-S8-2 W1 - W3 gap at 10000 steps,interval,75.0,60.0,85.0,75.94,HIT,0.94,, 2026-08-04,H-S8-4 f(W5) at 10000 steps,interval,0.45,0.2,0.65,0.458867,HIT,0.008867,, 2026-08-04,H-S8-8 f(W10) at 10000 steps,interval,0.25,0.05,0.45,0.25221,HIT,0.00221,, 2026-08-04,H-S8-6 median relative endpoint gradient norm at 10000,interval,0.0003,3e-05,0.003,0.006319,MISS,0.006019,, 2026-08-04,H-S8-7 median relative parameter travel at 10000,interval,0.35,0.2,0.6,0.197149,MISS,0.152851,, 2026-08-04,H-S8-5 f(W5) at 10000 minus at 300,interval,-0.15,-0.4,0.05,-0.042924,HIT,0.107076,, 2026-08-04,P-S8-A probability f(W5) at 10000 steps is still above 0.30,probability,0.7,,,1,—,,0.09, 2026-08-04,P-S8-B probability the W1-W3 gap at 10000 steps is still at least 50 points,probability,0.8,,,1,—,,0.04, 2026-08-04,P-S8-C probability the median relative gradient norm falls 10x from 300 to 10000 steps,probability,0.6,,,0,—,,0.36, 2026-08-04,A02-S8-1 probability P-S8-C resolves FALSE (the sweep does not reach stationarity),probability,0.72,,,1,—,,0.0784,registered in docs/prereg/S8-addendum-02.md (commit 117c5e6) before any 10000-step cell was decoded 2026-08-04,A02-S8-2 probability H-S8-6 misses high (grad norm at 10000 above 3e-3),probability,0.7,,,1,—,,0.09,registered in docs/prereg/S8-addendum-02.md (commit 117c5e6) before any 10000-step cell was decoded 2026-08-04,A02-S8-3 probability H-S8-7 misses low (travel at 10000 below 0.20),probability,0.75,,,1,—,,0.0625,registered in docs/prereg/S8-addendum-02.md (commit 117c5e6) before any 10000-step cell was decoded 2026-08-04,"A02-S8-4 probability H-S8-4 hits (f(W5) at 10000 inside [0.20, 0.65])",probability,0.85,,,1,—,,0.0225,registered in docs/prereg/S8-addendum-02.md (commit 117c5e6) before any 10000-step cell was decoded 2026-08-04,"A02-S8-5 probability H-S8-5 hits (f(W5) 10000 minus 300 inside [-0.40, +0.05])",probability,0.88,,,1,—,,0.0144,registered in docs/prereg/S8-addendum-02.md (commit 117c5e6) before any 10000-step cell was decoded 2026-08-05,H-S10-1 f(W12b),interval,0.6,0.35,0.8,0.602,HIT,0.002,, 2026-08-05,"H-S10-2 f(W12u) - f(W12b), the coordinate contribution",interval,0.26,0.06,0.48,0.2559,HIT,0.0041,, 2026-08-05,"H-S10-3 f(W12b) - f(W11a), the architecture contribution",interval,0.34,0.1,0.55,0.3367,HIT,0.0033,, 2026-08-05,P-S10-A probability f(W12b) < f(W12u),probability,0.85,,,1,—,,0.0225, 2026-08-05,P-S10-B probability f(W12b) > f(W11a),probability,0.9,,,1,—,,0.01, 2026-08-05,P-S10-C probability f(W12b) > 0.70 (architecture explains most of the step),probability,0.35,,,0,—,,0.1225, 2026-08-05,"H-S11-1 f(W12ub), MNIST",interval,0.55,0.3,0.78,0.5565,HIT,0.0065,, 2026-08-05,H-S11-2 the 2x2 interaction of grading and phasor lift,interval,0.0,-0.25,0.25,0.0131,HIT,0.0131,, 2026-08-05,"P-S11-C probability the 2x2 interaction is outside [-0.10, +0.10]",probability,0.3,,,0,—,,0.09, 2026-08-05,"H-S11-3 s(W12), FashionMNIST",interval,0.85,0.6,0.95,0.828,HIT,0.022,, 2026-08-05,"H-S11-4 s(W12), luminance CIFAR-10",interval,0.6,0.3,0.85,0.8263,HIT,0.2263,, 2026-08-05,"H-S11-5 s(W12), RGB CIFAR-10",interval,0.6,0.3,0.85,0.9645,MISS,0.3645,, 2026-08-05,P-S11-A probability s(W12) beats c_align on every dataset,probability,0.8,,,1,—,,0.04, 2026-08-05,P-S11-B probability s(W12) on RGB CIFAR-10 exceeds W10's 0.534,probability,0.55,,,1,—,,0.2025, 2026-08-10,P-S12-B probability all validity conditions are met on the first attempt,probability,0.7,,,0,—,,0.49,gate failed on the first attempt: median relative gradient norm 1.0e-4 to 7.3e-4 against a registered threshold of 1e-4 2026-08-10,"H-S13-1 W12 test accuracy, benchmark MNIST-INR",interval,88.0,78,94,93.082,HIT,5.082,, 2026-08-10,H-S13-2 W12 minus DWS,interval,2.3,-8,9,7.372,HIT,5.072,, 2026-08-10,H-S13-3 W12 minus ScaleGMN,interval,-8.6,-19,0.5,-3.488,HIT,5.112,, 2026-08-10,P-S13-A probability W12 beats DWS,probability,0.55,,,1,—,,0.2025, 2026-08-10,P-S13-B probability W12 beats NG-GNN,probability,0.2,,,1,—,,0.64, 2026-08-10,P-S13-C probability W12 beats ScaleGMN,probability,0.07,,,0,—,,0.0049, 2026-08-11,"H-S14-1 f(B), collapse",interval,0.86,0.72,0.95,0.7698,HIT,0.0902,, 2026-08-11,"H-S14-2 f(C), pad",interval,0.89,0.75,0.97,0.7557,HIT,0.1343,, 2026-08-11,H-S14-3 width effect f(C)-f(B) at fixed information,interval,0.03,-0.02,0.1,-0.0141,HIT,0.0441,, 2026-08-11,H-S14-4 information effect f(A)-f(C) at fixed width,interval,0.07,0.0,0.16,0.2088,MISS,0.1388,, 2026-08-11,P-S14-A probability the information effect exceeds the width effect,probability,0.7,,,1,—,,0.09, 2026-08-11,P-S14-B probability the width effect is at least 0.05,probability,0.35,,,0,—,,0.1225, 2026-08-11,P-S14-C probability arm B lands within 0.05 of the luminance arm,probability,0.6,,,0,—,,0.36,