8ad09b47a6
The three neighbouring applications out of the published branch: the NLMS step size (anc), the ODE step controllers (odestep) and the relaxation factor (relax), with the six stage scripts that drive them and the five tests that cover them. They are not part of the root-finding paper and no document refers to them. stage18_bulletproof.py keeps claim A and loses claim B; tests/test_bulletproof.py loses the one test that reached into anc. Server: verify 548 OK / 0 BAD, tests x2 1205 passed (was 1240 with the three).
main
vvsandClaude Opus 52026-09-25 10:02:18 +01:00
886b3f4244
README rewritten for a public mirror: what the method is, the licence and the archive DOI, how to use the compiled library, what it costs on the corpus fixed before the code was frozen, how to reproduce, the layout, and what was refuted. The old text still stated 72 checks, 490 tests and an order of 1.944, none of which has been true for some time.
vvsandClaude Opus 52026-09-24 12:29:55 +01:00
ce5b42f284
PDF rebuilt after the history rewrite. Server: verify 548 OK / 0 BAD, regression x2 1240 passed, PDF 32 pages, 0 errors, 0 overfull.
vvsandClaude Opus 52026-09-22 11:23:40 +01:00
2bc57fcc83
History rewritten to remove working notes not meant for publication; Appendix D and the README say what the rewrite changed. README: the archive is described as code, data and history for the paper (manuscript included), cited by its DOI.
vvsandClaude Opus 52026-09-22 11:17:23 +01:00
fc91d3b8aa
DOI архива Zenodo 10.5281/zenodo.22894337: статья, README; PDF пересобран
vvsandClaude Opus 52026-09-22 10:31:09 +01:00
1c63b829d9
Zenodo package: the bundle carries HEAD as well, so a plain 'git clone sim-numerics.bundle' checks the branch out (the clean-room check could not clone it).
vvsandClaude Opus 52026-09-22 10:20:02 +01:00
9c1638af07
Zenodo package (open release) and the PDF rebuilt with the published history's hashes.
vvsandClaude Opus 52026-09-22 10:17:07 +01:00
5c0550a0de
Open release of A' under the Apache License 2.0: LICENSE (verbatim Apache-2.0 text), NOTICE, licence stated in the paper's availability section and in simbracket's metadata. The next method is not described anywhere in the tree or in the history reachable from this branch. README no longer names the server. PDF rebuilt from the text as it stands. Server: verify 548 OK / 0 BAD, regression x2 1239 passed, PDF 32 pages, 0 errors, 0 overfull.
vvsandClaude Opus 52026-09-22 09:40:25 +01:00
1eb2c1664f
Revision 3: exact-zero stop in section 4; Table 12 with the post hoc Lagouanelle variant (branch m>=2: corpus 4 5568, six corpora 18372, multiple/fractional 6540, APS 2960; A' 178/34/28), the controls' missing exit in its caption; the reading of Table 12 restated in 7.1, 6.6 and the conclusion - the analogy order transfers to the classical frame and works better there than in A's split (2809 vs 4042 on multiple roots), and A's layer as a whole (detector with exit) earns its place (+1233 vs -3934). Server: verify 548 OK / 0 BAD, regression x2 1239 passed, PDF 32 pages, 0 errors, 0 overfull.
vvsandClaude Opus 52026-09-22 09:12:54 +01:00
aa304cd262
POSTHOC_lag_branch_2026-09-22: Lagouanelle-in-Chandrupatla restricted to the branch m>=2 (revision 3, point 5), written after K3 and before this run; marked post hoc, outside the K3 decision rule.
vvsandClaude Opus 52026-09-22 09:03:17 +01:00
34b2a30d6c
K3 outcome (PREREG_estimated_transform, Linux run): the 6.6 claim stands. Corpus 4: A' 3656 vs Chandrupatla on sign(f)|f|^(1/m_hat) with the analogy order 3806 (A' 126/38/76, p=5e-4) and with Lagouanelle 7526; six corpora 13032 vs 15733 and 30991. On multiple/fractional roots the analogy order inside Chandrupatla closes 59%, 58%, 31% of the A'->oracle gap (CHA 322 vs 654, corpus 1 721 vs 1229, corpus 4 1766 vs 2159) and pays on simple roots (APS 3453 vs 2595); Lagouanelle on Chandrupatla's side-alternating triples reads a fractional order on 115 of 154 simple APS roots.
vvsandClaude Opus 52026-09-22 08:50:37 +01:00
2fe6d729a8
PREREG_estimated_transform_2026-09-22 (revision 2, point 5), before code and run: Chandrupatla on sign(f)|f|^(1/m_hat) with m_hat estimated during the run - Lagouanelle (K3-L) or A's analogy order (K3-A), hysteresis on the branches m>=2, m<=2/3; decisive on corpus 4; outcome rule including withdrawal of the 6.6 claim. Also diag_revision2.py (points 3, 4: the three unchanged Table 2 rows exhaust the budget on aps.13.00 under both criteria; exact-zero stops per method in Table 11).
vvsandClaude Opus 52026-09-22 08:46:50 +01:00
0a62f7babb
Revision step 3: restructured and shortened paper, full listing of A', code availability.
vvsandClaude Opus 52026-09-22 08:18:30 +01:00
fbf13cd592
Revision step 2 into the paper: section 9.1 (Tables 9b, 9c) as PREREG_controls prescribes. The analogy detector beats Lagouanelle-with-memory, Chandrupatla's band and no detector on corpus 4 (3656 vs 4488, 6236, 6165); with the multiplicity given, the classical transformation is 2.4-6.8x cheaper on every multiple or fractional root - stated as the price of estimating the order, in the abstract, 8.3, 8.6, the conclusion and as the first future-work question. verify_controls(): 26 checks. Server: verify 526 OK / 0 BAD, regression x2 1232 passed, PDF 30 pages, 0 errors, 0 undefined, 0 overfull.
vvsandClaude Opus 52026-09-22 07:53:58 +01:00
92c07d8a40
PREREG_controls outcome (Linux run): C1 passed for every detector on corpus 4 - A' 3656 against Lagouanelle-with-memory 4488 (84/121/35, p=8e-6), Chandrupatla's band 6236, no detector 6165; six corpora 13032 against 15171, 20668, 20749; 0 bound violations. K1: with the true multiplicity, library Chandrupatla/Brent on sign(f)|f|^(1/m*) beat A' on every multiple or fractional root (CHA 96 vs 654, corpus 1 358 vs 1229, corpus 4 901 vs 2159); level elsewhere.
vvsandClaude Opus 52026-09-22 07:38:26 +01:00
60294e84d9
Controls K1/K2: code and two PREREG amendments, committed before the full run.
vvsandClaude Opus 52026-09-22 07:37:41 +01:00
77312bedcf
PREREG_controls_2026-09-22: the controls, written before code and run.
vvsandClaude Opus 52026-09-22 07:30:48 +01:00
862feaff86
Revision 1, step 1: one scoring criterion, the correct order, and the internal inconsistencies.
vvsandClaude Opus 52026-09-22 07:01:36 +01:00
eeb9102d47
simbracket: A' as a compiled library (Cython, abi3 wheels for Windows and Linux, Python >= 3.10).
vvsandClaude Opus 52026-09-22 07:01:13 +01:00
0e05c92a1b
Title (Valeri): 'Root Finding by Similarity and Analogy: A Bracketing Method with One Declared Constant' - the method rests on both, and 'parameter-free' overstated it (one declared constant, two declared choices). .tex and PDF regenerated; verifier 467 OK / 0 BAD.
vvsandClaude Opus 52026-09-21 22:15:01 +01:00
23d0c73231
The .tex is no longer stale: regenerated from the canonical .md by build_tex.py (pandoc 3.1.3 + pdflatex on the server; marked GENERATED, never edited by hand). PDF: 28 pages, 0 errors, 0 undefined references, 0 overfull boxes. For LaTeX the .md got braces on four n_\max subscripts (fix_md_for_latex.py; rendering identical) and Table 10 was split into totals (10) and per-problem counts (10b) so that it fits a page; numbers unchanged. verify_article on the server 467 OK / 0 BAD; regression x2 1226 passed, 0 failed.
vvsandClaude Opus 52026-09-21 22:07:48 +01:00
d2a142c39e
All-corpora wins/losses table: ModAB added (totals, violations, A' per problem and total difference against it on every corpus).
vvsandClaude Opus 52026-09-21 21:43:43 +01:00
f80aeb5d13
Article: ModAB added - Table 8 row (2914 / 2436 / 5350; A' 118/11/25 and 32/9/4), Table 10 column (totals, bound violations, A' vs ModAB per corpus), how it is run (authors' package pymodab 1.0.9, same stopping rule, counts agree with the package's own), the evidence paragraph (ahead on every corpus p <= 0.0013, the one test 3656 vs 5535 183/34, ModAB's own benchmark all 91: 1507 vs 1588, 9 bound violations), 8.6: ModAB's linearity check at the midpoint named as the precursor of our exit judgment, abstract, conclusion, introduction, future work (the comparison no longer missing), reference 20. verify_article: every ModAB number checked; server 467 OK, 0 BAD; regression x2 1226 passed, 0 failed.
vvsandClaude Opus 52026-09-21 21:42:15 +01:00
37c81b027a
PREREG_modab outcome (server): M1 PASSED - A' ahead of ModAB (authors' package pymodab 1.0.9, same stopping rule, counts equal to the library's own counter). Corpus 4: A' 3656 vs ModAB 5535, per problem 183/23/34 p=7e-26. Paper protocol: APS 2438 vs 2914 (118/11/25), CHA 2005 vs 2436 (32/9/4). Six corpora: 13032 vs 16657, significant on every one. ModAB's own benchmark (Roots.jl f01-f91): all 91 - 1507 vs 1588 (56/13/22 p=1.5e-4); 78 - 1371 vs 1446. ModAB breaks the 2x-bisection bound on 9 problems (1 on corpus 1, 8 on corpus 4), A' on none. ModAB wired outside the published grid (EXTRA_BASELINES), which stays at 264.
vvsandClaude Opus 52026-09-21 21:35:36 +01:00
bfa23b223d
PREREG before install and run: comparison with ModAB via the authors' own package pymodab 1.0.9 (native wrapper of their C code); the same stopping rule as every rival (aTol 1e-12, rTol 8.9e-16, no |f| stop - read in ModAB.c), evaluations counted by our wrapper, budget 200. A' and ModAB frozen. M1 on the fourth corpus (total and sign test), M2-M4 reported (six corpora, paper protocol, guarantee, ModAB's own Roots.jl benchmark). Priority note: ModAB's linearity check at the midpoint is a precursor of our exit judgment, named in 8.6 whatever the outcome.
vvsandClaude Opus 52026-09-21 21:27:06 +01:00
b677283afc
Article: the similarity in the step is earned, not scaffolding; literature in place.
vvsandClaude Opus 52026-09-21 21:22:41 +01:00
667f243f54
Article proof-reading: the introduction no longer says the four new corpora were unseen (three were used to close the last two gaps; only the fourth was run once after the code was frozen); a repeated clause in 8.2 merged. verify_article also recomputes every sign-test p printed in Section 9. Server: 411 OK, 0 BAD; regression x2 1226 passed, 0 failed.
vvsandClaude Opus 52026-09-21 21:03:15 +01:00
16b9ab87f1
Article (canonical .md), step 2: abstract, introduction, a note in section 3, the close of 7.4, negative results 12-16, conclusion, future work and references 8-12 rewritten for the proposed method A'; two statements that could not be recomputed rephrased. verify_article: expected values updated to the regenerated sweeps, the gate cells renamed (progress.*; their CHA values had moved unnoticed, 2991 -> 3047, 2851 -> 2888, not quoted in the paper), and verify_flat_regime_and_oos() added - Sections 8-9, Tables 8-11, items 12-16 and the abstract, recomputed from server files and fresh diagnosis runs. Verifier on the server: 401 OK, 0 BAD, 'every number stated in the paper matches the sweep data'. Server regression x2: 1226 passed, 0 failed.
vvsandClaude Opus 52026-09-21 20:56:58 +01:00
8bf595d9be
Article (canonical .md), step 1: 42 stale numbers brought to the regenerated server sweeps (Tables 3-7, 5.3, 7.1-7.4, section 10), statements that no longer hold rephrased ((2)+(3) not worse than either alone on the wide-bracket set; the best constant-free configuration is no longer defect); the rule of section 3 renamed the base rule; new Section 8 (the flat regime recognised by analogy: the flag, the analogy with its own past, the translation limit, the split from the order, five more readings, the proposed method A', Tables 8-9) and Section 9 (out-of-sample evidence on four more corpora, Tables 10-11); negative results and conclusion renumbered 10-11. Abstract, introduction, conclusion and verify_article still to update.
vvsandClaude Opus 52026-09-21 20:44:31 +01:00
92fb8c5c04
A' on the paper's own protocol (published roots, server): APS 2438, CHA 2005 (4443) vs Chandrupatla 2787/2441 (5228); per problem 103/20/30 and 29/4/7; earned-budget variant 3608/2807; guarantee held under every safeguard. Leave-one-out ablation of A' on all six corpora: regime +5336, point +3115, sub +959, close +473, half +376, exp +338, G2 +89 (2/6 corpora), contra +28 (3/6) - every component positive in sum, G2 and contra marginal.
vvsandClaude Opus 52026-09-21 20:37:09 +01:00
2c458c51dc
Fix: _local_order divided by fp*fp that underflows to 0 for a tiny nonzero slope (ZeroDivisionError, caught by the sweep harness and charged as a failure - this is what had moved Table 6's order cells after the sign fix). Pinned by a test that fails on the old code. Every sweep file regenerated on the server (run_all_sweeps.sh). verify_article: skip result files that are not grid lists (it crashed on the out-of-sample dicts). Verifier now completes: 224 OK, 57 stale numbers in the paper to update (Table 6 again 7 of 7).
vvsandClaude Opus 52026-09-21 20:34:19 +01:00
80610e5c27
table_all_corpora: the problem count was hardcoded as 1055; it is 855 (154+45+206+78+132+240) - now computed. Totals unchanged.
vvsandClaude Opus 52026-09-21 20:20:11 +01:00
94d81f24d8
Wins and losses of A' on all six corpora (1055 problems), server numbers only: the seen-corpora run is replaced by its Linux rerun (differences of 1-3 from Windows). Totals: A' 13032, Chandrupatla 16481, Brent 24565, TOMS748 23649. A' vs Chandrupatla per problem significant on 5 of 6 corpora; the applied+GSL corpus is parity (1382 = 1382, 48/52/32 p=0.09).
vvsandClaude Opus 52026-09-21 20:18:54 +01:00
6d2531e6ea
PREREG_close_sub outcome - the ONE check on the fourth corpus (server): A' TRANSFERS. 240 problems: A' 3656 vs Chandrupatla 5717 (-36%), per problem 177/35/28 p=1e-27, guarantee held (Brent 50 and TOMS748 37 violations), all solved; A' below A (4547). Without the targeted classes: 2403 vs 3569, 107/31/22 p=1e-14. Ablation: sub earned (+893, D3 1807 -> 924), close NOT earned on this corpus (-2) though it earned on the seen applied corpus. Per class vs Chandrupatla: D1 random polynomial simple root 20/8/12 (433 vs 419), D2 multiple 39/1/0, D3 fractional power 37/1/2, D4 exp+sin sums 23/14/3, D5 compositions 25/8/7 (503 vs 1017), D6 odd around the root 33/3/4.
vvsandClaude Opus 52026-09-21 20:08:29 +01:00
8027d7f5ef
A' code frozen BEFORE the corpus-4 run (PREREG_close_sub): close (borrowed Dekker/Brent tolerance step, only after a kept promise) and sub (overshoot branch m <= 2/3; the shallow memory is not cleared by the exit - amendment documented). Development on the five seen corpora (uniform sign-change criterion): A' below A everywhere; vs Chandrupatla APS 103/31, CHA 31/10, new-1 109/65, Roots.jl 44/20 p=0.004, applied 48/32 p=0.09 with total 1382 vs 1380. Tests: 1223 passed (full), 358 (gate files).
vvsandClaude Opus 52026-09-21 20:02:43 +01:00
657df10634
PREREG before code and run: repair the two base-step gaps - close (Dekker/Brent's tolerance step, named as borrowed) and sub (the second branch of |1 - 1/m| > 1/2: enter the regime for m <= 2/3 on an overshooting unkept promise) - and test A' = A + close + sub once on a FOURTH corpus (corpus_fourth.py, seed 20260922, 240 problems in six new families, two of them targeted), committed now; development only on the seen corpora. F1-F3 transfer, F4 the repairs earned.
vvsandClaude Opus 52026-09-21 19:54:13 +01:00
c2b0eea0f6
PREREG_applied outcome (server): variant A does NOT transfer - total 1580 vs Chandrupatla 1382 (+14%), per problem 40/46/46 p=0.59; guarantee held, 132/132 solved. Wins where roots are (near-)multiple: Kepler e->1 20/11/4, van der Waals near-critical 9/7/4. Losses traced to two gaps in the BASE step (the plain rule loses the same): (1) no bracket closing at the end - on sin[3,4] the root is reached in 7 evaluations (|f| 1e-16) but the far end stays 1.3e-7 away and 17 halvings follow, where Chandrupatla steps a tolerance past the root (Dekker/Brent's device); likely also the IRR losses (1/0/11); (2) roots of order < 1 (sqrt|x| sgn x): the step overshoots and oscillates (+0.15, -0.10, +0.066 ... ratio 0.8), 84 vs 4 - the unkept-promise flag only sees undershoots, the regime only m >= 2.
vvsandClaude Opus 52026-09-21 19:49:36 +01:00
8ea89dd38d
PREREG before any run: third corpus - the GSL root-finding test suite (10, published) and 122 physics/engineering equations on grids fixed beforehand (Kepler incl. e up to 0.999, Colebrook-White, Lambert W, Planck/Wien, van der Waals CO2 incl. near-critical, IRR, finite square well). T1-T3 as before; GSL part reported separately; per equation breakdown.
vvsandClaude Opus 52026-09-21 19:47:21 +01:00
554ec0c0ae
PREREG_roots_jl outcome (server): variant A does NOT transfer by the stop rule on the Roots.jl / ModAB benchmark. 78 out-of-sample problems: total 1464 vs Chandrupatla 1566 (-6.5%, R1 holds), per problem 38/10/30 p=0.4 (R2 fails), guarantee held and all solved (R3); strict f24-f91 the same picture (1340 vs 1454, 35/6/25 p=0.25). B does not transfer either (32/11/35). Same structure as the first new corpus: multiple/flat roots 331 vs 481 (10:3); the rest 1133 vs 1085, per problem 28/10/27 - parity. Largest loss f32 (4 - tan x near the pole, 41 vs 12). Against Brent and TOMS748 A wins (47/25 p=0.013, 55/20 p=6e-5); they break the 2x-bisection bound on 11 and 9 problems, A on none.
vvsandClaude Opus 52026-09-21 19:40:05 +01:00
97af3ae767
PREREG before any run: second independent corpus - the Roots.jl bracketing benchmark (from the ModAB benchmark, Ganchovski & Traykov), f01-f91 transcribed; 13 problems equal to APS/Chandrupatla functions excluded objectively (7-point identity test), verdict on the other 78; strict view f24-f91; success = sign change within 1e-6 (no roots given, jumps and poles present). R1-R3 as in the first out-of-sample check.
vvsandClaude Opus 52026-09-21 19:38:26 +01:00
a86a9ed263
PREREG_out_of_sample outcome (server). Variant A (G2 + contra + remembered analogy + order point + exp limit + half) TRANSFERS: new corpus 3492 vs Chandrupatla 3899, per problem 105/34/67 p=0.0046, guarantee held, solved 205 = 205. Variant B (defect + remembered + point + exp) does NOT (total 3587 below, but per problem 95/33/78 p=0.22; loses the exponential class 9/5/26). Per class the transfer is carried by the multiple/flat-root class C3 (1194 vs 1792, 37/0/3); on C0-C2 A is at par per problem and ~6% dearer in total, C4 ~ par, C5 (near-cancelling) 691 vs 574. Ablation: every component earns its place on at least one set (contra only on APS/CHA: +9/+22, NEW -8; G2 only on NEW: +82). Addendum: C0.05 is SciPy's random step function - seeded and its cache cleared before every run for reproducibility; two unseeded runs gave the same verdicts.
vvsandClaude Opus 52026-09-21 19:33:07 +01:00
47361dabea
PREREG before any run: out-of-sample corpus and ablation for the two strongest variants. corpus_new.py (seed 20260921): SciPy 'original' (6) + five generated classes x 40 (smooth simple, transcendental simple, multiple/flat |x-r|^p p~U(2,20), exponential flank, near-cancelling terms), built and validity-checked only (no method has seen it). O1-O3 = transfers (total below Chandrupatla, per-problem sign test p<0.05, guarantee); O4 vs the plain rule; per class. Leave-one-out ablation of every component on APS, CHA and the new corpus.
vvsandClaude Opus 52026-09-21 19:30:27 +01:00
bbb16ec81c
PREREG_half_contra outcome: ACCEPTED (server = smoke in every projection cell; 0 violations in the guaranteed family).
vvsandClaude Opus 52026-09-21 19:24:31 +01:00
3d74b4531b
PREREG before code and run: remove the false promise 'the root is here'. Two models of which half holds the root: the step (S >= 1/2 -> the near half) and the bracket line (smaller |f| at the far end -> the far half); if they disagree, halve. Component 'half'; pairs without/with on none and the two accepted exp cells, +/- G2; H1-H4, goal kept; the post-crossing risk written down.
vvsandClaude Opus 52026-09-21 19:16:24 +01:00
cc82156cde
Why the translation limit fired late on the fun9.5 flank (diagnostics only). The limit itself reads correctly: at x = 10.89 the pair has lx 1.09, lf 1.5e-3 -> m = inf, and the regime is entered at the next flag (10.39), after which the flank is halved (5.19, 2.60, 1.30). It fires late because the flank configurations do NOT recur from 20.25 to 10.89: (1) the near-zero step 20.2502 -> 20.2501 (overconfident S ~ 1 from a slope dominated by f(81) = 1.5e35) makes the next triples degenerate (xi ~ 0, then -1.7e7) and pollutes the memory; (2) our three-point step on exp is not a translation - it alternates long/short (xi -1.43, -0.46, -4.10, -0.03, -13.9) and first repeats at lag 5 (state 10 ~ state 5). 9 evaluations from 20.25 to 10.39 where one halving would do: ~8 of the +10 gap on fun9.5.
vvsandClaude Opus 52026-09-21 18:49:19 +01:00
baa93963cf
PREREG_flat_exp_limit outcome: ACCEPTED (server numbers equal to the smoke in every projection cell; 0 violations in the guaranteed family).
vvsandClaude Opus 52026-09-21 18:44:23 +01:00
e887691d09
PREREG before code and run: the exponential limit of the analogy. When the analogous pair's f shrank but its x-scale did not (translation self-similarity, exp(x-1) = exp(x)/e), read m = inf - the boundary of the same definition, no new constant; the order point then is the midpoint. Composite component 'exp'. Pairs without/with exp on the four order-point variants +/- G2; E1-E4, goal retention; false-entry risk written down.
vvsandClaude Opus 52026-09-21 18:36:36 +01:00
485f25f165
Why the fun9 well is closed only partly (diagnostics only, method unchanged). The well itself IS closed: reaching the root's side takes 4 evaluations with contra/defect against Chandrupatla's 3 (13 and 33 before). The remaining +20 (75 vs 55) is after it: (1) the exponential flank on the widest brackets - our step moves ~1 per evaluation down exp(x) (f/f' = 1), 10 evaluations where Chandrupatla halves 3 times; the analogy entry comes late because exp is self-similar by TRANSLATION, not scaling - the pair's x-scale does not shrink, so m = ln(f ratio)/ln(x ratio) is undefined or negative; (2) overconfident near-zero steps - on a steep flank the three-point slope is huge, z ~ 0, S ~ 1, the model promises the root at a point that is not it: 8 wasted evaluations on fun9 (27 without the decisions).
vvsandClaude Opus 52026-09-21 18:35:07 +01:00
08863c170c
PREREG_flat_order_point outcome: ACCEPTED, and the GOAL is reached - below Chandrupatla on BOTH sets (server numbers, all projection cells equal to the smoke).
vvsandClaude Opus 52026-09-21 18:13:10 +01:00
d1be3bbd05
PREREG before code and run: split at the point from the analogous pair's order m (t = |FA|^(1/m) / (|FA|^(1/m) + |FB|^(1/m))) instead of the midpoint, in the flat regime entered by analogy (flatanalogpt, flatanalogmempt). Exit generalised to any regime point: f(x) nearer to FA(1-t) + FB t than to 0. Paired against the midpoint twins with the same entry; P1-P4, goal reported; the one-sided stall risk written down beforehand.
vvsandClaude Opus 52026-09-21 18:00:24 +01:00
63cf9fb768
Comparative table of wins and losses (server numbers): five variants of the proposed rule against library Chandrupatla, Brent, TOMS748 - totals, per-problem sign tests, wide-bracket families. Server defcontra grid matches the smoke in every projection cell but one (defect+G2 CHA 3183 vs 3182).
vvsandClaude Opus 52026-09-21 16:00:05 +01:00
3d9287ec2c
PREREG_defect_contra_flat outcome (smoke, all 199 problems). Goal (below Chandrupatla on both) NOT reached by any cell: best CHA 2537 vs 2441. Best total: defect+flatanalogmem APS 2530 / CHA 2561 = 5091 < Chandrupatla 5228 (first time), C1-C3 hold; ablation: defect earns APS +284 (costs CHA 9), remembered analogy earns CHA +649 (costs APS 7). contra earned without defect (APS 2806 -> 2785, CHA 3220 -> 3198; with the remembered analogy 2792 / 2537, the best CHA), redundant with defect (defect+contra = defect). fun9: 105 -> 81 (Chandrupatla 55). Remaining CHA gap 120: flat +88, fun1 +20, fun9 +26. Server regression x2: 1116 passed, 1 failed = the paper-number check.
vvsandClaude Opus 52026-09-21 15:57:15 +01:00
3c92d610d6
PREREG before code and run: both directions. (1) defect decision combined with the flat regime (analogy / remembered analogy) for canonical headroom; (2) contra: halve when the local secant's sign contradicts the bracket's - the fun9 cause (a near-cancelling left end, |f| doubling 17 steps in a row); traces: 3/5 fun9, 7/154 canonical, 0 flat. Composite prestep 'a+b+c', one per category; factorial grid on the proposed rule +/- G2; decided on the smoke; goal = below Chandrupatla on both; ablation of every component of the best cell.
vvsandClaude Opus 52026-09-21 15:45:03 +01:00
441373800b
PREREG_flat_analogy_mem outcome (decided on the smoke, all 199 problems): NOT accepted. Remembering the analogy until the next flag: CHA 2635 -> 2552 (G2: 2625 -> 2548; flat fun3-6 1234 -> 1151, M4 holds) but APS 2806 -> 2814 (G2: 2782 -> 2787 = Chandrupatla, M2 fails) - the extra false entries. Server regression x2: 1069 passed, 1 failed = the expected paper-number check.
vvsandClaude Opus 52026-09-21 15:06:37 +01:00
4bf8634757
PREREG before code and run: entry with the analogy remembered. The judgment (nearer to own past than to the line, m >= 2) is made at every gate state; once it holds, the run is flagged analogous and the regime is entered at the next unkept promise; an exit (midpoint judged linear) clears it. Stop rule M1-M4 (M4: CHA below flatanalog, 2635 / 2625 with G2); goal vs Chandrupatla reported.
vvsandClaude Opus 52026-09-21 15:00:19 +01:00
b7271a1ce4
Why the analogy entry is late on the flat families (diagnostics only, method unchanged). At flag moments before entry: unlike-past 118, no-memory 7, m<2 2 - the order is not the bottleneck. The period-4 cycle: the flag rises in the phases whose configuration does NOT recur (the undershoot whose relative step shrinks every cycle, and the crossing); the recurring phases have the flag down and read the true order (fun4 = 6(x-2)^5: m ~ 4.9). The entry demands both at the same state. Remembering the analogy until the next flag: flat entry state 12.9 -> 8.2; problems entered APS 3 -> 8, fun9 0 -> 1.
vvsandClaude Opus 52026-09-21 14:58:20 +01:00
1842eb0765
Flat regime: exit by midpoint similarity (flatstickexit), one judgment at entry and exit (flatjudge), entry by analogy (flatanalog).
vvsandClaude Opus 52026-09-21 14:55:45 +01:00
42bfab4330
PREREG before code and run: entry into the flat regime by analogy. The dimensionless configuration (xi, phi) of the last three points is nearer to a past one than to the line phi = xi, AND the pair's similarity exponent m = ln(f shrink)/ln(x shrink) >= 2, on top of the unkept promise; midpoint exit as flatstickexit. Basis (diag_flat_analogy.py, traces only): entry would fire on 3/154 canonical, 20/20 flat, 0/15 fun2/8/9. Stop rule N1-N3; goal (below Chandrupatla on both) reported; expectation written: APS ~2806, CHA ~2454.
vvsandClaude Opus 52026-09-21 14:48:30 +01:00
1dbda21182
PREREG before code and run: one similarity judgment at entry and exit of the flat regime. At every new point, before the bracket moves: flat if |f| < |f - L(x)| (L = the line through the bracket ends), linear if the reverse. Entry = unkept promise AND flat; exit at a midpoint = linear. Goal: below Chandrupatla on BOTH sets (J1 CHA < 2441, J2 APS < 2787); decided on the smoke (all 199 problems), server only to confirm. Basis (where_we_lose.py, saved results): with the exit the flat families are at par with Chandrupatla; the rest of the loss is false entries on simple roots.
vvsandClaude Opus 52026-09-21 14:40:12 +01:00
f8dd8d9931
PREREG before code and run: exit from the sticky flat regime by similarity at the midpoint. In the regime every step halves; at each midpoint f(M) is compared with the two predictions, linear L = (FA+FB)/2 and flat 0; the regime is left when |f(M) - L| < |f(M)|. Ordinal, no constant; re-entry by the same flag. Stop rule X1-X3; X6 = the exit must beat the sticky midpoint without exit.
vvsandClaude Opus 52026-09-21 14:32:57 +01:00
f8c17c4ed1
PREREG_flat_stick outcome: NOT accepted. Sticky regime under the projection: CHA 3220 -> 3019 (W1 holds), flat fun3-6 1823 -> 1057 (W4, near pure bisection 1044), guarantee holds (W3), but APS 2806 -> 6781 (W2 fails): the unkept-promise flag enters on 134/154 APS problems and 20/25 non-flat CHA problems, and a false entry now sticks to the end of the run. W6 fails: the order read from the promise is m = inf in 163/174 entries (r >= 1/e; our step is not a pure tangent), so the derived split equals the midpoint on CHA (3019 both). Audit bug fixed and pinned: an infinite |f| at an end made the split nan. Server regression x2: 999 passed, 1 failed = the expected paper-number check.
vvsandClaude Opus 52026-09-21 14:14:35 +01:00
1192c68501
PREREG before code and run: sticky flat regime. Entry = the unkept promise (r >= 1/4); from then on every step splits the bracket. Order m from the promise itself, r = (1-1/m)^m (r >= 1/e -> m = inf); split point t = |FA|^(1/m) / (|FA|^(1/m) + |FB|^(1/m)), which tends to the midpoint at m = inf. Ablation flatstickmid (sticky midpoint). Stop rule W1-W3; W6 for the derived ratio. Ceiling from sweep_full: flat problems bisected -> CHA 2441; split 1:k near side k=2.5 -> 2146, non-monotone in k, so k is derived, not chosen.
vvsandClaude Opus 52026-09-21 14:06:46 +01:00
5a1edffcfc
PREREG_flat_creep outcome: NOT accepted. Unkept promise AND creep under the projection: APS 2806 -> 2878 (V2 fails), CHA 3220 -> 3354 (V1 fails), flat fun3-6 1823 -> 1991 (V4 fails), guarantee holds (V3). With G2: 2782 -> 2842, 3169 -> 3318. Under safe CHA 3390 -> 3149, weaker than the single flag (3037). None of the three flat-regime gates pays under the projection: its budget already does their work. Server regression x2: 952 passed, 1 failed = the expected paper-number check (46 numbers after the signfix, pending the paper decision).
vvsandClaude Opus 52026-09-21 13:55:37 +01:00
35eb9416d3
PREREG before code and run: flat gate = unkept promise (r >= 1/4, no crossing) AND creep (last two moves the same way). Decided by the full grid directly (V1-V5, stop rule); firing rates reported for information only.
vvsandClaude Opus 52026-09-21 13:46:58 +01:00
9884b21204
PREREG_flat_twice outcome: stage 1 passed (twice-in-a-row fires on 27.8% of model steps on the flat families, 8.7% on APS, 4.3% on fun1; flat/APS ratio 3.20 against the single flag's 2.47), stage 2 FAILED: under the projection APS 2806 -> 2803 (W2 holds, the per-problem edge over Chandrupatla kept at 83/28/42) but CHA 3220 -> 3341 and the flat families 1823 -> 1975 (W1, W4 fail); with G2 the same. Under safe/step the twice gate is inert (3390 -> 3388) where the single one gave 3037.
vvsandClaude Opus 52026-09-21 13:45:20 +01:00
6a8bb2315f
PREREG before code and run: flat regime by the promise failing TWICE in a row (same r* = 1/4, a counter of consecutive unkept model steps, any other step resets it). Stage 1 detectability (T1, T2 against the single flag's 2.48 ratio), stage 2 in the method only if stage 1 passes (W1-W5, stop rule).
vvsandClaude Opus 52026-09-21 13:35:37 +01:00
e7e74d71e5
Flat regime (PREREG_flat_regime): the unkept-promise flag - after a model step that did not cross the root, r = f_new/f_old >= 1/4, the tangent-versus-halving break-even at m = 2 - RECOGNISES the regime (stage 1: 53.6% of model steps flagged on the flat families against 21.6% on APS and 17.0% on fun1; recall ten times the m>=2 estimator's 5.8%) and cuts the wide-bracket set by 350-490 wherever the budget does not already bind: safe 3390 -> 3037, step 3524 -> 3037, unguarded 4104 -> 3198.
vvsandClaude Opus 52026-09-21 13:25:56 +01:00
9df1aa5f13
PREREG before code and run: flat regime recognised by the model's unkept promise - after a model step that did not cross the root, r = f_new/f_old >= 1/4, the break-even of a tangent step against a halving at m = 2 ((1-1/2)^2). Stage 1 detectability (F1), stage 2 in the method only if stage 1 passes (Z1-Z5, stop rule).
vvsandClaude Opus 52026-09-21 13:02:40 +01:00
68df1d25db
Sign tests by product underflowed: FIXED everywhere (branch, not master - the paper is not yet updated). FA * fnew < 0.0 with FA ~ -1e-200, fnew ~ 1e-150 is -0.0 and '-0.0 < 0.0' is False, so on flat functions (x^19) the bracket update moved the wrong end, the root left the bracket and the run still reported convergence - the proposed rule returned 4.9e-9, bisection 3.5e-9, ITP 1.6e-8 with a claimed 1e-12 bracket; the 1e-6 harness scored them solved. Same defect in our bisection, false-position family, Ridders, Brent, ITP and twelve copies in stage scripts, simroot/sim.py and one test; all replaced by _opposite / _same_sign, pinned by test_sign_underflow.
vvsandClaude Opus 52026-09-21 12:49:01 +01:00
29ab49599c
PREREG before code and run: when is lengthening appropriate - G1 confirmed geometric series of three steps, G2 repeating step pattern with the period from the data (gate S_R > 1/2), G3 steps AND results (1/(1-min(S_step, S_res)), S_res = 0 after a sign change). Five-way paired grid, criteria Y1-Y6, stop rule: accept only if Y1, Y2 and Y3 all hold.
vvsandClaude Opus 52026-09-21 12:16:25 +01:00
99be33a632
A in the method (extrapolate='steps', default 'none' = published bit for bit): NOT accepted into the proposed rule by the declared stop rule - X2 fails. Proposed rule without -> with A: APS 2806 -> 2967 (X2 FAIL), CHA 3220 -> 2925 (X1 pass), flat fun3-6 1823 -> 1524 (X4 pass), guarantee intact, 0 violations in 39 family cells (X3 pass). The per-problem edge over Chandrupatla on APS is lost (87/24/42, p=9e-5 -> 42/62/49, p=0.53).
vvsandClaude Opus 52026-09-21 12:11:06 +01:00
a25cffad1d
PREREG before code and run: step extrapolation (A) inside the method - step/(1-S) with S the ratio similarity of the proposed step and the previous displacement, capped at the far end; paired grid; criteria X1-X5 and stop rule declared against the library baselines.
vvsandClaude Opus 52026-09-21 11:56:36 +01:00
2498c66358
Metric defect found by B's first run (mean = inf): an exact hit made the new width 0 and the floor (ulp of the root, 5e-324 when the root is 0) overflowed. Fixed to XTOL - an exact hit ends the run, so it is worth the halvings still needed to reach xtol - identical for every candidate; A re-run with it.
vvsandClaude Opus 52026-09-21 11:49:39 +01:00
2b0ee228b4
PREREG B before the run: outcome memory - bisect when the most similar past state (ratio similarity of the contraction rho, select not weight, ties -> most recent) bought less than one halving. Replay check, criteria B1-B4 and falsifier declared.
vvsandClaude Opus 52026-09-21 11:48:13 +01:00
3cce10bfdf
Outcome of PREREG A: PASSED on every declared criterion, both configurations. Paired, per one evaluation from the same state, bracket halvings: flat families 0.26 -> 2.87 (proposed) and 0.36 -> 1.16 (reading 3), i.e. above bisection's 1.00 (A1, A3); canonical set 8.21 -> 10.05 and 12.15 -> 12.96, no harm (A2); fun1 7.36 -> 9.01 and 7.28 -> 11.46. S > 0 on about half the canonical states (A4). Step A leaves the bracket on 6-28% of the states where it acts; excluded here, projected in a real method. One-step replay only - the whole-run effect needs the rule in the method, a separate pre-registration.
vvsandClaude Opus 52026-09-21 11:47:42 +01:00
781bd15253
PREREG A before the run: similarity of successive steps as the multiplier, step d_n/(1-S), S = ratio similarity of the two displacements (0 if opposite). Replay check, one extra evaluation per state for our step, step A and the midpoint; criteria A1-A4 and falsifier declared; the one-step limitation declared.
vvsandClaude Opus 52026-09-21 11:46:59 +01:00
d754689480
Outcome of PREREG_empty_step: both features FAIL by the declared criteria. Empty step (similarity of |f_n| to |f_n-1|): AUC 0.45/0.53/0.34 (proposed) and 0.46/0.51/0.31 (reading 3) against APS/fun1/other CHA; on the proposed rule it fires MORE on APS (36.0%) than on the flat families (18.9%). Repeating pattern with the period chosen from the data: separates the flat families from APS and fun1 (AUC 0.69/0.77 and 0.66/0.67; fires 77.1% vs 37.7%/38.2%, and 18.0% vs 6.2%/9.5%) but not from the other CHA families (AUC 0.41 and 0.22; fires 72.4% and 68.6%), so R1 fails and no rule is built.
vvsandClaude Opus 52026-09-21 11:23:47 +01:00
8279fb6341
PREREG before the run: the empty-step feature (similarity of |f_n| to |f_n-1|, rule S > 1/2, the bisection constant) and the repeating-pattern feature with the period chosen from the data (lag maximising the mean ratio similarity of rho_j and rho_j-p over the history, lags up to half the history). Detectability only; criteria E1/E2/R1/R2 and the falsifier declared; no rule in the method.
vvsandClaude Opus 52026-09-21 11:23:02 +01:00
0ec7142baf
Two audits, both on existing data or traces, method untouched. (1) Where S is evaluated: only in the step length (and simpair, and two refuted modes); the slope gate, the triple by shape, the far end and 'defect' decide by comparing - which is a similarity judgment of the ordinal kind ('which is more similar'), where every decreasing S gives the same decision. Their gains depend strongly on the S in the STEP: triple by shape -493 with the tent, up to +250 with others; defect -283 with the tent, up to +215. The records are the pairing of the tent step with those judgments.
vvsandClaude Opus 52026-09-21 11:16:08 +01:00
1b6c626f67
Valeri's decisions (1) competitors from the library, (2) variant a. (1) Chandrupatla is scipy.optimize.elementwise.find_root, Brent is brentq, no more per-problem best_of. Our Chandrupatla stopped on a bracket twice as wide as every other method (the factor the protocol corrects for ITP); fixed (+0.5*xtol) in baselines.py and in the two copies found by grep (stage16_blend, stage19_modern) and the traced mirror, and pinned: ours == library on every problem. Regenerating sweep_full.json changed exactly three rows - chandrupatla APS 2785->2787, CHA 2423->2441, brent CHA 4099->4115 - and nothing else (diff_sweeps.py: 0 unexpected).
vvsandClaude Opus 52026-09-21 10:53:00 +01:00
de90c7a089
Audit on Valeri's challenge: our Chandrupatla implementation stops below 2*xtol (tl = tol/span > 1/2) while every other method, SciPy's included, stops at xtol; the protocol corrected that factor for ITP and not here, and best_of let the looser run win. Corrected it costs Chandrupatla 2 evaluations on APS (2785 -> 2787) and 18 on CHA (2423 -> 2441): real, in the competitor's favour, and not what changes the picture. The public scipy.optimize.elementwise.find_root, the private _chandrupatla and ours corrected agree on all 199 problems exactly.
vvsandClaude Opus 52026-09-21 10:14:42 +01:00
8830765454
§3.1 and Tables 2-3 of the canonical .md brought to current data (Valeri's request). §3.1 said the projection's doubling 'is not a tuning constant' and that zero slack 'behaves as bisection'; both are refuted by measurement (stage12 part 3, rerun as measure_budget_factor.py): the canonical total is 2806 for every multiplier >= 1.25, but the wide-bracket total moves monotonically, 2752 at 1.25 to 4006 at 4, and multiplier 1 costs 4817/2997, nowhere near bisection's 7530/3090. So the proposed safeguard is the EARNED projection (credit, C0 = 1 bit: bisection + 1 step, the cheapest endowment within that promise, 8395 -> 7658 -> 6588), as code and verifier had already fixed but no text said; the scheduled projection stays as the one-constant end of the trade. The text says openly that this choice was made after a question, and that it was not made for the numbers - on the canonical set it costs 762 more.
vvsandClaude Opus 52026-09-21 08:03:21 +01:00
de1f8f726d
The new sections move into the canonical text, article/sim_numerics.md (Valeri's decision): section 7 'Where the similarity is read' with 7.1 per-place similarity, 7.2 the decision place, 7.3 the fun7 scoring artefact, 7.4 the gap decomposition; negative results 8-11 appended so items 1-7 keep their numbers; abstract paragraph; Future work rewritten (the refuted log transform removed); item 6 now says aps.13.00 is a scoring artefact, not a method failure. Translated into the .md's revised names: code 'ratio' = the tent max(0,1-|z|), code 'rational' = the ratio similarity 1/(1+|z|); the proposed method stays the .md's (tent, 3-point, projection).
vvsandClaude Opus 52026-09-21 07:48:55 +01:00
f5d036c9d8
Why the wide-bracket gap remains, taken apart on the SAME evaluations: 'order of the model' explains only a third. Of the 262 evaluations lost on 33 problems, 178 sit on fun3-fun6 - pure powers of order 3, 5, 9, 19 - where no interpolating model works (our step 0.41-0.81 bracket halvings per evaluation, IQI from our own points 0.01-0.17) and Chandrupatla wins by BISECTING 93-100% of his steps; our step is slower than bisection there and our 2x-bisection guarantee licenses it. His band inside our step confirms it per family: fun4/5/6 573/456/456 -> 296/239/235 against his 293/228/228, while +164/+111/+170/+144 on fun1/2/8/9 where it refuses good steps - slot 2's total was two opposite effects summed. Only fun1 (58) is model order: IQI from our three points 10.37 halvings against our 0.62; a convex two-point step cannot carry IQI's negative weight (428/428 non-degenerate states; five underflowed to -0.0).
vvsandClaude Opus 52026-09-21 07:38:16 +01:00
46c713cdd3
Regression runner on the server with pytest-xdist -n 14 (installed into ~/torchenv with Valeri's consent) and BLAS threads pinned to 1: 755 passed in 2:00 and 2:04, against 5:49 single-process on the server and 3:42 on the laptop. Single-process pytest on the server is slower than the laptop; the server only pays with parallelism.
vvsandClaude Opus 52026-09-21 07:10:31 +01:00
2357379c2d
Slot 2, the decision place: Chandrupatla's own acceptance band, put inside our step rule as an instrument, does NOT close the gap - CHA 3066 (project) / 2899 (credit) against his 2423, while APS doubles 2806 -> 5432, and the band is not idle (63.6% refusals). The fifth independent test of one conclusion: the remaining gap is the ORDER of the model the step is built on, not the place the similarity is read in. By the declared stop rule (M1 and M2 both false) the 'where is the similarity read' line is closed: all four places measured.
vvsandClaude Opus 52026-09-21 06:58:59 +01:00
15b94242e0
Valeri's second point measured: the similarity function belongs to the PLACE it is read in, not to the method. False position's three-century stall is a similarity read at order one on an approach of order m: with sigma = |f(far)| the residual ratio is (e/E)^m, so raising it to the measured 1/m recovers the distance ratio exactly. Paired inside each cell, the correction is cheaper in 7 of 7 safeguards for S = 1/(1+z) - the one function that makes that place exact inverse interpolation - and in only 29 of 60 cells over all nine functions, i.e. a coin. It does NOT beat Chandrupatla (2884 against 2683 and 2423), it HARMS the slope reading (2806 -> 2854), and it does not cure false position: 127/154 and 10676 become 128/154 and 10043, where Illinois buys 153/154 and 7045 with one constant. Estimator bias measured and stated: a pure cube reads 2.830, not 3
vvsandClaude Opus 52026-09-21 01:17:25 +01:00
cc2c499f63
Valeri's objection measured: the principle is right (a similarity step should be large far away and small near), and our sigma = |f'|D defeats it — at x=1e4 on a 2e4 bracket the normalised residual is 0.167, so S=0.833 'similar' and the step is small five thousand step-lengths from the root. Selecting sigma = min(|f'|D, max|f| at the ends) — no new constant — improves 403 to 390 on eight hand-picked wide problems but is WORSE over the grid: APS 3288 vs 2726, CHA 3111 vs 2683, so the mode is rejected by its pre-registered rule. Its extra solves are two fun7 problems landing inside the flat zero region, i.e. the scoring artefact and not a capability. Written into Negative results as the answer to the natural objection
vvsandClaude Opus 52026-09-21 00:19:03 +01:00
08ffa42e98
CORRECTION, prompted by Valeri: the step IS a fraction of the bracket — with the ratio similarity 1-S = min(1,|f|/(|f'|D)) so step = min(D,|f|/|f'|) — and my 'absolute length cannot march' was wrong, as was the follow-up guess that we creep along one side: measured, 91.5% of our probes DO shrink the bracket, more often than Chandrupatla's 87.5%. What differs is how much each probe removes: 0.93 halvings per evaluation against 2.26. A linear model of the value axis undershoots on a wide bracket of a convex function, so the new bracket is about half; inverse quadratic interpolation lands near the root and removes four fifths. The limit is the ORDER OF THE MODEL the similarity is read against, and all three parameter-free ways round it are now measured failures (cbrt, log change of variable, derived3_m). Paper corrected
vvsandClaude Opus 52026-09-21 00:12:58 +01:00
0fb75be3cf
the wide-bracket gap is NOT closed, and now it has an address: the STEP RULE, not the guarantee and not the reading places. fun1 is one function at growing widths — 7/8, 12/12, 19/15, 59/24, 63/44 — so we tie when narrow and lose as it widens, while our observed order is the HIGHER one (3.63 vs 2.02); 59 on a 2e4 bracket is what bisection needs, i.e. one halving per evaluation against their 2.3. With the safeguard removed entirely we still spend 46/57/61/68 against bisection 57/77/45/45 and Chandrupatla 24/44 — worse than bisection on two of them. |f|/|f'| is an absolute length and cannot march across decades; written into the paper as the stated limit. Also fixed a tracer bug that had best_of double-counting evaluations (2993 vs 1487)
vvsandClaude Opus 52026-09-21 00:09:18 +01:00
8d20816163
the scoring artefact is in BOTH sets: aps.13.00 joins the five fun7 problems — 6 of 199 problems where f underflows to exactly zero on a ~3.2e-2 neighbourhood of the declared root, so every method returns a genuine root (|f(x)| below the smallest normal) and every method is scored a failure. About 1150 evaluations of EVERY two-set total are this artefact; re-scored, all methods solve 199/199 except Brent 197. Ranking unchanged: ours-by-shape 2337 on APS against chandrupatla 2593, chandrupatla 1487 on CHA against our 1740. Written into the paper as a table with both criteria; pdflatex clean, verifier green
vvsandClaude Opus 52026-09-21 00:01:35 +01:00
6b5870f815
chasing Chandrupatla found a property of the BENCHMARK: on all five fun7 problems f vanishes identically near the declared root (f(0+t)=0 up to 3.2e-2), every method stops at a point where f(x)=0 exactly and is scored a failure by the |x-root|<=1e-6 rule — so 1000 evaluations of EVERY wide-bracket total are an artefact and the universal '40 of 45' is a scoring statement. Re-scored: 45/45 for everyone but Brent, ranking unchanged, gap to Chandrupatla 253 (17%) and diffuse across 8 families. Written into the paper with its reproducing scripts; pdflatex clean, verifier green
vvsandClaude Opus 52026-09-20 23:57:49 +01:00
ee816933eb
paper updated with the placement results and machine-checked: new section 'Where the similarity is read' (2527 on APS below every method measured, 5511 over both sets, the non-composition of the two readings, and the two negatives with their numbers), abstract extended, verify_article.py now also checks every new number against its own sweep file — all green, pdflatex clean
vvsandClaude Opus 52026-09-20 23:52:13 +01:00
22f19c57f0
similarity in two more places MEASURED: triple by shape gives APS 2527 — below the published 2806 and below EVERY baseline including chandrupatla 2785 — and far-end by dissimilarity gives CHA 2683 (the ITP target met) with 5511 over both sets (project target 6026 beaten, ITP-oracle 5871 passed). Guarantee intact everywhere, 193/199. Neither mode is inert (38.6% / 26.9% of steps changed). A+B do NOT compose: 2884/2859 is worse than either alone — select, never weight, and never assume composability
vvsandClaude Opus 52026-09-20 23:45:55 +01:00
04ee5c89b9
two more places for similarity: points='shape' (WHICH triple the model uses, argmin |phi-xi| over declared candidates — select, never weight) and direction='dissimilar' (the far end read by |f| rather than by distance); defaults bit-identical, P1-P5 pinned (7 green), inertness counters make an inert mode a failed EXPERIMENT rather than a pass; each grid writes to its own file
vvsandClaude Opus 52026-09-20 23:39:41 +01:00
e4769be9a3
slope-stability gate: the PROJECT'S OWN target is met for the first time — 5922 < 6026 over both sets at 193/199 with the guarantee intact (slope.ratio.derived3.project: APS 2848 + CHA 3074), CHA improves 4516 -> 2903 (x1.56), repairs 77.5% -> 55.3%. Honest cost: APS 2806 -> 2848, which moves us from 1st to 3rd on the canonical set (chandrupatla 2785, toms748 2820), so my stricter S1 fails while the project target passes — a positioning decision, recorded not resolved. Refusal CURRENCY is what decides: bracket progress 4449 vs model shape 2848, a 1601-evaluation difference for the same 'refuse before the step' idea
vvsandClaude Opus 52026-09-20 23:27:43 +01:00
a6116cd027
pre-step gate MEASURED: repairs 77.5% -> 15.9%, CHA hits the ITP target for the first time (2851 <= 2857), two sets 7440 (best of the line) — but APS grows 2806 -> 4449, so by the pre-registered stop rule the mode is NOT accepted. Cause named: my gate refuses on BRACKET PROGRESS, and this project had already measured that currency to be wrong (a secant step collapses the error while barely moving the bracket: 150 vs 86). Chandrupatla refuses on the SHAPE of the configuration. Next candidate derived, not borrowed: slope-stability gate
vvsandClaude Opus 52026-09-20 23:15:48 +01:00
0617243ed4
pre-step gate: prestep='none'|'progress' (default bit-identical), refuse the similarity step BEFORE spending an evaluation when it promises less progress than bisection — two lengths already in hand, no new constant; R1-R4 pinned by tests/test_prestep_gate.py (6 green); the gate grid writes to its OWN sweep_gate.json
vvsandClaude Opus 52026-09-20 23:12:34 +01:00
be10fc0b22
explained, with numbers: Chandrupatla wins the probe by REFUSING its model cheaply, not by trusting it — it accepts the interpolation on only 28% of steps, its gate costs zero evaluations (read from data already in hand) and its refusal branch still halves the bracket; ours repairs 77.5% of evaluations after the fact and drops to 0.92 halvings/eval on fun8/fun9, worse than bisection. In similarity terms: their similarity is a PRE-step GATE on dimensionless ratios (a neighbourhood of the linear configuration), ours is an IN-step SCALE on the residual magnitude — the PLACEMENT of the reading matters more than its form
vvsandClaude Opus 52026-09-20 23:05:22 +01:00
vvsandClaude Opus 5