Skip to content

Latest commit

 

History

History
288 lines (258 loc) · 21.8 KB

README.md

File metadata and controls

288 lines (258 loc) · 21.8 KB

Regularized Views Registered as Verified Upper Bounds

This directory holds the certificates of T-155: the exact views of the regularization run of 11 and 12 October 2026 whose sides lie below their cases’ verified upper bounds, registered as results under the regularization policy (row 3, “register first”, and the owner’s answer to its question 3 of 12 October 2026, “regularize and correct always”). Under grid-slack-2026-10-13/ it holds the second batch, T-156, the grid-slack view of $n = 290$ (below). Under the policy’s question 16, as the owner answered it on 13 October 2026, a side below the verified upper bound is registered as a result however it was made, and is a new upper bound only where no earlier certificate stated that side exactly, as Evan Daniel’s had at 33 of T-155’s counts, and no larger count’s verified upper bound is at or below it. T-156’s bound was not new: T-155’s certificate at $n = 291$ is at the same side and already implied it by monotonicity, and T-156 is kept as the certificate of $n = 290$’s own packing at that side. devtools.register_regularized_bounds wrote everything here and holds it to the case records.

What Is Here

Path What it is
views/n-NNN.yaml.gz Each registered view, a rational centre-basis Witness/v2, byte for byte as the run wrote it, gzip-compressed
selection.json One row per count: the view, its exact side and display, the bound the record held before and the margin, the printed side and closed form, the drawing decision, the identity measured on the retained view against the witness, and the regularization in the register’s terms; a grid-slack row also records the next count’s ceiling where it already implies the bound
exact-decision.json Each retained view decided over $\mathbb{Q}$ by both maintained exact routes
negative-controls.json Three mutants of each view, decided by both routes
grid-slack-2026-10-13/ The second batch (T-156), with the same four kinds of file

How the Views Were Chosen

The run drove Evan Daniel’s exact solver and regularizer, vendored under vendor/evand-regularize/ (MIT), over every count’s known-best witness with devtools.regularize_evand run (conservative, coincide, the source orientation, 80 digits), and this project’s same-side re-solve fixed each result at the witness’s side where it could. Each count yields up to two views, the regularizer’s certificate at the side its solve reached (solved) and the same-side view at the witness’s side $s_W$ (same). A view whose side lies below the case’s verified upper bound $U$ is held at row 3 of the policy’s decision table until it is registered.

select read the reports of the full run and its re-runs, a later run’s report replacing an earlier one’s at the same count, and took every count with a view at row 3 except three sets: $n = 17$, which the owner holds (think-x4v4), the 42 counts #518 re-houses, and the nine #553 re-housed after the run, whose witnesses those pull requests replace. At each count it registered the view with the least exact side, so that no view of the count lies below the new bound. A later select records every row-3 view’s side in its row (row_three_sides), so the choice can be checked without the run’s report. This selection predates that field. Its rows show the choice at 34 counts: at 33 one view alone was at row 3, and at one a solved view was taken over the same-side view, whose side is the witness’s own and which --check compares with it. At the other 25 a same view was taken over a solved one, whose side is in the run reports alone. It refused any report whose witness or bound was not the checkout’s.

Of 71 counts with a view at row 3, 59 are registered here and 12 are left out: $n = 17$; 70, 102, 103, 123, 126, 129, 146, 232, 267 and 295, which #518 re-houses; and 176, where #553’s certificate proves $s(176) \le 13.8128562532770564$ (T-151), below the side of the run’s view. At 33 counts the registered view is the regularizer’s certificate, and its side is exactly the side of Evan Daniel’s earlier certificate of the same packing (T-101), which the record had carried rounded up at the fourteen decimals the catalogue prints. Those 33 bounds are previously published, in Evan Daniel’s certificates of 5 October 2026, and T-101’s replay by two checkers that share no code with each other or with the source already proves each; there the case records cite that replay and the source’s own beside this package’s decision, and what is new is a second certificate at the same side and the record carrying the side exactly. At the other 26 it is a same-side view at the witness’s side: at 25 of them $6.6 \times 10^{-20}$ to $1.8 \times 10^{-19}$ below that certificate’s side, and at $n = 29$ $5.4 \times 10^{-20}$ below the endpoint of the interval certificate (T-009). Every registered side lies $5.4 \times 10^{-20}$ to $9.9 \times 10^{-15}$ below the verified upper bound it replaces.

At 37 counts the view draws the same picture as the witness; at the other 22 it redraws the packing. Each case record carries the registered side as its verified upper bound’s exact_form and keeps the decimal it printed before as value, which lies above the side and so still bounds it from above, as the records at $n = 70$ and elsewhere print a rounded value beside a rational exact form. The bound’s agreement with the catalogue’s printed side at the printed precision is therefore what it was: it agrees wherever the catalogue prints a decimal alone, apart from $n = 29$, and trails wherever the catalogue prints a closed form, which each side lies above.

How Each View Was Decided

decide decompresses each view and decides it, every pair and every wall over $\mathbb{Q}$, by sqpack.witness.exact_verify on its centre-basis witness and by devtools.check_rational_witness_independent on its rational corners. The run that made the views decided them with the same two checkers, so this is a reproduction with the producer’s code, not a second implementation.

controls makes three mutants of each view and decides each on both routes: the tightest pair’s second square moved along the pair’s best separating face axis by its gap plus one unit of the side’s denominator, doubled until the independent checker’s pair gap is negative; the box shrunk past its least top or right clearance by that unit; and the box shrunk by the unit alone. Both routes must refuse the first two and accept the third wherever the clearance exceeds the unit. The overlap is judged by the independent checker’s own pair gap, so its refusal of that mutant is circular; the mutant is invalid regardless, since a negative gap on every face axis means, by the separating-axis theorem, that the two interiors meet.

Both routes accepted all 59 views on 12 October 2026, 1,051,241 pair decisions each, in 438.4 CPU seconds for the two. The 33 regularizer certificates clear every wall by $5 \times 10^{-21}$ and every pair by about $10^{-20}$; the 26 same-side views touch a wall and a neighbour exactly, and both routes accept contact, which is sound for $s(n)$. All 177 controls reached their required outcomes on both routes (1,307.2 CPU seconds): the overlap and the shrink past the clearance were refused at every view, and the one-unit shrink was accepted at the 33 regularizer certificates and refused at the 26 same-side views, which touch the top or right wall. At those 26 views the least top or right clearance is 0, so the shrink past the clearance and the one-unit shrink are the same mutant, refused under both names: the 177 decisions are of 151 distinct mutants.

Reproduce

From packing/, with the project’s Python:

uv run --frozen --all-extras --group dev python -m devtools.register_regularized_bounds decide
uv run --frozen --all-extras --group dev python -m devtools.register_regularized_bounds controls
uv run --frozen --all-extras --group dev python -m devtools.register_regularized_bounds --check

--batch grid-slack after decide or controls decides the second batch. select needs the run directories, which are outside the repository; the solved points they hold are not retained here.

The Counts

Exact side is the registered side, cut and rounded up where long; Printed as the record’s value; Below the earlier bound by the margin under the verified upper bound the record held before; Below T-101’s side by the margin under the side of Evan Daniel’s certificate of the same packing; and Agrees whether the bound as printed and the catalogue’s printed side agree at the printed precision. Drawing is the run’s word for the view as a drawing, then its word once the view is registered; the regularized layer decides the drawing on its own.

n View Exact side Printed as Below the earlier bound by Below T-101’s side by Agrees Drawing Moved, largest
28 solved 5.82444461667405929702… 5.82444461667406 7.0e-16 0 yes no-change, then no-change none
29 same 5.93383346267692918969… 5.93383346267692918974379895098 5.4e-20 — no held, then admit 3, 1.3e-2
37 same 6.59861960924436011618… 6.59861960924437 9.9e-15 6.6e-20 yes held, then admit 7, 1.7e-1
39 solved 6.81072208306864885177… 6.81072208306865 1.1e-15 0 yes held, then review 11, 7.3e-2
41 same 6.92669309446880223159… 6.92669309446881 7.8e-15 6.9e-20 yes held, then admit 5, 2.4e-2
50 same 7.57142857142857142858… 7.57142857142858 8.6e-15 7.6e-20 no held, then admit 6, 9.6e-3
53 same 7.82287565553229529526… 7.82287565553230 4.7e-15 7.8e-20 no held, then refuse 3, 3.7e-2
54 same 7.84666719284348978295… 7.84666719284349 2.2e-16 7.8e-20 no held, then admit 8, 6.9e-2
55 solved 7.94577100750391297904… 7.94577100750392 7.0e-15 0 yes no-change, then no-change none
69 solved 8.82719465572973891502… 8.82719465572974 1.1e-15 0 yes no-change, then no-change none
71 same 8.94407155757031550657… 8.94407155757032 4.5e-15 8.9e-20 yes held, then admit 3, 2.8e-2
83 solved 9.63475764863108202936… 9.63475764863109 8.0e-15 0 yes held, then admit 14, 3.1e-2
87 solved 9.83881526994826226037… 9.83881526994827 7.7e-15 0 yes held, then admit 6, 2.9e-3
101 same 10.5355339059327376221… 10.53553390593274 2.4e-15 1.1e-19 no no-change, then no-change none
104 solved 10.7071067811865475246… 10.70710678118655 2.5e-15 0 no no-change, then no-change none
107 solved 10.8466671928434897831… 10.84666719284349 2.2e-16 0 no held, then admit 4, 1.1e-1
109 solved 10.9497474683058326710… 10.94974746830584 7.3e-15 0 no no-change, then no-change none
122 same 11.5355339059327376221… 11.53553390593274 2.4e-15 1.2e-19 no no-change, then no-change none
124 same 11.6568542494923801953… 11.65685424949239 9.8e-15 1.2e-19 no no-change, then no-change none
125 solved 11.7071067811865475246… 11.70710678118655 2.5e-15 0 no no-change, then no-change none
128 same 11.8250919682136870292… 11.82509196821369 3.0e-15 1.2e-19 yes no-change, then no-change none
145 same 12.5355339059327376221… 12.53553390593274 2.4e-15 1.3e-19 no no-change, then no-change none
147 same 12.6568542494923801953… 12.65685424949239 9.8e-15 1.3e-19 no no-change, then no-change none
148 same 12.6568542494923801953… 12.65685424949239 9.8e-15 1.3e-19 no no-change, then no-change none
149 solved 12.7071067811865475246… 12.70710678118655 2.5e-15 0 no no-change, then no-change none
151 solved 12.8228756555322952954… 12.82287565553230 4.7e-15 0 no held, then refuse 1, 8.3e-1
170 solved 13.5355339059327376222… 13.53553390593274 2.4e-15 0 no no-change, then no-change none
171 solved 13.5714285714285714288… 13.57142857142858 8.6e-15 0 no held, then admit 12, 2.1e-2
173 same 13.6568542494923801953… 13.65685424949239 9.8e-15 1.4e-19 no no-change, then no-change none
174 solved 13.7071067811865475246… 13.70710678118655 2.5e-15 0 no no-change, then no-change none
178 same 13.8466671928434897830… 13.84666719284349 2.2e-16 1.4e-19 no held, then admit 16, 1.3e-1
197 same 14.5355339059327376221… 14.53553390593274 2.4e-15 1.5e-19 no no-change, then no-change none
198 solved 14.5714285714285714288… 14.57142857142858 8.6e-15 0 no held, then admit 12, 2.1e-2
200 same 14.6568542494923801953… 14.65685424949239 9.8e-15 1.5e-19 no no-change, then no-change none
201 solved 14.7071067811865475246… 14.70710678118655 2.5e-15 0 no no-change, then no-change none
202 solved 14.7279220613578554394… 14.72792206135786 4.6e-15 0 no no-change, then no-change none
203 solved 14.7781745930520227686… 14.77817459305203 7.2e-15 0 no no-change, then no-change none
204 solved 14.8228756555322952954… 14.82287565553230 4.7e-15 0 no held, then review not one to one
205 solved 14.8244511461240882237… 14.82445114612409 1.8e-15 0 yes no-change, then no-change none
226 same 15.5355339059327376221… 15.53553390593274 2.4e-15 1.6e-19 no no-change, then no-change none
227 same 15.5710678118654752441… 15.57106781186548 4.8e-15 1.6e-19 no held, then admit 1, 2.9e-1
229 same 15.6568542494923801953… 15.65685424949239 9.8e-15 1.6e-19 no no-change, then no-change none
230 solved 15.6829268292682926831… 15.68292682926830 7.3e-15 0 no held, then admit 32, 3.5e-2
231 solved 15.7071067811865475246… 15.70710678118655 2.5e-15 0 no no-change, then no-change none
233 solved 15.7781745930520227686… 15.77817459305203 7.2e-15 0 no no-change, then no-change none
234 solved 15.8228756555322952955… 15.82287565553230 4.7e-15 0 no held, then admit 6, 9.9e-2
235 same 15.8266056334285680696… 15.82660563342857 1.9e-15 1.6e-19 yes no-change, then no-change none
257 solved 16.5355339059327376222… 16.53553390593274 2.4e-15 0 no no-change, then no-change none
260 same 16.6568542494923801953… 16.65685424949239 9.8e-15 1.7e-19 no no-change, then no-change none
262 solved 16.7071067811865475246… 16.70710678118655 2.5e-15 0 no no-change, then no-change none
264 solved 16.7781745930520227686… 16.77817459305203 7.2e-15 0 no no-change, then no-change none
265 solved 16.7781745930520227686… 16.77817459305203 7.2e-15 0 no no-change, then no-change none
291 same 17.5355339059327376221… 17.53553390593274 2.4e-15 1.8e-19 no no-change, then no-change none
293 solved 17.6341463414634146344… 17.63414634146342 5.4e-15 0 no held, then admit 27, 3.5e-2
294 same 17.6568542494923801953… 17.65685424949239 9.8e-15 1.8e-19 no no-change, then no-change none
296 solved 17.7071067811865475246… 17.70710678118655 2.5e-15 0 no no-change, then no-change none
298 solved 17.7781745930520227686… 17.77817459305203 7.2e-15 0 no held, then refuse 2, 7.8e-1
299 solved 17.8228756555322952955… 17.82287565553230 4.7e-15 0 no held, then review not one to one
300 same 17.8241233884785434709… 17.82412338847855 6.5e-15 1.8e-19 yes no-change, then no-change none

The Grid-Slack Batch

The owner wrote on 13 October 2026: “Any new upper bounds should also be recorded as results. Pure regularization should not be considered a new result, but it is a credited part of the frontier record” (the policy’s section 9, question 16). The policy reads it so: a side below the case’s verified upper bound $U$ is registered as a result however it was made, and it is a new upper bound only where no larger count’s verified upper bound is at or below it; otherwise the result says that its bound follows by monotonicity and names the result that proves it. A view of devtools.regularize_grid_slack, this project’s port of Evan Daniel’s coincidence placement, whose side lies below $U$ is registered so, as a batch of its own; T-156 registered the first, and its bound was not new (below).

The view. The grid-slack run of 13 October 2026 at 5 degrees, recorded in the regularized layer’s index as grid-slack-deg5-1013-a, made a view of $n = 290$ at the witness’s own side, $s_W = 70142135623730950488016887242097/4000000000000000000000000000000$, which is $17.53553390593273762200422181052425$ exactly. That lies 2.4e-15 below the $U$ the record held, Evan Daniel’s certificate side rounded up at fourteen decimals (T-101), and 1.8e-19 below that certificate’s own side. The catalogue’s closed form, $14 + (5/2)\sqrt{2}$, lies 4.8e-33 below it, so the bound still trails the printed side. It is the same packing certified at a tighter rational side, not a new one: no square moves by more than 6e-31 and none turns. Evan Daniel’s regularizer gave no view of $n = 290$ below $U$ in the full run, whose selection does not list it. Run again here at 0.5 degrees, the generator’s default, the generator made the same view byte for byte.

A bound already implied. $s(290) \le s(291)$, since deleting a square from a packing of 291 leaves a packing of 290 in the same square. T-155 registered $n = 291$ on 12 October 2026 at exactly this side, so from that day its certificate proved the bound T-156 registers, while $n = 290$’s verified upper bound still lay above the next count’s. Review B of #587 found this, deleting square 139 from T-155’s certificate, the square the $n = 290$ witness lacks, and deciding the rest valid on both exact routes. The owner decided on 13 October 2026 to keep T-156 and say so: its bound follows from T-155 by monotonicity, and the grid-slack view is a certificate of $n = 290$’s own packing, the catalogue’s, at that side. The selection row records the next count’s ceiling where it implies the bound (implied_by), and the case record says both.

Selection. select --batch grid-slack --run-dir RUN ... --n 290 decides each named run’s view by the policy’s table against the case record’s $U$ (regularize_grid_slack.decided) and takes the row-3 view with the least exact side, the first run named among equals. Its row keys every row-3 side by run (run_sides, runs_at_row_3), where a row of the first batch keys them by output (row_three_sides, outputs_at_row_3), and records the witness’s own side as read from the witness (witness_side), which --check reads again. It refuses a view of another generator, one not derived from the checkout’s witness, one not at that witness’s own side or labelled other than the same-side view, one whose side lies above the next count’s verified upper bound, which already proves more than the view and whose count is registered by monotonicity instead, two runs of one name, and a count the owner holds or another batch registers, and it writes the view and the selection only once every refusal has run.

Decision. Both maintained routes accepted the view on 13 October 2026, 41,905 pair decisions each, in 14.9 CPU seconds for the two; it touches a wall and a neighbour exactly, and both accept contact. The three controls reached their required outcomes (48.2 CPU seconds): the overlap and the shrink past the clearance were refused, and so was the one-unit shrink, which at a least clearance of 0 is the same mutant.

Drawing. No change, before registration and after: the view draws the same picture as the witness, so the atlas draws nothing new at $n = 290$.

n View Exact side Printed as Below the earlier bound by Below T-101’s side by Agrees Drawing Moved, largest
290 same 17.5355339059327376221… 17.53553390593274 2.4e-15 1.8e-19 no no-change, then no-change none

Credit

The packings are the ones the Kingbird catalogue credits to their finders, as T-101 and T-009 record them, and the register entry’s regularizations name those results. The solver and the regularizer are Evan Daniel’s (evand/square-packing, MIT); the run, the same-side re-solve, the exact decisions and the registration are this project’s. T-156’s view was made by this project’s port of his coincidence placement, not his code; the method stays credited to him, and the regularization, the decision and the registration are this project’s.