Skip to content

Latest commit

 

History

History
823 lines (727 loc) · 58.4 KB

regularization-policy.md

File metadata and controls

823 lines (727 loc) · 58.4 KB

Regularization Policy

A regularization redraws a packing the record already holds so that it reads cleanly: near-axis squares straightened, contacts closed, sub-blocks aligned, symmetric where the packing allows it, and turned into a canonical pose by flip regularization. This policy governs every regularized drawing and every regularization the register records, whichever generator made it: this project’s regularize_axis_components/2, Evan Daniel’s regularizer as vendored here, or a reviewed adjustment of either. It says what a regularized view must satisfy, what the automated pass decides by itself, what a reviewer decides, and how each decision is recorded and credited.

Neighbouring documents own the rest. attribution.md owns credit and epistemics.md the rungs; the atlas README describes the layer and how to rebuild it; development.md places the gates; and the assessment of Evan Daniel’s process is the evidence behind the rules below.

1. Purpose and Scope

Regularization is for drawings. It never changes a side, a frontier value, an evidence tier or a credit under a drawing’s name. A drawing made from a regularized view carries the record’s tier from the source witness, never the view’s own verification, and the view never replaces the witness.

A run produces two kinds of output, and they take different paths:

Output What it is Where it lives Path in
Regularized view An exact packing of the credited arrangement at the side the record already states packing/atlas/known-best/regularized/ and its index.json; the atlas and site drawings made from it The automated pass (section 4), then review (section 5)
Register result An exact certificate whose side is below the case’s verified upper bound: a new verified upper bound, however it was made An entry in results.yaml with one regularizations item per count The import runbook’s record and validate stages, as T-146 took

A certificate that proves more than the record holds is registered as a result first, and its drawing enters the layer only after that. Once validated, a register result below the case’s house best packing replaces it, by the import runbook’s house rule.

2. Definitions

Term Meaning
Source witness The atlas witness packing/witnesses/known-best/n-NNN.yaml a view is derived from
$s_W$ The source witness’s side: its exact side, a rational or an element of the number field it is stated in, or for a decimal witness the printed side of its stated rational promotion, the 36-digit pose at dilation 1
$U$ The case’s verified upper bound when the view is decided
$s_V$ The view’s exact certified side
Generator The code that made a view: regularize_axis_components/2 in packing/devtools/regularize_axis_components.py, or Evan Daniel’s exact solver and regularizer vendored under vendor/evand-regularize/, driven by packing/devtools/regularize_evand.py and brought to the witness’s side by packing/devtools/regularize_same_side.py. A generator’s version is its own version, or for the vendored code the pin plus its ordered local patches.
same Side relation: $s_V = s_W$ exactly, in the witness’s frame or a recorded isometry of it
bound Side relation: $s_V = U$ exactly, where that is not $s_W$. The view certifies exactly the side the register already verifies, so it is a drawing frame (owner, 12 October 2026); at $n = 206$, $U$ is the side of Evan Daniel’s exact solve of the same packing (T-098), and the regularized certificate reproduces it. A regularizations item, whose side knows same and solved, writes it solved: the view was solved to that side
solved Side relation: any other exact side, including Evan Daniel’s certificate, which scales the solved point by $1 + 10^{-20}$ and rounds the side up. Never written as same or bound.
Identity Whether the view is still the credited packing, measured by arrangement_gap in packing/devtools/check_half_angle_area.py between view and source witness under the best of the box’s eight symmetries: one_to_one, moved (beyond $10^{-3}$), max_move and max_turn_deg, recorded as identity: {basis: measured, measured_in, …} in the register’s terms
Non-regression No square’s contact count falls under the house rule (gap $2 \times 10^{-6}$) or the stage rule (gap $0.01$, half a degree), as census_atlas_contact_shades.witness_shades measures it
Free square A square with no load-bearing contact in the exact solve: a rattler, or a force-free square
Level Evan Daniel’s classification of a run, recorded as his: 0, nothing moved; 1, only free squares touched; 2, the same packing by his sampled path test; 3, that test blocked. It informs review and decides nothing.
Alternate An arrangement at the record’s side that this project does not draw as the credited packing

3. Admission Obligations

A view enters the drawing layer only if every obligation holds. They are checked in this order, and the first that fails is the refusal the index records.

  1. Exact validity. The retained view is decided valid by both maintained exact routes. Over $\mathbb{Q}$ they are sqpack.witness.exact_verify and the independent checker packing/devtools/check_rational_witness_independent.py. Over a real number field $\mathbb{Q}(\alpha)$, the field an algebraic house’s witness is stated in, they are sqpack.witness.exact_verify over sqpack.field and the independent checker packing/devtools/check_algebraic_witness_independent.py, which reads the view’s own text and shares no arithmetic, root isolation or geometry with the first (think-mbdw, 26 October 2026). A view converted from Evan Daniel’s format is also decided by the third route, check_half_angle_area.py. His own checkers may be replayed as evidence about the source and never decide. A view that two routes here cannot decide is not admitted. The index’s digest binds the verdict to the view’s bytes.
  2. Side relation. $s_V$ is computed exactly and the relation recorded. A drawing needs same or bound, and $s_V \ge U$. A view with $s_V < U$ is registered as a result first and drawn after (owner, 12 October 2026: “regularize and correct always”), and a solved view with $s_V > U$ is not a drawing frame (the decision table). Where the witness’s side is above $U$ and a view at the bound exists, that view is the one decided and drawn, and the same-side view is set aside with its measurements (owner, 13 October 2026: a valid view has “equal or smaller side”).
  3. Identity. Measured and recorded for every view, from every generator: the one-to-one pairing with the witness, each square turned by more than $10^{-6}$ (as a sine) that the run’s recorded angle changes do not name, and the largest move. A changed pose no longer stops a view (owner, 13 October 2026: “the pose doesn’t matter if it changes as long as it’s optimal and conforms with our regularization guidelines”): a view without the pairing, with an unnamed turn or with a move over $1/2$ is admitted when every other obligation holds, and the reason it is admitted with names the change. The move and the turns are measured on the pairing, so a view without one records only that it has none (one_to_one: false), with no move or turn to state (16 October 2026, a clarification of the owner’s answer of 13 October, first applied to Evan Daniel’s view at n = 176).
  4. Non-regression. Neither the house nor the stage total of shared sides falls, against the source witness or against the admitted view it would replace. A view in which some square’s count falls while neither total does goes to review (owner, 12 October 2026).
  5. Labelling. The index entry says label: regularized, the view’s id ends in -regularized, the drawing’s title and description say “regularized”, the container drawn is the view’s certified side, and the credit line of section 7 is carried.
  6. Provenance. The index entry names the generator and its version; the method and every parameter that differs from the generator’s defaults; the inputs by digest (the source witness and, for the vendored generator, the exact-solve input); the run receipt; the review record, if any; and the measurements of obligations 2 to 4. The index header cites this policy.

The move bound is a measurement, not a gate. Until 13 October 2026 a move over $1/2$ sent a view to review. The bound came from the fifteen views T-146 measured, where no view that paired one to one moved a square more than $0.496$ and each of the three that did not needed a move of at least $0.589$, and from the then layer’s largest move, $0.421$ at $n = 181$. It is still measured and named, and decides nothing.

4. The Automated Pass

What Runs

Stage Tool Default
1. Input regularize_evand.py writes each witness as the generator’s text input Full precision: a rational witness printed exactly, never reduced to binary64
2. Exact solve The vendored exact solver, through the driver 80 digits and a timeout per count; a count that will not certify is reported, never polished
3. Regularize The vendored regularizer, through the driver conservative, coincide placement, the source orientation
4. Import The driver converts each certificate to a rational Witness/v2 view Decided on three routes; identity and shades measured
5. Same side regularize_same_side.py re-solves the view’s tight contacts over $\mathbb{Q}$ at $s_W$ Wherever every rotation in the view is rational in the witness’s frame; the grid-slack generator re-solves over the witness’s own number field where its house is algebraic
6. Reference regularize_axis_components/2 Unchanged: the fallback, and the reference for non-regression
7. Classify The driver applies the table below Index entries, the review queue, the candidate list and the receipt

The generator runs on request, never as a gate (section 8). Every run writes a receipt: the generator version, the interpreter, each count’s status, sides and seconds, and the run’s cost. The regularization process gives the commands that run the generators, choose one view per count and admit it.

The Decision Table

A count can yield two outputs, each classified on its own: the vendored generator’s certificate at its own side, and, where stage 5 succeeds, a same-side view at $s_W$. The first row that matches decides.

Row Measured outcome Decision Index record
1 No output: no valid certificate, an infeasible start, a stall or a timeout Refuse; the current view or the witness stands refused, generator-failed, with the status
2 A route here refuses the output Refuse, and open a defect bead refused, invalid
3 $s_V < U$, under any relation Register first: a result through the import runbook, in batches, with its identity measured for its regularizations item; drawn once registered None; the receipt lists it
4 solved, $s_V > U$ Not a drawing frame None, dilated-side
5 same or bound, and the view’s pose equals the witness’s exactly No view unchanged
6 Retired 13 October 2026: a view not one to one, or with a turn the recorded angle changes do not name, goes on to rows 7 to 10, the change named None None
7 same or bound, a square’s house or stage count falls against the witness or the view it would replace Refuse when a house or stage total falls; review when neither does refused, regression, or the review queue
8 same or bound, variant fewest with a merge of determined groups, or level 3 Alternate: review only, never drawn as the packing Review queue
9 Retired 13 October 2026: a move over $1/2$ goes on to row 10, named None None
10 same or bound, no regression, conservative, level 2 or below Admit, naming any change of pose regularized

Row 10 is the only automatic admission. Rows 6 and 7 were refusals until the owner’s answers of 12 October 2026: a case the rules cannot settle goes to review, and only an exact route’s refusal or a fall in a total refuses. Rows 6 and 9 were retired by the owner’s answers of 13 October 2026: the pose may change, so only rows 7 and 8 still send a view to a person. Views from regularize_axis_components/2 are classified by the same table; having no variant or level, they meet row 10’s last two conditions by construction. A view in the review queue is not drawn until a review record admits it.

How a decision is reported (owner, 11 October 2026). “Refused” is kept for a view that would change the drawing and fails an obligation. A run reports every count in one of these words, in this order of precedence:

Reported as When Rows
no view The generator produced nothing to judge; the witness or the current view stands 1
no change The view would draw the same picture: no square moves beyond $10^{-3}$ or turns, and the angle classes and the house contact counts are the source’s. Whatever its side relation, nothing is drawn differently, so there is nothing to refuse; a side below $U$ is still listed as a register-result candidate in the receipt 3 to 7 and 10, when nothing visible changes; 5
refused The view changes the drawing and an exact route refuses it (2), or a house or stage total falls (7) 2, 7
held The view changes the drawing, passes every route, and waits on the record: a register result first (3) or a side the record does not hold (4); the reason is shown with it 3, 4
review or alternate A square loses a contact while no total falls (7), or as row 8 says 7, 8
admitted Row 10 10

This changes how decisions are worded, not what is drawn: only an admitted view, or a reviewed one, enters the layer.

Two kinds of admitted view are listed first in the gallery for the reviewer:

  • A tie-break of taste: the generator dropped a symmetry for a coincidence, fell back from its coincidence placement, or reported an optimum that is not unique.
  • Two candidates: both generators yield an admissible view. The vendored generator’s view is kept, and the project’s is recorded beside it as the reference.

5. Manual Review and Adjustment

Who reviews: the owner, or a person the owner names in writing as the layer’s reviewer. An agent prepares the gallery and may draft a recommendation, and the review record names the person who decided: the rule No Blind Trust sets for the rungs, evidence from an agent and the decision from a person.

What the reviewer sees: a gallery the driver renders for each pass. For each count in the review queue or flagged, it draws the source witness and the view side by side with the house renderer at one scale and orientation, and gives the side relation and $s_V - s_W$; the identity measurements; each square’s house and stage counts before and after, with changed squares marked; the generator, variant, level and recorded angle changes; any symmetry imposed or dropped; and the other generator’s view, where there is one.

What a reviewer may decide:

Decision How it is made reproducible
Admit a view row 7 sent to review (a square’s count falls while no total does) A review record naming that very view: its count, generator and version, run, output and row, and the source witness as it identifies itself; layer then draws it, and the entry names the record
Admit a row 8 alternate Only once question 2 is answered yes, and then labelled an alternate, never the credited packing
Choose symmetry or contact, attached or separate sub-blocks, or a placement style A per-count generator parameter in the review record, which the driver reads on every later run
Choose an orientation None: flip regularization draws every default picture in its canonical pose, and the witness’s frame is never changed; the one isometry devtools/atlas_orientation.py records, at n = 211, is history
Choose between the two generators A per-count choice in the review record
Reject a view The review record; the previous view or the witness stands
Change how the generator works A reviewed patch (section 6)

What a reviewer may not do:

  • Edit a view’s coordinates or angles by hand. Every pose comes from a generator run and, for same, from the exact re-solve.
  • Admit a view that fails any obligation of section 3, or override a refusal or a hold of rows 1 to 7; row 7’s review, where no total falls, is the reviewer’s to decide.
  • Change a side, a tier or a credit.
  • Draw an alternate as the credited packing.

The review record: each decision is to be one record in packing/atlas/known-best/regularized/reviews.yaml, a hand-written input beside the index. A record holds the count, the date, the reviewer, the decision and its reason; any parameters or generator choice; the generator version and the source witness it was made against, as the witness identifies itself; and the gallery it was made from. The index entry names its record, and the records tier’s check of the layer (regularize_axis_components --check-atlas) holds the two to each other. The first review created both: on 16 October 2026 the owner admitted Evan Daniel’s same-side view at n = 211, which row 7 had sent to review, as recommended by the coordinating agent that asked. regularize_evand layer draws a view the table sent to review only as a record admits that very view: its count, generator and version, run, output and row, and its source witness as the witness identifies itself (its id, its source’s key, path and revision, and its side), never a digest of its bytes (OR-16). A review of a whole pass may also be written up under docs/project/reviews/ and mapped in the document map.

Re-review: a record lapses, and its count goes back through the automated pass, when

  • the source witness changes, by a new record at the count or a corrected witness;
  • the vendored generator is re-pinned or patched, or the project’s generator changes version;
  • this policy changes an obligation or a threshold; or
  • an owner answer to an open question bears on the count.

A lapsed adjustment is re-applied only if its parameters still produce a view that meets every obligation, and the reviewer confirms it in a new record. A lapsed view goes back through the pass by the regularization process.

6. Forking the Vendored Generator

Evan Daniel’s exact solver and regularizer, search/exact and search/regularize in evand/square-packing, are MIT-licensed and vendored under vendor/evand-regularize/ at the pin its manifest vendored.yaml records, with his licence notice kept. The project may patch them, on four conditions:

  • A patch is a recorded, reviewed local change. The manifest’s local_changes lists each one (the path, the kind of change, the date, the bead and the reason), the vendored README describes it under Local Changes, and the review that accepted it is the pull request that made it, so the unpatched pin can always be rebuilt. devtools.check_vendored_evand holds the tree to the manifest. An unrecorded edit under the vendored tree is a defect.
  • The version travels with the view. A patched generator’s version is the pin plus its ordered local changes, and every view and receipt it produces carries that version.
  • A patch changes how views are made, not what admits them. Section 3 applies unchanged, and every view a patch could change goes back through the pass with its review records lapsed.
  • Credit stays with the method. A view made by the vendored code, patched or not, or by a port of it, credits Evan Daniel’s method (section 7).

A patch that fixes a defect in the upstream code is reported to its author, as epistemics.md asks of every defect found here.

7. Credit

Whose work How it is credited
The packing’s finder Keeps the packing’s credit through the register result that holds it, which a regularization names as of. A regularized drawing keeps the finder’s citation line.
The regularization by: project, with method naming the generator and variant, for this project’s runs; another regularizer’s bibliography key for theirs, as T-146 credits Evan Daniel’s lists
Evan Daniel’s method Whenever his generator or a derivative made the view: method names it (“Evan Daniel’s regularizer, conservative, coincide”); the index and every drawing made from the view carry the line Regularized by the Squares Project with Evan Daniel’s regularizer (evand/square-packing, MIT); and a register result is credited Levy after Daniel through builds_on, which cites a bibliography key for the vendored code
This project’s tool by: project, method: regularize_axis_components/2, credited as the Squares Project
This project’s modifications to his regularizer Named beside his credit wherever his regularizer or the grid-slack port of his placement made the view, as modifications: the retry rungs its solve climbed (dps-100, force-kkt, the first-pass shim), the vendored copy’s patches its method names (the replica patch probe-flat-stall-stop), and the exact re-solve that brought it to the witness’s side (devtools.regularize_same_side); a grid-slack view is this project’s port of his coincidence placement (regularize_grid_slack), not his code. Where none applied the record says unmodified

The record states a regularization as the same packing in its own voice only where the identity basis is measured. A regularization taken in from elsewhere states its identity as its source’s claim until a measurement here decides it.

The owner asked on 13 October 2026: “If we’re using our own modifications to Evan Daniel’s regularizer, we should note that as well.” Each drawn view’s item in results.yaml’s regularized_drawings and its case record’s regularization note (question 16) name them, read by devtools.regularized_drawings from the view’s index entry alone: the rungs from run.rungs, or before the retry ladder from run.retried_with, and the patches from method. A view admitted after the ladder and before run.rungs existed has its rungs recorded from its run’s reports by python -m devtools.regularize_evand rungs, which writes them only where a report reproduces the drawn view byte for byte, never over rungs or a retry already recorded, and names the run that supplied them as run.rungs_from. A registered result’s own regularization (by: project in its regularizations, as T-155’s 59 are) names them too, written by the same tool from the item’s method and side: the patches and the same-side re-solve. An item records no solve, and T-155’s selection kept none while its runs’ reports are not retained, so its items say in modifications_unrecorded that whether the solve climbed a retry rung is not recorded, rather than reading as unmodified. Where the layer draws that very solve, the item reads the rungs from the drawn entry instead and records nothing as unrecorded, so the register and the drawing read one solve one way: the result registers the drawing itself, and its registration receipt (T-155’s selection.json) names, at that count, a run whose receipt the entry’s run block records, with the entry’s output and exact side. At fourteen counts T-155’s certificate is the full run’s view the layer draws, which ran no retry, so its item reads unmodified or names the same-side re-solve alone. At 87 the layer draws a later run’s view at the same side, which says nothing of the solve T-155 registered, so its item still says the rungs are not recorded.

8. Gates

The generator costs about 13.7 CPU-hours for $n = 1..324$ (49,302 CPU-seconds in the evand-regularize-full receipt, 3.8 hours of wall at four workers), far above any routine gate’s wall ceiling. So it runs on request with a receipt, and the gates hold the decisions it produced (OR-13, OR-17). An adoption is such a request: the intake pass that changes a count’s house best packing runs the generators on the new witness in the adoption’s pull request, by the import runbook’s house rule, which says how the runs fit the compute cap and which deferrals --check-state allows.

Check What it holds Where it runs
regularized atlas views match their index The index against the manifest, the witnesses and the views, by digest; with the layer’s contract, also the generator, method, side relation, identity and review-record fields; every drawn view’s exact side against $s_W$ and against the case record’s $U$, decided exactly, with the views above $U$ named in ABOVE_U Records tier, every pull request
regularized atlas drawings match their index Every drawing re-rendered from its retained view Records tier, every pull request
default pictures improve on their house drawings Every drawn view measured against its house drawing again, each count’s default picture decided under the improvement rule (section 11) and held to default-drawings.json Records tier, every pull request
every drawn picture is regularized, regular as published, decided or exempt on the owner's word Every count’s regularization state in default-drawings.json against the index; no count regularization-owed, no decision that still owes something, and no exemption without the owner’s recorded word or past its review date; no count pending unless default_drawings.PENDING_ALLOWED names it with its bead, and no named count that is no longer pending; and each house the change moves since its base regularized and flip-posed in it (section 12) Records and fast tiers, every pull request
results rungs are earned and the view agrees devtools.check_results on each register result’s regularizations items, and on regularized_drawings and every case record’s regularization note against the layer’s index Records tier, every pull request
regularized atlas views re-derive exactly Every regularize_axis_components/2 view re-derived byte for byte; every other generator’s view decided again over $\mathbb{Q}$ on the maintained routes, its square count, identity and shades re-measured and its digests held Deferred: the regularized-views job, whose 450 s ceiling bead think-uomm revisits
The generator run The views reproduced from the pinned code and inputs On request only, with a receipt

9. Owner Decisions

The assessment’s eight open questions, the four the pilots of 11 October 2026 raised, the three the owner settled on 13 October 2026 after the stack’s re-posed views were audited, and the owner’s rule of the same day on credit and results, each with its answer or, where none is recorded, the default the policy applies until there is one.

  1. Variant: conservative only; fewest is not run. Answered by the owner on 12 October 2026: yes. At $n = 129$ fewest lost 26 of 79 four-side squares and the one-to-one pairing.

  2. A fewest merge of determined groups as the same packing: no. Such views are alternates (row 8) and are not admitted.

  3. Exact views below the verified bound: answered by the owner on 12 October 2026, “regularize and correct always”: every such view is registered as a result through the import runbook, in batches with a regularizations item per count, and drawn once registered (row 3). The first batch is the exact packings at the decimal witnesses’ printed sides, about $10^{-14}$ below bounds rounded up at fourteen decimals.

  4. Irrational optima: a dilated certificate is not a drawing frame (row 4). A witness stated over a real number field is regularized in that field: since 26 October 2026 the grid-slack generator starts from such a witness’s own exact pose, re-solves the placement’s contacts over the field at $s_W$, and writes the view in the witness’s field, where the two number-field routes of obligation 1 decide it (think-mbdw). The owner, 26 October 2026, on seeing algebraic houses drawn unregularized on the home page: “When something lands on our main rendered lists, the regularization should be done. There’s no excuse not to, because it should be cheap to calculate. We should do flipper regularization as well as the general regularization, and that process should be well defined.” $n = 11$, whose proved optimum lies in a field of degree 8, is drawn from its witness, which reads regular as published.

  5. Orientation: every witness and view keeps its source’s frame; any isometry of a frame is recorded as atlas_orientation records one. Answered by the owner on 23 October 2026: the default picture is drawn in the canonical pose that flip regularization chooses, and the house drawing as published keeps its source’s pose.

  6. Tie-breaks of taste: the generator’s defaults, flagged in the gallery for the reviewer.

  7. Method credit: the owner directed on 11 October 2026 that the vendored code be properly credited, so its notice and the vendored README carry Evan Daniel’s credit. Until the owner says where else, the index, the drawings and a register result’s builds_on carry it as section 7 states. Since the regularized view became the default picture (section 12), the composites carry it too, and so does the case page that draws one; the atlas README names it in its account of the layer. Both composites print the owner’s approved two-line wording:

    Drawn packings are regularized for cleaner display using Squares Project rules,
    based on Daniel’s regularizer
    

    Their SVG metadata retains the full generator credits in data-generator-credits and the depicted regularized cases in data-regularized-cases. The compact printed credit does not replace that attribution.

  8. Retention of the solve: kept. The solved points and contact structures go in a pinned packet or release asset outside Git (OR-18), named by digest in the receipt: they can be discarded later, and a discard cannot be undone.

  9. A view at exactly the verified bound: answered by the owner on 12 October 2026: admitted, as the side relation bound (rows 5 to 10). It certifies exactly what the register holds.

  10. Cases the rules cannot settle: answered by the owner on 12 October 2026: they go to review, not to refusal. Row 7 sends a view to review unless a total of shared sides falls, as row 6 did until question 14 retired it; only an exact route’s refusal or a falling total refuses.

  11. A view that changes nothing visible: answered by the owner on 12 October 2026: as now. It is reported as no change and no file is written for it.

  12. Registered refinements the case records have not adopted: answered by the owner on 12 October 2026, as recommended: the register’s lower exact certificates (T-146 at its fifteen counts first) become the atlas witnesses before their counts are admitted, once the case-record rewrite (#403) and T-146’s custody item allow, and after the imports in flight at the same counts land. A view derived from a witness that later changes is re-reviewed (section 5). The intake’s #518, merged on 12 October 2026, adopted these houses on the case records directly, citing an owner decision of 11 October 2026 not to wait for #403; #559 regularizes its 42 counts again on the new witnesses.

  13. What a valid view is: answered by the owner on 13 October 2026: “valid means a valid packing with equal or smaller side”, “with exact verification”, where the side is “optimal as best as known”. A view is valid when both maintained exact routes accept it and its side is at most $U$. A same-side view above $U$ gives way to a view at the bound where the solve reached it (obligation 2), and the grid-slack generator tries a rational $U$ first wherever the witness’s side is above it. A view at a side below $U$ is a new result and is registered, as question 3 answered. Four drawn views sit above $U$ and stay, by name, until a view at $U$ draws as well: at 18, 26, 66 and 85 Evan Daniel’s views are rationals about $10^{-30}$ above a closed-form optimum, the status quo of question 4. Three more did until their houses moved: at 110, 241 and 297 the drawing was at Francisco Couzo’s decimal side, $3 \times 10^{-12}$ to $1.3 \times 10^{-11}$ above Evan Daniel’s exact optimum, until 241 moved to T-182 on 15 October 2026 and 110 and 297 to T-174 on 17 October 2026. packing/tests/test_regularize_select.py reads every drawn view’s side against $U$ and fails on a view above it at any other count, and on a listed count once it is drawn at $U$.

  14. A changed pose: answered by the owner on 13 October 2026: “once things have been verified correct, that they are equal or better, we can commit and use them. the pose doesn’t matter if it changes as long as it’s optimal and conforms with our regularization guidelines.” Rows 6 and 9 are retired: no pairing, an unnamed turn or a move over $1/2$ is named in the admission’s reason and decides nothing. Rows 7 and 8 stand, so a per-square loss of contact still goes to review. “Equal or better” is the choice rule of section 10, item 4, now an answer, and since the owner’s further answer of the same day it is the improvement rule of section 11, which is stated there and nowhere else: a view replaces the drawing only when it improves on it. Of the rule’s seven measures, classes and hues joined after a review of #559 found a view at 236 that tied the other five and drew one more hue; exactness counts as a goal because an exactly verified view at a side no larger than $U$ replaces a numerically checked witness (question 15). devtools.regularize_select applies the rule, and the regularization process says how a pass runs. The grid-slack generator’s straightening threshold is a per-run setting: a view from a run at 1 or 5 degrees is admitted on the same terms as one from the default half degree.

  15. Exact verification of every packing: answered by the owner on 13 October 2026: “Every packing, even after regularization, should be exactly verified.” Every view the layer draws is decided by both exact routes, at admission and again by --verify-atlas. The packings drawn without a view, and the decimal witnesses the case pages draw, are not; the verification gaps review of 13 October 2026 records them as a serious gap to be filled, as the owner asked.

  16. Credit and results: answered by the owner on 13 October 2026: “can we review how we’re doing credits? It seems like we should put a note on our frontier records about which ones have done regularization and if the regularization is due to us (squares project)”, and “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.” A pure regularization, one whose exact side is not below $U$ and which so proves no new bound, redraws a registered packing and is never a result: it changes no side, frontier value, evidence tier or credit, and no register result is made for one. It is credited in results.yaml’s regularized_drawings and on the case record of every count the atlas draws regularized. The record’s generated regularization field says who regularized the drawing (by: project, with the generator, its credit line and this project’s modifications, section 7), at which side relation (same or bound), which registered result holds the packing (of, or that none does), and the view’s path. Where the drawn regularization is itself registered, as T-155’s are at fifteen drawn counts, the field names that result (registered_in) and the note says the exact side is registered in it; the drawing adds no further result. T-155 registers 59 exact sides: 26 are new and prove new upper bounds, and 33 equal the sides of Evan Daniel’s earlier certificates (T-101), which the record had carried rounded up and which T-101’s replay already proves. At those 33, eight of them drawn, the item names T-101 as first_certified_in, and the note says that the regularization is the Squares Project’s, that its exact side is registered in T-155, and that the same side was first certified by Evan Daniel’s certificate (T-101), so it proves no new bound. Where a result registers this project’s own regularization of a count that the atlas does not draw, as T-155 does at 43 counts, or draws by another method, as at 227, the field carries it too, after any drawing: in the register’s terms (method, side, of, registered_in) with the generator its method names, its credit line and its modifications, and the note credits the Squares Project and names the result, so every record whose count this project regularized says so whether or not the view is drawn. Where the house is itself a regularization its source published, as T-146’s certificates are, the field credits the source first, in the register’s terms (by its bibliography key, of the result whose packing it regularizes, registered_in the result that registered it). devtools.regularized_drawings writes it from the layer’s index, its check in the records tier holds every case record to the index, and the site’s case page prints its note. An exactly verified side below $U$ is always registered as a result however it was made (row 3, question 3), as T-155 registered the exact views below the bound. It is a new upper bound only where no earlier certificate already stated that side exactly and no larger count’s verified upper bound is at or below it, since $s(n) \le s(m)$ for $m > n$. Otherwise the result says that its bound is not new and names the result that already proves it: at the 33 of T-155’s counts where Evan Daniel’s certificate stated the side first, the item names T-101 as first_certified_in, and T-156, which registered the grid-slack view of $n = 290$ on 13 October 2026 at the witness’s own side 2.4e-15 below $U$, says that its bound follows by monotonicity from T-155’s certificate at $n = 291$, at the same side. The owner decided on 13 October 2026 to keep T-156, to say so, and to keep the grid-slack view as the certificate of $n = 290$’s own packing at that side (#587); that view draws no change, so the atlas draws nothing new there.

An answer, like any other change to this policy, is an owner decision recorded here with its date. It replaces the default, and the counts it bears on are re-reviewed.

10. Implemented on the Owner’s Standing Instruction, for Confirmation

On 12 October 2026 the owner asked for the work to run “autonomously end-to-end overnight to make sure that we have regularized versions of everything by the time we’re done and everything is PR’d and ready to merge”, with “all versions that are regularized … properly captured in our frontier records”. The stack implements the choices below under that instruction. None is an answer to a question of section 9 until the owner confirms it; each is recorded where it is implemented, and reversing one is a change to that layer.

  1. Rotation matching: a regularized view’s rotation matches the witness’s within a sine of $10^{-9}$, the tolerance of #511’s re-solve, a thousand times below the turn threshold of obligation 3 (#511).
  2. Retrying a solve that gives no view, once per rung and each attempt listed in the receipt: a failed solve at 100 digits; a witness not jammed only at noise level with upstream’s --force-kkt, kept only where the side falls by at most $10^{-10}$; a converged solve whose certificate upstream refused, or whose active-set step failed after a converged first pass, passed on to the regularizer; a timeout solved again with a recorded replica patch that stops upstream’s zero-mode probe at its floor, applied outside vendor/ (#511). The views are decided by this project’s exact routes as always.
  3. A third generator: regularize_grid_slack, this project’s port of Evan Daniel’s coincidence placement, started from the witness’s exact frame (near-axis squares straightened under 0.5 degrees where a placement stays feasible), brought to the record’s side by the same exact re-solve, and credited to his method (section 7) (#530). Over a witness stated in a number field, the frame and the re-solve are in that field (think-mbdw, 26 October 2026).
  4. Choosing between generators: where both yield an admissible view, Evan Daniel’s is drawn unless the grid-slack view is better on a measure of the drawing (light, uneven or near-axis squares, or a house or stage total) and worse on none (#548). The owner confirmed this rule on 13 October 2026, with the measures section 9, question 14 lists.
  5. The layer’s contract and checks: every index entry names its generator; the records-tier check reads each non-axis view against its entry, and the full tier re-derives the axis tool’s views and decides every other view again on both exact routes (#548).
  6. The frontier record of a drawing: each drawn view’s regularization is recorded in the register’s terms in results.yaml’s generated regularized_drawings, of resolved to the registered result that holds the packing; a count whose packing no registered result holds is listed as unregistered (#548). A drawing proves no bound, so no register result is made for one. The owner confirmed this on 13 October 2026 (section 9, question 16), and each drawn count’s case record now carries the same credit in its generated regularization note.
  7. Registering exact views below the bound (question 3’s answer, carried out in #549): the row 3 views outside the counts #518 and #553 re-housed are registered as one batch result, T-155, with the case records’ exact forms moved to the registered sides and their printed values kept, so no citation weakens. Where a registered side is exactly the side of an earlier exact certificate, as it is of Evan Daniel’s (T-101) at 33 counts, that bound was already published, and the records cite the earlier certificate’s independent replay beside the batch’s.

11. The Improvement Rule

The owner, 13 October 2026: “tie is fine on exactness or regularity as long as it is making some improvement wrt our regularity or heuristic goals”, and “that should be codified in our regularization standards doc”. The rule below is that answer. It decides whether a view is drawn in place of today’s drawing (question 14), and whether a drawn view is a count’s default picture in place of its house drawing (section 12).

A view improves on a drawing when it is worse than the drawing on none of the goals and better on at least one. Ties on the other goals are fine, and a tie on every goal is no improvement.

  • light squares: fewer is better
  • uneven squares: fewer is better
  • near-axis squares: fewer is better
  • angle classes: fewer is better
  • hues: fewer is better
  • house total of full shared sides: more is better
  • stage total of full shared sides: more is better
  • exact side: smaller is better
  • exactness: exactly verified is better than numerically checked

The marked text is the rule’s one statement: two sentences and its goals, one line each, with the direction in which each is better. The first five goals are regularity as the house renderer’s colouring shows it, counted by devtools.census_drawing_regularity; the house and stage totals are the full shared sides of regularize_evand.contact_partners; the exact side is a larger one worse; and exactness is an exactly verified view against a numerically checked witness. devtools.improvement_rule reads the statement from this file and decides it, and its readers take it from there: devtools.regularize_select prints it in the rule it reports, and devtools.default_drawings prints it with its check. That check parses the goal lines and compares them with the goals and directions the code judges, both ways, so a goal added or dropped here, or a direction changed, fails it until the code agrees. Every view the layer draws is exactly verified, so exactness is never worse. A change to the rule is an owner decision recorded here, made to the statement and to the code together.

12. The Default Picture

Owner decision, 13 October 2026. Asked “why do we even have ‘house’ on the diagrams? are there any benefits other than the historic accuracy? I’d be inclined to make all diagrams just default to regularized ones”, the owner approved the plan “regularized drawing is the default where one exists; house kept as ‘as published’ on case pages and the toggle; posters redrawn”, adding that “having original records seems good but it seems like regularization should be a strict improvement if we’ve done it correctly”.

So a count’s default picture is its regularized view wherever the layer draws one that is valid and improves on the house drawing:

  1. both maintained exact routes accept it, as its index entry records;
  2. its exact side is at most $U$, or it is one of the views question 13 names above it;
  3. it improves on the house drawing under the rule of section 11, measured as the selection measures a view against today’s drawing.

Everywhere else the house drawing is the picture. That includes a drawn view that ties an exactly verified house drawing on every goal at its own side, whose picture differs from the house’s by little more than a square’s small slide, and at most counts by less than a pixel at tile size; the layer keeps such views, admitted before this rule, and draws none of them by default.

Where the default is the view, the homepage and atlas tiles, the frontier survey’s thumbnail, both posters and the case page draw it, and the case page keeps the house drawing one click away as as published, with the credit of the source whose packing it pictures. The owner’s follow-up on 13 October 2026 adds Regularized | Original to the homepage and atlas page. Each page starts with Regularized, using the defaults above and the house drawing where no view passes the improvement rule. Original shows the current house packing as published, with its source credit. Switching preserves the tiles and their links, and the drawing choice lasts for that page only; a new page starts with Regularized. Without scripts, the default drawings remain available. Where the layer holds a view that is not the default, the case record’s regularization note opens by saying that the house drawing is drawn as published, and why. The house witness stays the record’s witness, and the view changes no side, value, tier or credit (section 1). devtools.default_drawings writes each count’s default and its reason to packing/atlas/known-best/default-drawings.json; its check, in the records tier on every pull request, measures every view against its house drawing again, fails when the record or a default departs from the rule, and lists the counts that keep the house picture and why.

The owner asked on 15 October 2026 that the records make it evident which drawings are regularized. So the same record states, for every count, whether the picture it draws is regularized (a view of Evan Daniel’s regularizer or the grid-slack generator), regular-as-published (the house drawing has no uneven and no near-axis square, no strong view of the count has lapsed, and PENDING_ALLOWED does not hold it for its bead’s runs), decided-no-better (the strong generators ran on that witness and drew nothing the improvement rule takes), exempt (on the owner’s word alone, below) or pending, with the witness it was decided on (its source’s key, path and revision and its side, never a digest of the file, under OR-16) and the runs that decided it.

The owner, 26 October 2026: “When something lands on our main rendered lists, the regularization should be done. There’s no excuse not to, because it should be cheap to calculate. We should do flipper regularization as well as the general regularization, and that process should be well defined.” The record had a sixth state, regularization-owed, for a count whose runs drew no view the rule could judge, its regularization owed to an open bead. The seven exact algebraic houses rested there from 23 to 26 October, and n = 259 and 302 were drawn unregularized on the homepage, which the owner called a serious process failure that must never recur. So no count rests owed on main. The state’s name stays in the record’s totals at 0, and the records tier fails a count in it and a SETTLED decision that still owes runs or an outcome. A count whose witness no strong generator can draw waits for that capability in the change that moves its house, or carries the owner’s exemption. An exemption is the owner’s alone: its entry carries the owner’s words exactly, the day they were given and the repository file that records them, which must hold them, with the reason, the bead that will settle the count, and a review date at most 14 days after that word. The records tier fails an exemption without the owner’s word or past its review date, and the intake sweep lists each one for the owner. The four decimal Kingbird witnesses at 39, 55, 290 and 291, set aside by an agent on 15 October 2026 without a strong run, stood without the owner’s word until 9 November 2026, to fail after it. Both strong generators settled all four on 31 October 2026 (think-yk2h): 39 is regularized, and at 55, 290 and 291 every view ties the house drawing on every measure and is better only on exactness, row 10 reports no change, and the count is decided no better. No count is exempt, and default_drawings.UNRATIFIED is empty (the runs’ evidence; whether exactness alone should earn the picture is the owner’s question, think-j68m).

A change that moves a house regularizes it in the same change, generally and by flip (section 13): against the record of the commit it is built on, each count whose witness has changed must land regularized, regular as published or decided no better on the new witness, or exempt on the owner’s word, or pending only as a corrected witness default_drawings.PENDING_ALLOWED names; and its flip pose, decided again from its picture, must be the one the flip record holds. The import runbook’s Adopt a House, Then Regularize It gives the step: the tools, the receipts kept, the states allowed and what a reviewer checks.

The index’s status: regularized says only that the layer holds a view. A view that lapses with its witness is recorded on the axis tool’s record (lapsed), and the count is pending until the generators decide the new witness. So is a count that default_drawings.PENDING_ALLOWED names, whatever a new house reads: the list’s first entries are the record of strong views that lapsed before the axis record kept that block. The records tier fails on a pending count unless default_drawings.PENDING_ALLOWED names it with its bead, and on a named count that is no longer pending, so the list equals the pending set, and bead tree holds each bead it names open. Since the list also holds its counts pending, it changes only by reviewed changes: a count leaves it when the strong generators or a SETTLED decision decide it, or when a reviewed change drops its entry, which the gate accepts where the count then reads regular as published; and a count joins it with its bead, whatever its house reads. The gate cannot tell a dropped entry from a decided one, so the pull request that drops one shows its bead’s runs or what the count draws before and after.

13. Flip Regularization

Owner decisions, 23 October 2026. Packings arrive in their sources’ poses, rotated or mirrored against one another, “so they look a little chaotic”. The owner asked that each drawing be rotated and flipped by quadrant weight, “first rotate to put the heaviest quadrant, and then flip if necessary”, with ties broken in “some other arbitrary way to make it deterministic”, in floating point, “since this is only for the purposes of regularization”. The original source is not adjusted, the regularized picture is “fully canonical pose”, and the pose is kept as “a separate piece of metadata” rather than as a second copy of the packing. The owner then put the heaviest quadrant at the lower right, so that “the really well-known ones should not change, like n = 11”, and the rule matches “what the conventions are for the common ones as much as possible”; and asked that the stage be “completely separate from the other regularization” and fast.

The owner named the stage flip regularization; it is the last of the regularization process. It turns every count’s default picture (section 12) into a canonical pose: one of the container’s eight symmetries, chosen from where its squares lie.

  1. Split the container down the middle into four quadrants, and weigh each by the area of the squares inside it over its own area, in binary64.
  2. Rotate the heaviest quadrant to the lower right. Then, if the upper right is heavier than the lower left, mirror across the diagonal through the lower-right corner, the one flip that keeps that quadrant in place. The quadrant opposite the heaviest ends upper left, and the upper right takes what is left, so empty space gathers at the upper right, as the catalogues draw the integer grids: rows filled from the bottom, the empty cells at the top right. Equivalently, of the eight poses keep those whose weights, read lower right, lower left, upper left, upper right, are lexicographically greatest, weights within $10^{-9}$ counting as equal.
  3. Where poses tie, compare the same weights inside each quadrant, and inside each of those, on grids of 4×4, 8×8 and 16×16 cells taken quadrant by quadrant in the same order; then the squares themselves, each as its centre and its tilt, sorted from the bottom row and the right, the least list kept.
  4. After each of those stages, the quadrants included, the poses still tied are matched square for square. Once they all draw the same picture, the packing having that symmetry, the comparison stops: that stage decided the pose, and the first of them in a fixed order is kept, the identity first and the rotations before the flips, so a symmetric packing keeps its source’s pose. Poses that draw different pictures after the last stage stop the stage with an error, rather than a symmetry being assumed.

On the 324 default pictures of 23 October 2026 the rule keeps the published pose at 252 counts, n = 11 among them, and turns 72; no rule that puts the heaviest quadrant lower left keeps more than 195. No packing catalogue states an orientation convention, as the plan records, and the turned counts are mostly larger tilted packings whose sources publish them in no common pose.

The pose is kept as metadata, in its own record, packing/atlas/known-best/flip-regularization.json: for each count, the picture drawn (the view where a view is drawn, the house elsewhere) and its retained rendering; rotate, 0, 90, 180 or 270 degrees counter-clockwise, and flip, whether the mirror across the upper-left to lower-right diagonal follows; the stage that decided it; and the turned picture’s quadrant weights. python -m devtools.flip_regularization --update decides every count from that rendering’s coordinates in a few seconds, and --check, the records-tier step flip regularization record is current, decides them again. Every surface that draws the default picture applies the pose when it draws: both composites, the homepage and atlas tiles in Regularized mode, the frontier survey’s thumbnails, the drawing that heads a case page, and the social card. No witness, view, certificate, side, tier or credit changes, and no turned copy of a packing is stored. The house drawing as published, in Original mode and under a case page’s As published, keeps its source’s pose; the recorded pose, or its inverse, can be applied to it later with nothing new stored.