Skip to content

Latest commit

 

History

History
788 lines (708 loc) · 50.8 KB

README.md

File metadata and controls

788 lines (708 loc) · 50.8 KB

Frontier: What Is Known About s(n), Case by Case

$s(n)$ is the side of the smallest square holding $n$ non-overlapping unit squares. This folder is the structured claim register for each $n \le 324$. Each case keeps the best bounds reported by the named public sources separate from the strongest bounds supported by formal evidence. It also records numerical checks, verification origin, conflicts, blockers, review date, and links into the local literature archive.

Start with STATUS.md. Its one row per $n$ shows reported and verified upper and lower bounds, formal status, verification origin, conflicts, and freshness. Follow a case link for the evidence and the reason any formal bound trails its report.

Verified means formal here. A finite calculation is numerically checked even at hundreds of digits or tolerance 1e-100. An exact or interval-verified feasible witness proves an upper bound, not global optimality. A case is proved only when its verified upper and lower bounds coincide exactly.

Rendered Cases

The frontier and the figure atlas meet through ../atlas/rendering/manifest.json. Each manifest entry binds an $n$, evidence tier, view level, alt text, retained SVG, matching frontier file, and exact regeneration command. That keeps a figure discoverable without making presentation data authoritative over a case’s mathematical frontmatter. The complete $n = 1\ldots324$ visual index lives in the known-best atlas; specialized examples and diagnostic views live in the renderer gallery.

From the exploration root, list or rebuild every discoverable example with:

uv run --frozen --all-extras --group dev python -m devtools.render_packing_gallery --list
uv run --frozen --all-extras --group dev python -m devtools.render_packing_gallery --update

Every currently rendered case embeds its own artifact below the frontmatter. The deterministic SVG gate checks those links, the gallery documentation, the manifest, and the regenerated bytes together.

Structure Here, Tables There, Generated in Between

The research documents keep their tables—a reader should get the whole picture of the problem from the report alone, without opening this folder. This folder holds the same facts in a form a tool can read, so they can be queried, plotted, and checked.

The duplication is deliberate and it cannot drift, because the Markdown tables in the reports are generated from these files by devtools.render_research_tables, between <!-- BEGIN GENERATED: … --> markers, and packing-validate fails if any is stale:

uv run --frozen python -m devtools.render_research_tables
uv run --frozen python -m devtools.render_research_tables --check

Five reader views are generated this way: STATUS.md, the open frontier (261 rows), the solved cases (63), and the search and proof strategy catalogues (28 and 30). Editing a fact means editing the data here and re-rendering; editing a generated table by hand will be caught.

Generating the frontier table from structured data caught a headline error in its first hand-written version (see Smallest Open Gaps below).

Each n-NNN.md is a softschema artifact: YAML frontmatter carrying the structured payload, validated against square-packing-case.schema.yaml, followed by a Markdown body that is purely reader-facing.

The frontmatter is authoritative. A consumer reads the YAML and must not parse the body prose for structured values. The body holds judgement, history, and caveats that do not fit structured scalar fields.

Source Datasets and Their Schemas

Each source dataset declares a contract, points at a compiled JSON Schema, and validates at status: enforced.

File(s) Profile Schema
n-001.md … n-324.md frontmatter-md square-packing-case.schema.yaml
search-strategies.yaml, proof-strategies.yaml pure-yaml strategy-catalogue.schema.yaml
asymptotic-waste-bounds.yaml pure-yaml asymptotic-waste-bounds.schema.yaml
source-availability.yaml pure-yaml source-availability.schema.yaml
evidence.yaml pure-yaml frontier-evidence.schema.yaml
source-coverage.yaml pure-yaml source-coverage.schema.yaml
results.yaml pure-yaml results.schema.yaml
covering-values.yaml pure-yaml covering-values.schema.yaml
uv run --frozen python -m devtools.validate_schemas

packing-validate runs that check, so an artifact cannot drift from its schema unnoticed. The loader also rejects duplicate YAML mapping keys instead of silently keeping the last value; this protects the evidence field the author actually wrote, not a formatting preference.

A caveat about how the pure-YAML files are validated. softschema’s spec defines the pure-yaml profile, and its library implements it correctly—but as of 0.6.1 the CLI reads every artifact with the frontmatter reader and builds its contract without a profile, so softschema validate rejects a conforming pure-YAML file with no_frontmatter. Filed upstream as jlevy/softschema#38, with a reduction showing the library returns valid for the same file when the profile is set explicitly.

Until that lands, devtools.validate_schemas applies the CLI’s compiled schema directly to the pure-YAML datasets, so status: enforced reflects an executed check. Negative controls reject wrong fields, invalid enumerations, inconsistent counts, known-invalid exponents, and case statuses outside {proved, open}.

The Exact-Side Register

exact-side-register.json is a generated view, not a source. Each entry is one exact side fact the record stands behind:

  • its degree;
  • its minimal polynomial, inline in exact_sides.polynomial_text notation or held by a digest-pinned asset with the facts measured from it;
  • its interval;
  • who states it;
  • whether it is about the house, a staged house or a named configuration, such as T-246’s s(29).

devtools.exact_side_register derives it from the case records, select_houses, the intake ledgers and the retained packets. On every run it decides each inline polynomial irreducible and its root isolated. --check, the records-tier step exact-side register is current, fails when the register differs from its inputs, so a change that moves a house runs --update. Case records keep what adopt_houses writes; the register writes nothing on them. The plan and its later slices are in plan-2026-10-25-exact-side-register.md.

What Is in a Case

The payload fields are documented in the schema. The ones that carry the most weight:

  • reported_upper_bound and reported_lower_bound preserve the strongest literal claims in the named source set. They do not endorse those claims.
  • verified_upper_bound and verified_lower_bound contain only formally supported bounds. They are a ceiling and a floor, not the value of $s(n)$. The certified ceiling may be a weaker rational construction or the exact grid when the tighter public pose is only numerical. A gap exceeding the reported side’s declared precision carries a mathematics blocker in the case record; the generated status table lists those gaps. A rational certificate can lie just above the source’s printed side, as Evan Daniel’s catalogue certificates do (T-101), or reach one exact side in both lanes, as his refinements of the earlier packings do (T-098). bounds_agree_at_declared_precision accepts a ceiling within one unit of the printed side’s last place as the same displayed bound. An exact_form on the ceiling is the exact form of the ceiling; $s(n)$ is known exactly only when status is proved. An algebraic ceiling the house writer writes is root(p, low, high): $p$ its minimal polynomial in $s$, highest degree first, as minimal_polynomial writes it, and low and high the side cut down and rounded up at forty places, an interval in which it is $p$’s one root (exact_sides.form, which exact_sides.read_form reads back exactly). A side of degree two is written in its closed form instead, (a+b*sqrt(d))/c with $d$ square-free and the fraction in lowest terms, as n = 51’s (16+5*sqrt(2))/3 and n = 295’s (22+22*sqrt(2))/3 are. n = 11’s root(P_trump11, …), which names its polynomial, is not this form.
  • Both upper lanes hold the certificate of the case’s house best packing, as the synopsis defines it. By the import runbook’s house rule, the pass that validates a smaller packing, whoever found it, adopts it as the house best packing and regularizes it in the same pull request. Where the house has an algebraic side and does not read regular as published, its count draws its exact house and its regularization stays owed to a bead (regularization-owed), since no regularizer yet draws an exact view of a number-field witness (the house rule’s algebraic bullet). devtools.select_houses derives it from the register and devtools.adopt_houses writes it.
  • reported_status is the source-set view; status is proved only when the verified bounds match exactly. There are currently 81 proved and 243 open formal cases.
  • reported_upper_bound.catalogue_rigid transcribes what the catalogue says about rigidity, in the catalogue’s own three-valued vocabulary: rigid, semi-rigid, or not-stated. not-stated is source silence and never a claim that a packing can move. The case-level rigidity field is separate and is this repository’s own finding, with an assurance level, a stated scope, and replayable evidence; null there means not assessed. The tutorial defines the related distinctions among contacts, stationarity, rigidity, and rattlers.
  • evidence points into evidence.yaml, where assurance, method, performer, independence, origin, actual precision, tolerance, certificate, replay, proof scope, and limitations remain separate fields. Common classification fields are required; method-specific fields appear only when they apply rather than being repeated as null placeholders.
  • conflicts preserve a failed replay or scope disagreement; blockers say whether the gap is source evidence, an importer, a checker, a field precondition, or mathematics.
  • resources—citation keys with paths into ../resources/ and a retrieved flag.
  • regularization, generated by devtools.regularized_drawings from the regularized layer’s index and never written by hand, is present where the atlas draws the count from a regularized view, where a registered result holds this project’s own regularization of the count (as T-155 does at 59), or where the house is itself a regularization its source published. Its note says who regularized the drawing, with which of this project’s modifications to Evan Daniel’s regularizer, and that a pure regularization, one that proves no new bound, is a credited part of the record but never a result; where the drawn regularization’s exact side is registered as a result, as at fifteen of T-155’s counts, the note says so (the regularization policy, question 16).

The assurance states are reported, numerically-checked, and verified. The method is separate: numerical-f64, numerical-multiprecision, interval-certified, exact-algebraic, published-proof, proof-audited, or proof-assistant-checked. Reader views qualify verified evidence as external, independently external, replayed here, or audited here. Replaying the source author’s own numerical program does not become an independent verification.

Source Coverage and Freshness

source-coverage.yaml is a dated claim about a named source set, not a claim that a web search exhausted every publication. It records each source’s scope, review date, retained first-party material, evidence, replay disposition, in-horizon overrides, and relevant claims beyond $n = 324$. Every source dated from 22 August 2026 that a registered result or the evidence it cites names has an entry here, which carries its key or cites that evidence, papers and tools included; devtools.check_source_coverage refuses one without.

The current baseline reparses the retained Kingbird catalogue through $n = 324$, applies Francisco Couzo’s certified packings at 49 counts from $n = 68$ to $307$ (T-056), seven of them lowered again by his revision of 3 October 2026 (T-092), and Joost de Winter’s at $n = 211$ (T-057), then Evan Daniel’s exact optima of 47 of those packings and of de Winter’s $n = 126$ (T-098), and records the Schadt $n = 29$ repository as a superseded numerical witness. UnitSquare’s reports at $n = 68, 103, 105, 110, 131$, Griffin Casson’s 39 packings, each larger than Couzo’s at its count, Couzo’s own earlier sides at $n = 208, 209, 228, 263, 272, 303, 306$, and the printed sides of the 47 Couzo and de Winter packings Daniel solved exactly are listed as superseded reports, and so is UnitSquare’s at $n = 69$, larger than the catalogue’s later side for the same packing (T-088). Overrides may come from any retained source: each must beat the catalogue baseline, and every claim a source’s retained record makes — the UnitSquare release, a packet’s acquisition record, or a directory of exact certificates — is reparsed and accounted for exactly once, as selected, superseded or beyond the horizon. Claims beyond $n = 324$ are listed by count and source. A count has at most one current row; an earlier dated report that a later one beats stays listed as its source printed it, marked superseded and naming the source of the row that beats it, which must be dated no earlier and strictly better by exact comparison. Göbel’s classical constructions are earlier reports beyond $n = 324$ (the owner, 23 October 2026, #613), so a row above Göbel’s construction at its count also stays listed as its source printed it. It is marked above-classical-construction, states the construction $p + q\sqrt 2$ it is above, and holds no count. The standing side there is the construction, which has no row: in the walks, and from 23 October 2026 in every start, claim record and replay reading beyond the horizon (several reports at one count). Rows are tracked beyond $n = 324$ with no ceiling, each recorded as its source reported it (the owner, 23 October 2026, D1). devtools.beyond_horizon_replay holds every row beyond $n = 324$ to Göbel’s construction exactly, from the maintained module devtools.goebel_construction, which proves each side final by the area bound: a member at count $k$ packs $k$ unit squares, so its side is at least $\sqrt k$. A construction an index or table states with no certificate file of its own (a tiling or L-border recipe on a certificate) is a row once the maintained construction checker, devtools.construction_checker, decides it exactly from the recipe and its parent certificate has been replayed on both routes (D2). A two-strip L-packing (#799), Ryu’s Theorem 1.3 shape with its two strips chosen independently, has no parent: the maintained devtools.two_strip_checker decides it exactly from its parameters. A deletion bound is no such row: one square below the next count’s certificate is an earlier report (D6), and a deletion-made bound gets a row only where a report claims the count with a certificate file of its own, as #941’s 809 does (the owner’s answer of 17 November 2026, think-gmoj), so an index deletion bound has none, claimed or not; any bound deleting two or more squares, claimed or not, waits on the owner’s answer to whether it is an earlier report (think-kzme; several reports at one count). A certificate kept as hosted data is read through its hosted-data manifest (devtools.hosted_shards), while its receipts and rows stay in Git (OR-18). Nine rows of 8 to 13 October at $n = 369, 375$ and $376$ are so marked, and 369 and 376 have no current row. The check is local and deterministic; refreshing a public source is a dated W1 research survey, not a network operation hidden inside ordinary validation.

SQUISH’s ten-packing release and its later $n = 153$ supplement, published on 7 October 2026, carry the upper bounds at eleven counts (T-113 and T-114). Nate Chaoweeraprasit credits David Ellsworth’s records and tools and states human-directed Claude Opus 5.5 assistance. The certificates are retained as derived facts and were replayed here by two independently re-implemented exact-rational checks, with two accepted project AI reviews. Both bound lanes display each exact rational side rounded upward at sixteen decimals; both exact_form fields retain that original certified rational side. The source’s finite displays remain attributed metadata in the packet: four were one unit too low, five eight or nine units too high, and two agree with the safe display. Neither release claims optimality.

The catalogue was last captured on 2026-09-30. The capture of 2026-08-22 is kept beside it byte for byte, because the UnitSquare and certified-packet intakes read it and their paragraphs about the packing each replaced are drafted from it; devtools.diff_kingbird_catalogue --before REV compares two captures count by count. Where a newer capture prints a side below a record’s and taking it is an intake of its own — a result by others published on or after 22 August 2026, which the register must hold first — the count is listed under pending_catalogue_intake, with the side the record still reports, the side the current capture prints, and the capture the record transcribes. None is listed today; $n = 69, 83$ and $87$ were, until their results were registered on 2026-10-05 as T-088 and T-089. The check reconciles such a count against that earlier capture, requires its record to carry a source-evidence blocker naming pending_catalogue_intake and the newer side (which STATUS.md shows as “catalogue ahead, intake pending”), and fails the entry as soon as either capture or the record stops bearing it out, so an entry is removed in the commit that takes the side.

Each entry also names the open bead that owns the intake and the day it was recorded, and check_bead_tree fails one whose bead has closed: three counts held with no owner trailed the catalogue from 2026-09-30 until the owner noticed on 2026-10-05. An intake pass captures the catalogue into scratch and compares it with the retained capture; taking a new capture into the record remains the dated survey above.

uv run --frozen python -m devtools.check_source_coverage

The separate DS7 lower-bound audit, with its retained comparison report, checks every case through $n = 324$ against Green’s reported Theorems 9 and 10, usable exact Table 2 entries, and the already indexed $4.5705$ external report for $n = 17,18$. It uses exact rational enclosures for theorem comparisons and keeps Table 2’s three decimal-only rows outside that arithmetic. The $n = 21$ decimal remains opaque reported metadata; the malformed $n = 82\text{–}85$ table expression is excluded in favor of the separate Theorem 9 specialization. None of these source corrections changes a verified bound or supplies Green’s missing proof. This is an audit of those named sources, not every publication.

Adding or Reviewing a Result

A result from any source, this project’s parallel projects included, enters these records through the result import process, and epistemics.md → Results by Others is its policy. To ask for a registration, open an issue with the registration request form.

The process relies on these rules of the records:

  • The reported lane takes the literal public claim, with a typed evidence entry of assurance: reported whose novelty is previously-published, which is what the register’s coverage gate reads. Disagreement is preserved as a conflict or a typed blocker; the source’s claim is never edited to match a checker.
  • A formal source build is separate evidence. An inspected, retained complete producer build can support V under the formal-proof rule, recorded with origin: external and the producer’s passing status. Keep its build and axiom receipts, scoped statement audit, full coverage and trust assumptions. A later independent replay has its own confirming entry; source execution alone does not establish C3 or change the case’s verified bounds here.
  • The register entry is made when the result is imported, at its derived rung, with a kind, a headline, attribution and a next_rung naming the replay and review it waits on. Its date is attribution.published; established is this project’s own results’ date and the checker refuses it here. Its status is derived, and reads recorded until a read or a replay of it is on file.
  • The claim is short paragraphs: the statement, the certificate, how the source checked it, what was replayed here, and the credit with a link to the source. The revision and digests of the retained copy belong to the evidence entry and the packet, not to the claim or the case record’s prose; devtools.check_prose_ceremony holds both to that.
  • activity says who has the next move. Once a replay or review is under way here, or a question is with the authors, the entry says so, dated and linked to the bead or issue that shows it, and the line is removed when the work lands or the answer arrives.
  • regularizations says whose packing a regularized certificate is. Where the certificate an entry states rearranges a packing another registered result holds, the entry lists each count: the result of whose packing it regularizes, whose finder keeps the credit; the regularizer by, one of the entry’s source keys or project; the regularizer’s method; the side, same or solved to an exact side first; the changes the source reports; and what the identity rests on, the source’s claim or a measurement here, with what was measured and where (measured_in). devtools.check_results holds each item to the register, and the third exact route measures each certificate against the packing its item names.
  • Geometry is adapted once to Witness/v2, and then checked with packing-witness inspect or check under explicit arithmetic, precision, and tolerance.
  • The verified lane takes a value only after an exact proof, exact witness replay, or rigorous interval certificate discharges its assumptions. For a result by others, that is a complete replay here and a review of its mathematics under docs/project/reviews/. External evidence and a local replay are recorded as separate entries. The case-lane check refuses a machine run not made here, and devtools.check_results a verified bound on a result by others dated from 22 August 2026 that no register entry covering it lists an accepting review of the mathematics for (accepted or defects-resolved, by a kind other than oversight), and whose evidence records no defect-free read in a mapped review document. An upper bound that passes them below the case’s house best packing replaces it (the house rule).
  • Beyond the corpus, the report-order ledger decides the rows (several reports at one count), with the owner’s answers of 23 October 2026. A side below an earlier report by less than the 12-place print, at the same print, restates it and is no row; a side at or above Göbel’s classical construction at its count is above an earlier report; a certificate its author withdrew from an issue is not a report; and a side obtained by deleting one square from the next count’s published certificate is an earlier report, which on the D6 decision’s recommendation, not yet confirmed, has no row of its own. Rows are tracked with no ceiling: Göbel’s construction comes from the maintained module at every count, a recipe with no certificate file is a row once the maintained construction checker decides it from a replayed parent, and a hosted certificate is read through its manifest.

Registering a First-Party Result

For a result produced here, register the claim supported by the retained evidence. Use epistemics.md for its classification and New Result Publication for publishing it. T-026 is a worked example of a machine-certified lower bound; proofs, numerical findings, and structural results use the fields and rungs appropriate to their evidence.

Before registering a machine-confirmed result, run every required verification route on the retained certificate bytes and keep the commands, outcomes, and scope in the receipts. Record replay_status: passed only after that replay succeeds. Schema and register checks inspect the recorded support; they do not execute the certificate’s replay command.

  1. Take the next id and put the row last. devtools.check_results compares the register’s ids to T-001..T-N in file order, so an id is its row’s position and nothing else. Check the integrated register and known parallel assignments, agendas, transports, and open PRs before allocating an id, as conventions.md requires. Coordinate any pending rows so the integrated register remains contiguous; a new result is appended one past the current last row. Two branches that each add a row collide on merge, and the one landing second is renumbered—the only renumbering conventions.md allows, recorded as an annotation on the affected artifacts rather than made silently. T-032 was T-031 on its own branch until the other T-031 landed first.

  2. Write the evidence entries first, one per decision, in evidence.yaml. The evidence schema requires id, claim, scope, assurance, performed_by, relationship_to_generator, replay_status, limitations, and source_reviewed, plus the fields conditional on the method and assurance. An entry whose method runs code, or that names a replay, lists the programs that verified it in verifiers, by their ids in verifiers.yaml, and its relationship_to_generator says whether they are the producer’s own code or a re-implementation (epistemics.md). The rung is derived from them, so what each carries decides what the row may declare: a machine entry has method exact-algebraic or interval-certified, origin audited-here or replayed-here, certificate, replay, and replay_status: passed, and counts toward C3 only with all five and not as a sample (epistemics.md). assurance: verified needs a formal method and may not carry precision or tolerance. Choose novelty from the retained source search, independently of who produced the result. For apparently-novel, record novelty_basis: the corpus, search, narrow novel_object, and known gaps, dated by source_reviewed. Write local certificate paths relative to packing/ and replay as a command runnable from packing/; every path in results.yaml is repository-relative. Verify that the certificate resolves and the command actually runs: the evidence schema accepts nonempty strings, and check_rung_figures can skip unrecognized or unresolved certificates. T-026 cites five entries: the exact decision and the interval decision of the 1440-step certificate, the dilation-limit record that turns its shrink into the registered bound, and the 720-step pair retained as its control.

    Say whose work the result rests on. The result is credited Levy. Where it rests directly on another’s proof, method or tool, give the row a builds_on with the names to print after Levy after and the bibliography keys they come from, and name those keys as the source_key of the evidence entries that use them; the checker refuses a name the cited evidence does not support. A proof the result repairs or corrects is a basis; a packing whose property it proves is its subject and is not listed.

  3. Keep first-party certificate packages under cases/. T-026’s certificates, limit records, proof note, and verifier are all in ../cases/n11_threshold_certificate/, one package per certificate family, named by the evidence entries’ certificate and the row’s artifacts alike. That is convention, not a predicate: the checker reads only that a certificate is named, and E-side2-center-lower cites one under campaign/series/…/results/. The package is what readers are pointed at: the register row’s artifacts name it, and the self-contained verifiable-claim documents live there.

    A verifiable-claim document is a separate deliverable, not a registration step. T-025 and T-026 each carry one, and render_verifiable_claim writes both from one template — but as two hard-coded branches of one function, keyed to those two files, with the witness direction, the title, the surd and the proof note written in. A third claim document needs that branch parameterized and its own proof note, which is work to schedule rather than a box to tick. Registering without one is what T-033 did; the row then cites the argument’s existing proof note rather than inventing an artifact.

  4. Update affected cases. When the result improves a formal bound, the case’s verified_lower_bound or verified_upper_bound cites the evidence that directly establishes that bound, with value and exact_form, and its body must restate the new bound: devtools.check_case_prose reads the prose against the front matter. Controls may remain result-only: T-026 cites five entries, while the case bound at that rung cited the three that established it. At the historical T-033 rung, n-011.md cited its three establishing entries rather than its control leg; the current case bound cites Marc Leonard Overbeck’s replay evidence and retains the earlier entries as history. A numerical finding or a result about a restricted family does not by itself change an unconditional verified bound. Preserve earlier evidence and decisions.

  5. Write the row with what results.schema.yaml requires: id, kind, headline, established, claim, scope, verification, confirmation, significance, novelty, evidence, artifacts, and next_rung, plus controls at C3 or above. kind is the line after id, then headline and established. kind says what the result is, one of the result kinds: the checker reads a headline that opens with a relation on $s(n)$ as a lower bound (≥), an upper bound (≤) or optimality (=), and holds the cited evidence to the same kind. headline is the claim shortened for a table cell, at most 100 characters of inline Markdown with the mathematics in backticks: the claim’s relation exactly (≥ is not >), its exact form with its decimal or a truncation of it marked …, and its $n$ values. The checker refuses a number the claim does not state. established is the day the certificate or proof first passed here, read from the commit that landed it, not from the day of registration. claim is the one statement the rungs attach to, in full and with exact values where established; devtools.check_rung_figures checks recognized a/b = d.ddd expressions and supported mass, atom-count, and margin phrases against the local fractional certificates it resolves. It is a targeted consistency check, not a proof of every numerical statement or certificate format; review the remaining figures explicitly. significance is a score from 1 to 5 against the rubric, a rationale, the date scored, and by; it never gates. artifacts and controls are repository-relative paths that must exist, and controls proves only that a control is retained—that it is adversarial is the test suite’s business, not the checker’s inference from a filename. produced_by is optional and, when present, must name a recorded hypothesis, agenda cell, session, or experiment. next_rung names the next evidence-improving action or says why no independent one applies. Where it names a confirmation not yet run that would raise confirmation, list it also in unfinished_confirmations: what it is, the rung it would reach, the bead or issue that owns it in link, and, wherever a price is known, an estimated_cost in CPU-hours with its upper end, its basis and the day it was priced (epistemics.md). The field is optional; remove an item when its confirmation lands, since the checker refuses one at or below the result’s rung. A registration also dates last_reviewed at the top of both files.

  6. Derive the rung; do not choose it. verification and confirmation are declared in the row and re-derived from the cited entries by the checker, which refuses a declaration above what they support and a declaration below it that has no composition note. V0 and V2 are declared-only and require explanatory notes. Review evidence relevance, scope, and coverage of every premise before accepting the derived rung; the checker does not establish those relationships. Result-level novelty is also declared and reviewed, not derived from the entries.

    • C3 is a machine certificate replayed here, or a third party’s retained replay, with a passing status and a named control: machine-checked, review record pending. Distinct methods are an attribute, not a rung: T-026’s exact event-cell sweep (exact-algebraic) and interval branch and bound (interval-certified) decide one certificate two ways, which the composition note states and the register shows beside the rung; two implementations of one method are one method.
    • C4 is C3 plus the review record of epistemics.md → Review Records: two adversarial AI reviews by distinct reviewers, the latest accepting, and a human oversight record naming the person, their relation and what they checked, each a reviews entry whose path is mapped in document-map.yaml as a non-superseded review. That is a document predicate: the reviewer must establish that the review covers the complete stated claim and its premises, and a mapped review of another object satisfies the predicate but cannot justify the rung for this claim.
    • C5 is a kernel check rebuilt here from the repository at a pinned toolchain with its axiom receipt, an open_review pointer so anyone can review and replay it, and two human experts’ formalization reviews. V is earned the same way from evidence of any origin, with one expert at V5, and C never exceeds V from C2 up.
    • A compound or derived claim takes the minimum rung over its parts, which can sit below what the checker derives from the strongest entry: in T-036 an audited prose step holds a C3 reduction at C2. The composition note records that reading—the load-bearing parts and which one sets the minimum—and its presence is what lets the checker accept the lower declaration. T-026’s names the sweep, the interval decision, the containment lemma, and the density step.
  7. Regenerate and check, in the order New Result Publication gives: validate_schemas and check_results on the records, then render_results --update, render_evidence_inventory --update, render_research_tables and render_results_headline; never edit inside a generated view’s markers. The site’s overview renders the results table, the recent results and the frontier survey from the same records (render_overview). packing-validate --records runs those checks and the related record checks; it does not replace certificate replay or the full checkpoint. Complete the remaining publication steps too: regenerate affected atlas exports, reconcile the synopsis, and run the applicable validation tiers, including the full checkpoint before final review. Retain the checked source/base and state which publication surfaces were updated or checked current. Each apparently-novel or confirmed-novel result gets its row in RESULTS.md and the site’s results table from the register. A T-id may not appear in the README or the synopsis before its row exists, since the checker rejects unknown ids in that tier.

The Strategy Catalogues

search-strategies.yaml and proof-strategies.yaml carry two working inventories of the field and adjacent methods: 28 search strategies and 30 proof strategies or candidates. They are broad research maps, not claims of an exhaustive history; several entries are explicitly unused on this problem. Each entry has an id, name, mechanism, family, a status (outcome for search, status for proof), a note, and any citation refs.

Grouping the entries by family exposes the imbalance below:

search proof
entries 20 30
families 4 6
entries that have produced results here 11 17
largest working family constructive (9 of 20) unavoidable points (10 of 30)

Ten of the seventeen working proof strategies refine the same idea—place points, prove they are unavoidable, and count—while four different search families have produced records. The other ten-entry proof family, the transversal and wider packing-and-covering toolkit, is almost entirely unapplied to $s(n)$. This contrast suggests that the lower-bound side has underused methods.

What the Corpus Shows

Counts below are computed from the artifacts, not asserted.

Of the 243 open cases, 0 have Nagamochi’s formula as their verified lower bound, and 187 have Karakuş’s weaker general bound in its place (T-083). Until 2 October 2026 those 187 rested on Nagamochi’s formula; that day his Lemma 1 was found false (review, T-085), so his formula became a reported bound and each verified floor it held fell to the strongest registered bound below it: Karakuş’s general bound, Hiroaki Hosono’s registered point bounds carried by monotonicity at $n = 54, 57, 73$ (T-044), which the formula had masked, and Bašić and Slivková’s piercing bound at $n = 37, 61$ (T-087), registered on 3 October 2026. Replayed certificates recorded in parallel and merged on 3 October 2026 have since raised all five of those, and 46 more of the Karakuş floors, above them; they are counted below. Two others use certificates already integrated into the register: at $n = 12$ unnir’s (Numaro) extension of Evan Daniel’s points ($9929/2500$), and at $n = 17$ the current external certificate bound ($18641771/4000000$). Complete interval and exact replays of external certificates hold the other 54: Hosono’s mixed rectangle-measure certificates at $n = 37$, 50, 65, 66, 83 to 85, 87 and 90 to 92, and by monotonicity at $n = 51$ from $n = 50$, at $n = 86$ from $n = 85$, at $n = 88$ from $n = 87$ and at $n = 93$ from $n = 92$, and its rectangle-density certificates, directly or by monotonicity, at $n = 18$ to 20, 26 to 31, 38 to 43, 52 to 58, 67 to 75, 89 and 94, and its linear certificates at $n = 82$ and 101, and by monotonicity at 102 to 105. On 2026-10-02 those replays superseded this repository’s first-party bounds at $n = 18$, 19 and 20 ($4679/1000$, $24/5$ and $97/20$), Tokoharu’s at $n = 26$, 29 and 30, Hosono’s point certificates at $n = 39$ to 41, 52, 53, 55, 56 and 68 to 72, and Nagamochi’s formula, or the floor that replaced it, at 33 counts; on 2026-10-03 its afternoon mixed certificates did so at three more, $n = 83$, 91 and 96, its linear certificate for $n = 101$, replayed on 2 October and recorded on 3 October, at five, $n = 101$ to 105, and its linear $n = 82$ certificate, replayed on 3 October, at one more, where the floor it raised was Karakuş’s. $n = 32$ left the open cases on 2026-09-27, when a replayed external closed cover proved $s(32) = 6$, and $n = 21$ and $n = 45$ on 2026-09-29, when replayed external mixed covers proved $s(21) = 5$ and $s(45) = 7$; $n = 59$, 60 and 61 left them on 2026-10-02, when replayed mixed covers proved $s(59) = s(60) = s(61) = 8$, and $n = 77$ and 78 later that day, when the replayed $s(77)$ cover proved $s(77) = s(78) = 9$; and nine more, $n = 97$, 118, 141, 166, 193, 222, 253, 286 and 321, left them on 2026-10-03, when the replayed Valid7 and the built Lean reduction proved $s(k^2 - 3) = k$ for every $k \ge 6$ (T-064); and $n = 44$ left them on 2026-10-31, when a replayed closed mixed cover proved $s(44) = 7$ (T-252); and $n = 76$, 95 and 96 on 2026-11-02, when mutated-cover controls of zm_mixed.py completed the confirmation of Benjamin Greff’s replayed mixed covers, proving $s(76) = 9$ and $s(95) = s(96) = 10$ (T-184, T-191), the $s(95)$ cover’s total being below 96 as well. Within the original $n \le 100$ corpus, the corresponding Nagamochi count is 0, and 38 until 2 October 2026. The count is checked against the case records by devtools.check_nagamochi_bounds (D-430), because earlier hand-maintained counts outlived their case promotions.

Of the 243 open cases, 98 are still held by the trivial grid. The other 145 carry non-grid constructions. Within $n \le 100$, the 33 non-grid open cases comprise 13 hand-built, 11 from simulated annealing (ten of the eleven dated 2024–2026; $n = 53$ is Cantrell’s from 2002), 5 diagonal strips, 3 extensions of smaller records, and 1 whose method the source does not record ($n = 68$).

Recorded catalogue degrees for $n \le 100$ rise through 4, 5, 6, 8, 12, 18, 20, 38, 41, 42 and 672, while every proved case but $n = 11$, whose side has degree 8, has degree at most 2.

Smallest Open Gaps

Ranked by gap—the best-known packing minus the verified lower bound, which is what devtools.gap_ranking computes—the six smallest open cases at $n \le 100$ are below.

$n$ gap record note
17 0.0151 Bidwell carried to $\frac{18641771}{4000000}$ by Guzhou0806 after Overbeck
31 0.0225 grid $6^2 - 5$, carried to $\frac{2391}{400}$ by Hosono after Tokoharu and Levy (T-232)
94 0.0280 grid $10^2 - 6$, carried to $\frac{2493}{250}$ by Özden after Hosono, Tokoharu and Levy (T-266)
12 0.0284 grid $4^2 - 4$, carried to $\frac{9929}{2500}$ by unnir after Daniel, Ryu and Levy (T-274)
27 0.0521 Göbel carried to $\frac{1131}{200}$ by Hosono after Tokoharu and Levy (T-233)
19 0.0581 Wainwright $5^2 - 6$, carried to $\frac{1931}{400}$ by Hosono after Tokoharu and Levy (T-233)

The $n = 17$ bound is Guzhou0806 / N17 project’s R071, continuing Overbeck’s charge (Overbeck building on Squares Project (Joshua Levy), Mira and Guzhou0806): an exact weighted-certificate proof over 5,114 orientation intervals. A candidate proof that closes the gap at Bidwell’s side, T-153 of 12 October 2026, has been read here and partly replayed; until its complete replay passes, the verified gap stays as the table gives it. A second proof of the same equality by another route, Oscar Fitzharding’s Lean 4 proof T-154 of the same day, has been read here and not built, so it does not move the gap either. $n = 17$ and $n = 12$ moved in September 2026 on third-party weighted certificates, and the $k^2 - 4$ family now has three solved members above $k = 3$: $s(21) = 5$, $s(32) = 6$ and $s(45) = 7$, proved by Evan Daniel on the same method’s zero-margin form, the first and third with mass on the grid lines as well as on points. $n = 21$, in this table at $0.0050$ from 2026-09-27, left it with that proof. $n = 11$, which headed it at $0.0021$ once Overbeck carried it to $31/8$, left it on 2026-09-30, when T-060 proved $s(11)$ equal to Trump’s side, and $n = 61$ on 2026-10-02, when a replayed mixed cover proved $s(60) = 8$ and with it $s(61) = 8$; $n = 78$ left it later that day, when another proved $s(77) = 9$ and with it $s(78) = 9$. $n = 97$, at $0.0557$, left it on 2026-10-03, when Evan Daniel’s family $s(m^2 - 3) = m$ was proved here for every $m \ge 6$ (T-064): his exact checker of the finite premise Valid7 was replayed in full and his Lean reduction from it was built. $n = 31$ entered it second, at $0.0225$, when T-232 landed and raised its verified floor to $2391/400$, and moved $n = 95$ to seventh. $n = 44$, third at $0.0275$, left it on 2026-10-31, when Mohsen Khalili and Sara Gharabaghi’s closed mixed cover, decided by Evan Daniel’s zmx2 and replayed here in both its modes, proved $s(44) = 7$ (T-252), and $n = 95$ returned to sixth. $n = 96$, fourth at $0.0300$, and $n = 76$ and $n = 95$, fifth and sixth at $0.0350$, left it on 2026-11-02, when the mutated-cover controls of zm_mixed.py completed the confirmation of Benjamin Greff’s mixed covers (T-184, T-191), the $s(95)$ cover deciding $n = 96$ as well, and $n = 94$, 75 and 19 took their places. $n = 94$ narrowed from $0.0500$ to $0.0350$ on 2026-11-04, when the whole-net control of Doğa Özden’s dilation of Hosono’s certificate passed and its verified floor rose to $1993/200$ (T-262). Later that day T-233’s four fine-net replays were mapped, once its $n = 26$ control passed under the census’s enclosure rule: $n = 27$ entered fifth, at $0.0521$ from $0.0636$, $n = 19$ narrowed from $0.0606$ to $0.0581$, and $n = 75$, at $0.0600$, left it. On 2026-11-05 $n = 94$ narrowed again, from $0.0350$ to $0.0280$, and passed $n = 12$ for third, when the census replay and whole-net control of Özden’s second dilation passed and its verified floor rose to $2493/250$ (T-266).

With Bentz’s $s(6)$, $s(13)$ and $s(22)$, Evan Daniel’s family $s(m^2 - 3) = m$ is now proved exactly for every $m \ge 3$.

Corrected 2 October 2026. The gaps this section quoted against Nagamochi’s closed form for the last unproved members of $s(m^2 - 3) = m$, $0.0557$ at $n = 97$ (above), $0.0627$ at $n = 78$ and $0.0718$ at $n = 61$, were measured from a floor whose published proof rests on his Lemma 1, which Karakuş showed false (T-085); against Karakuş’s verified floor they were $0.1059$, $0.1185$ and $0.1345$. All three cases are proved now, so devtools.gap_ranking, which ranks the cases still open, leaves them out.

Among the cases with a non-trivial record, $n = 19$ follows $n = 17$ at $0.0681$, then $n = 27$ at $0.0721$, $n = 26$ at $0.0888$ and $n = 28$ at $0.1019$. First-party certificates moved $n = 11$, $17$, $18$ and $19$ beginning on 2026-09-04; the retained $n = 18$ ladder reached $4.679$ on 2026-09-19. External certificates now carry $n = 11$, $17$ to $20$ and $26$ to $28$, Hosono’s replayed rectangle certificates of 1 October among them.

Cross-References

Provenance and Regeneration

The baseline upper bounds, degrees, minimal polynomials, rigidity annotations, analytic status, and attributions were parsed from the retained Kingbird catalogue. uv run --frozen python -m devtools.audit_kingbird_catalogue --rigidity re-reads the rigidity annotations from that archive and checks them against the case records, so a dropped or invented transcription fails instead of being migrated forward. Beside it, uv run --frozen python -m devtools.diff_kingbird_catalogue --before REV compares the retained transcription with the one at REV, count by count: each count whose side, closed form, degree, polynomial, rigidity annotation or credit sentences moved, and where the new side stands against the record. Newer first-party claims are selected through the source-coverage record rather than silently overwriting that baseline. Lower bounds were computed from four sources and the strongest taken. The proved set and its attributions come from the research document’s own analysis, which is sourced to the individual papers.

Editorial bodies for $n = 5, 10, 11, 12, 13, 17, 22, 23, 46, 51, 100$ are hand-written. The remaining bodies are generated from the structured fields. Adding editorial to a case is just editing its body; nothing regenerates over it.

Known limits. The Kingbird catalogue is parsed as annotation text, so an entry phrased unusually can be miscounted; improved_by in particular under-reports where it uses “Refound” or prose. Above $n = 100$ it also credits an “Optimized by” author the entry names nowhere else (Tej Stead at $n = 179$), and the packing paragraph quotes the catalogue’s own AI-assistance sentences; the hand-written records below read neither. Tilt angles are recorded only for the handful of cases where this research established them. The coverage inventory is dated and named; a newly published source still requires a new research survey and disposition.