$n=304$$n=304$

$s(304)$$s(304)$ — open

Hosono’s closed form, 14 October 2026. Hosono’s closed form (wand125, 13 October 2026, T-180) states $s(N)\ge a(N)$$s(N)\ge a(N)$ for every non-square $N\ge 49$$N\ge 49$, from a measure of line segments on 34 intervals of the fractional part. At $n=304$$n=304$, $k=17$$k=17$ and interval 25 give $a(304)=17.490382985636778684148708282520\dots$$a(304)=17.490382985636778684148708282520\dots$, so the reported lower bound here is now $s(304)\ge 17.490382985636$$s(304)\ge 17.490382985636$, rounded down, above Nagamochi’s closed form, $17.46207763315$$17.46207763315$. The source replayed its certificate with its own checker; no record of it has been replayed here, so the verified lower bound is unchanged. The source states that its search, certificates, checkers and text were produced with AI assistance (Claude and Codex), under the direction of and reviewed by Hosono.

House adoption, 16 October 2026. Polkovskyi’s tighter certificate of Ostashev’s n = 304 packing (T-186) is now this case’s house: its exact certificate proves $s(304)\le 17.92942775807517497$$s(304)\le 17.92942775807517497$, the reported and the verified upper bound, rounded outward from its exact side. The exact certificate side improves the previous verified exact side by about 1.79×10−19$1.79\times {10}^{-19}$. The earlier ceiling, displayed outward as $17.9294277580751750$$17.9294277580751750$ from Ostashev’s packing (T-175), applied until this adoption and stays retained as history (the certificate).

House adoption, 15 October 2026. Ostashev’s packing (T-175) was this case’s house until 16 October 2026: its exact certificate proves $s(304)\le 17.9294277580751750$$s(304)\le 17.9294277580751750$, the reported and the verified upper bound, 1.38×10−4$1.38\times {10}^{-4}$ below the ceiling the case held until then, $17.9295659468485010$$17.9295659468485010$ from Chaoweeraprasit’s packing, improved by Couzo (T-152), which stays retained as history (the certificate).

House adoption, 12 October 2026. Couzo’s refinement of Chaoweeraprasit’s SQUISH Mondrian packing (T-152) was this case’s house until 15 October 2026: its exact certificate proves $s(304)\le 17.9295659468485010$$s(304)\le 17.9295659468485010$, the reported and the verified upper bound, 5.08×10−3$5.08\times {10}^{-3}$ below the ceiling the case held until then, $17.934650018390681724105084929881$$17.934650018390681724105084929881$ from Couzo’s packing, improved by Daniel (T-098), which stays retained as history (the certificate).

Open. Previously, the best known published packing, Couzo’s at Daniel’s exact optimum (T-098), gives $s(304)\le 17.9346500183906817242\dots$$s(304)\le 17.9346500183906817242\dots$, and the strongest lower bound independently verified here is $17.448451$$17.448451$ from Karakuş’s general theorem, leaving a gap of $0.4862$$0.4862$. General bound: s(N) >= 1/2 + sqrt(N - floor(sqrt(N)) + 1/4).

The exact optimum

Daniel’s square-packing published on 5 October 2026 an exact rational certificate of this packing at its exact optimum (T-098): the same 304 squares in a square of side $17.9346500183906817242\dots$$17.9346500183906817242\dots$, 5.2×10−12$5.2\times {10}^{-12}$ below the side Couzo prints. Square for square, its pose lies within 1.5×10−4$1.5\times {10}^{-4}$ of the binary64 pose the atlas pictures for this count. 3 squares move by more than 1×10−8$1\times {10}^{-8}$, all of them squares the source lists as carrying no force; every other square moves by at most 1.0×10−11$1.0\times {10}^{-11}$. The known-best witness this record lists is that binary64 pose, posed at its finder’s larger side; the side above is witnessed by the certificate itself. Daniel’s solver moves the binary64 pose to a nearby exact KKT point of the problem of minimizing the side under non-overlap, computed at 80 digits, and rounds it outward to rationals, each square a rational centre and a rational $t=\tan (\theta /2)$$t=\tan (\theta /2)$.

This repository decides the certificate exactly. Converted without rounding, every pair and every wall is decided over $\mathbb{Q}$$\mathbb{Q}$ twice, by sqpack’s exact separating-axis test and by an independent checker that shares no code with it, and the source’s own two checkers, run here as retained, accept it as well (receipts). That proves $s(304)\le 17.9346500183906817242\dots$$s(304)\le 17.9346500183906817242\dots$, the verified upper bound until 12 October 2026; it says nothing about optimality. The source reports the exact point as a KKT local minimum: multipliers that keep it in equilibrium, a reduced Hessian positive definite once its exact flat motions are set aside (its per-count report), and no first-order descent across corner-to-corner contacts, each computed numerically at that point; none of that is verified here, and it bears on this packing alone, not on $s(304)$$s(304)$. On jlevy/squares#375 its author wrote that the solver, its checkers and the batch “were written with Claude (Anthropic) as a coding and research agent, directed and reviewed by me.”

The packing

Couzo’s square-packing reports a packing of side $17.934650018395903$$17.934650018395903$ for this count, dated 26 September 2026 and unchanged when this record retained the repository on 27 September 2026 (T-056). Its first packing for this count, of side $17.936633693223122$$17.936633693223122$, is dated 24 September 2026. The repository names no method and no tolerance, and itself states no AI assistance; its author said on issue #227 that he found the 102 and 103 packings “with the help of Claude”.

This repository certifies it exactly. The retained decimal pose rounds to an exact rational packing at centre dilation 1, of side $17.9346500183959035566\dots$$17.9346500183959035566\dots$, 5.566×10−16$5.566\times {10}^{-16}$ above the printed side, and every pair and every wall is decided over $\mathbb{Q}$$\mathbb{Q}$ twice, by the promotion’s exact separating-axis test and by an independent checker that shares no code with it (receipt). That proves $s(304)\le 17.934650018395904$$s(304)\le 17.934650018395904$, which was the verified upper bound until Daniel’s exact optimum above replaced it; it says nothing about optimality. Interval arithmetic on the printed pose itself, with each angle’s true cosine and sine and no rational rounding, decides every pair and wall again and gives the same bound (interval route).

Casson’s square-packing, dated 23 September 2026 and made with the help of Claude as its README says, reports a packing of side $17.94880191214547338$$17.94880191214547338$ for this count, larger than Couzo’s by $0.01415189374957038$$0.01415189374957038$. Casson’s is the earlier of the two by the timestamps this record’s priority notes keep: Couzo’s first packing for this count, of side $17.936633693223122$$17.936633693223122$, is dated 24 September 2026. The record states both dates and infers nothing about whether either packing derives from the other.

The previous best known packing

Before this intake the best known packing was the Kingbird catalogue’s, of side $17.94917201919110$$17.94917201919110$.

Found by Ellsworth in 2026, via simulated annealing.

The lower bound

The strongest lower bound independently verified in this record is Karakuş’s general bound, which applies to every nonsquare $N\ge 8$$N\ge 8$:

Correction, 2 October 2026. Until that date the verified lower bound here was Nagamochi’s general closed form, s(N)≥min(⌈N​⌉,N−2⌊N​⌋+1​+1)$s(N)\ge min(\lceil \sqrt{N}\rceil ,\sqrt{N-2\lfloor \sqrt{N}\rfloor +1}+1)$, which is stronger at this $n$$n$ and is now recorded as a reported bound. Its published proof rests on Nagamochi’s Lemma 1, which Karakuş showed false; nothing is disproved, and no packing beating it is known (review of 2 October 2026). This register had recorded that proof as verified, its own error, logged as defect D-516.

Source-reported bounds are recorded separately from this independently verified theorem.

Verification Code

The programs behind this case’s verified bounds, by their evidence. The code column says how the code that ran stands to the code its producer used. VERIFIERS.md says what each program is and whose it is.

Francisco Couzo’s certificate (T-152)

Couzo’s refinement of Chaoweeraprasit’s SQUISH Mondrian packing, reported on jlevy/squares#506, at exact side $8964782973424250487190982671787/500000000000000000000000000000$$8964782973424250487190982671787/500000000000000000000000000000$, displayed upward as $17.9295659468485010$$17.9295659468485010$ (T-152). The certificate gives a rational side and, for each of the 304 unit squares, a rational centre and the tangent of its half-angle; the retained packet holds it with the complete native jobs that decided it.

Both maintained exact routes decided it here, sqpack’s exact separating-axis verifier and the independent rational corner checker, with its duplicate-square and outside-container controls, and a third exact route, half-extent containment and exact intersection area, decided it again. One maintained route is the verifier the source ran; the independent rational corner checker and the third route are independently re-implemented. The review of 12 October 2026 accepted the replay for finite feasibility. That proves $s(304)\le 17.9295659468485010$$s(304)\le 17.9295659468485010$, the verified upper bound until 15 October 2026: confirmed, independently re-implemented. It says nothing about optimality, rigidity or a local minimum.

The author says the search code and the request were written with Claude under his direction.

The house it replaces, $17.934650018390681724105084929881$$17.934650018390681724105084929881$ from Couzo’s packing, improved by Daniel (T-098), stays retained with its own replay as history. The other certificates confirmed here at this count, T-151 at $17.9296570421708801$$17.9296570421708801$, stay verified and are not selected.

Daniil Ostashev’s certificate (T-175)

Ostashev’s packing, reported on jlevy/squares#575, at an exact rational side, displayed upward as $17.9294277580751750$$17.9294277580751750$ (T-175). The certificate gives a rational side and, for each of the 304 unit squares, a rational centre and the tangent of its half-angle; the retained packet holds it with the complete native jobs that decided it.

Both maintained exact routes decided it here, sqpack’s exact separating-axis verifier and the independent rational corner checker, with its duplicate-square and outside-container controls, and a third exact route, half-extent containment and exact intersection area, decided it again. No route here is the source’s code, so the replay is independently re-implemented. The review of 14 October 2026 accepted the replay for finite feasibility. That proves $s(304)\le 17.9294277580751750$$s(304)\le 17.9294277580751750$, the verified upper bound until 16 October 2026: confirmed, independently re-implemented. It says nothing about optimality, rigidity or a local minimum.

The issue says the author drives an AI-assisted research loop with his own ideas, merging the tilted bands being his, and that AI agents wrote and ran the search and verification code.

The house it replaces, $17.9295659468485010$$17.9295659468485010$ from Chaoweeraprasit’s packing, improved by Couzo (T-152), stays retained with its own replay as history. The other certificates confirmed here at this count, T-174 at $17.9295659468485010$$17.9295659468485010$ and T-151 at $17.9296570421708801$$17.9296570421708801$, stay verified and are not selected.

Mykhailo Polkovskyi’s certificate (T-186)

Polkovskyi’s tighter certificate of Ostashev’s n = 304 packing, reported in his repository on 14 October 2026, with no issue, at an exact rational side, displayed upward as $17.92942775807517497$$17.92942775807517497$ (T-186). The certificate gives a rational side and, for each of the 304 unit squares, a rational centre and the tangent of its half-angle; the retained packet holds it with the complete native jobs that decided it.

Both maintained exact routes decided it here, sqpack’s exact separating-axis verifier and the independent rational corner checker, with its duplicate-square and outside-container controls, and a third exact route, half-extent containment and exact intersection area, decided it again. No route here is the source’s code, so the replay is independently re-implemented. The review of 16 October 2026 accepted the replay for finite feasibility. That proves $s(304)\le 17.92942775807517497$$s(304)\le 17.92942775807517497$, the verified upper bound: confirmed, independently re-implemented. It says nothing about optimality, rigidity or a local minimum.

The certificate is Polkovskyi’s tightening of the packing T-175 records, measured here pose by pose as that packing (T-186’s regularizations), so the packing keeps its finder’s credit and Polkovskyi is credited with the certificate.

The commit makes no statement about AI assistance; the author’s issue 593 discloses Gemini assistance with solver scripting, precision checks and formatting.

The house it replaces, $17.9294277580751750$$17.9294277580751750$ from Ostashev’s packing (T-175), stays retained with its own replay as history. The other certificates confirmed here at this count, T-174 at $17.9295659468485010$$17.9295659468485010$, T-152 at $17.9295659468485010$$17.9295659468485010$ and T-151 at $17.9296570421708801$$17.9296570421708801$, stay verified and are not selected.