I am writing this because I believe that prohibited collaboration occurred during this Challenge and affected the prize positions. To state my personal conclusion plainly: I believe this is cheating, not merely a collection of innocent coincidences.

The final positions relevant to the accounts discussed in this post were:

Where I report q below, it is a multiple-comparison-adjusted result from one fixed contest-wide screen. I use it only to prioritize source review, not as the probability that misconduct occurred.

1. paqi and Turritopsis_nutricula: a mirrored submission sequence

This is the clearest timing pattern I found. On August 24, Turritopsis_nutricula made two submissions ten seconds apart. Exactly 60 seconds after each one, paqi produced the same two-submission cadence and reached the same exact score:

The runtimes and memory are not identical, so I am not claiming these four records prove identical binaries. The notable fact is the exact score and the perfectly shifted 10-second, then +60-second cadence.

This is not their only common fingerprint. They share six exact score plateaus in total. Four are rare and pair-exclusive in the full contest data: 8296.577, 9280.584, 16022.487, and 16024.151. For example:

• paqi submitted 387374421 and 387374432 at 10:16:11 and 10:16:20 on August 17, both scoring 16022.487; Turritopsis_nutricula submitted 387374660 at 10:18:36 with the same score.
• On August 22, Turritopsis_nutricula submitted 388054132, scoring 11580.203. After 53 and 60 seconds, paqi submitted 388054167 and 388054175, scoring 11580.163 and 11580.164.

paqi had 325 positive submissions across 13 contest days. Turritopsis_nutricula had 148 positive submissions concentrated on six days, frequently in pairs separated by roughly 7–14 seconds. The repeated cross-account mirroring is what deserves a log and source review. I found no direct public team or group link between these two handles; the only profile-context match nearby is that Hyacinth and Turritopsis_nutricula both self-report Wuhan, which is weak by itself.

The rp++ team connection and Hyacinth

paqi, Hyacinth, and youlinaixu are established teammates in the public Codeforces team rp++, with multiple actual team-contest participations.

Hyacinth made only four submissions in this Challenge. The first two, made 12 seconds apart on August 16, were already very strong and each reached a rounded total also reached by paqi:

Both scores are pair-exclusive to paqi and Hyacinth in the contest data. The order is crossed: paqi first had 15936.908, Hyacinth then submitted both totals, and paqi later reached 15982.034. A direct teammate relationship plus two pair-exclusive shared totals among an account's first two submissions is much more informative than a shared self-reported city.

youlinaixu made only 20 submissions over three contest days. At 10:22:03 on the final day, 388664642 scored 15939.201. At 10:52:46, about seven minutes before the finish, 388667534 scored 16710.886, an improvement of 771.685 points and final rank 11. A late jump is not evidence by itself, and I found no exact paqi score match here. I include it because the organizer can compare this final source with the sources in the much stronger paqi/Hyacinth/Turritopsis_nutricula evidence.

2. mahiro_zcy, mojo__hugo, and Zinc-acetate

The fixed contest-wide screen found two repeated correlations involving final rank 3 mahiro_zcy that my initial report had missed.

mahiro_zcy and mojo__hugo

Across the contest, these accounts have nine event pairs with the same runtime and a score gap no greater than 0.3, representing eight distinct score pairs. With the wider fixed threshold of 1 point, there are 49 event pairs versus 9.44 expected from the empirical baseline among records with the same runtime, representing 23 distinct score pairs; the family-wide screening q-value is approximately 1.7e-16. In eight of the nine closest 0.3-point pairs, mojo__hugo submitted first.

One particularly close example is:

• mojo__hugo 387821595, August 20 21:14:01, score 16324.557, 609 ms, 356,352 bytes;
• mahiro_zcy 387947266, August 22 02:13:46, score 16324.556, 609 ms, 61,440 bytes.

The memory footprints differ systematically, and the lags are generally many hours or days. This is therefore not evidence of the same binary or a real-time hand-off. It is a repeated all-time runtime/near-score correlation that the organizers can resolve by comparing sources. The public team and group records I checked place the two accounts in separate networks and show no direct social link.

mahiro_zcy and Zinc-acetate

This pair shows a different pattern: synchronized score progress rather than matching resource measurements. Within one hour, the fixed screen finds 43 event pairs with score gap at most 1, versus approximately 13.0 expected under the baseline for that pair, spanning 13 distinct score pairs. The directions are mixed: Zinc-acetate is earlier in 26 pairs and later in 17. Within 15 minutes and a five-point gap there are 86 event pairs spanning 36 distinct score pairs, versus 23.3 expected.

For example, Zinc-acetate 388055634 scored 16371.944 at 22:35:50 on August 22, and mahiro_zcy 388057488 scored 16372.003 48 minutes 27 seconds later. Their runtime and memory differ. I found no direct or one-to-two-hop public team-party, shared-group, city, or organization connection between them. The effect could reflect parallel convergence by two highly active contestants on the same deterministic score landscape; nevertheless, its repetition makes it a legitimate source-review lead rather than a single coincidence.

Zinc-acetate separately shares the public group gN3WSMKicW with mojo__hugo and Retired_Zhisheng. That group edge does not establish communication in this Challenge, but it is a concrete public association adjacent to both repeated behavioral patterns.

3. mojo__hugo, Lao_mang, and LA0mAng

mojo__hugo and Lao_mang both publicly list Guangzhou University. More importantly, they share four exact plateaus; three of them are pair-exclusive: 16327.778, 16331.003, and 16336.595. On two different rare scores, I found submissions with the same public score, runtime, and memory:

The time gaps are hours, not seconds. That weakens the timing evidence. In the runtime-only contest-wide screen, at a 0.3-point threshold there are 23 event pairs versus approximately 2.54 expected, representing 11 distinct score pairs, with family q around 1.2e-11. The two exact score/runtime/memory matches above are descriptively specific, but their hours-long gaps remain an important limitation. The direction is consistent across all four shared exact plateaus: mojo__hugo reached each first, and Lao_mang's first occurrence followed by between 8 hours 15 minutes 31 seconds and 18 hours 47 minutes 5 seconds. This remains a hypothesis for source comparison, not public proof of copying.

They are also in the same broader public team-contest circle, although not on a direct common team that I found. mojo__hugo has an actual two-person team party with YGkingNan (team 195089); YGkingNan has an actual party with BCNT and maomaoc (team 203766); and BCNT/maomaoc repeatedly competed with Lao_mang (team 201850). This multi-hop path is context, not transitive evidence of communication in this Challenge.

The Lao_mang/LA0mAng sequence is temporally tighter. The handles are visually similar, and LA0mAng made only 11 positive submissions, all on one day. Two scores, 15963.734 and 15974.214, are exclusive to this pair. The sequence of submissions at 15963.734 alternates between the accounts:

• Lao_mang 387353444, 07:17:27, 15963.734;
• LA0mAng 387354567, 07:28:40, 15963.734;
• Lao_mang 387354956, 07:32:30, 15963.734;
• LA0mAng 387356131, 07:43:44, 15963.734;
• Lao_mang 387358276, 08:04:54, 15963.734.

On the second score, Lao_mang submitted 387373444 at 10:06:24, LA0mAng submitted 387373793 at 10:09:44, and Lao_mang submitted 387374765 at 10:19:49. The last two both scored 15974.214 with exactly 968 ms and 307,200 bytes.

I found no direct public team shared by Lao_mang and LA0mAng. Their profiles list the same city but different universities. However, there is a short actual-party path through BCNT: BCNT repeatedly competed with Lao_mang in team 201850 and with LA0mAng in team 234098. Again, a two-hop social path does not prove that the endpoints collaborated here, but it makes the behavioral correlation less isolated and gives the organizers a concrete comparison set.

4. al1swedel, Tanya160, and Exiled_Code2022

This cluster is weaker than the first two, but it is still worth checking.

The al1swedel account was registered at 08:36:34 UTC on August 21, while the Challenge was in progress. Its first submission came 21 minutes later: 387864060, scoring 12480.305 with a 15,000 ms runtime. The only other account with that exact score was Exiled_Code2022: 387127370, also at 15,000 ms and submitted six days earlier. The long gap makes this a weak link. I preserve the chronology because it was the new account's first submission, but I do not have a population baseline that makes “new account plus first submission” independently suspicious.

al1swedel and Tanya160 share two rare plateaus:

• 16102.414: six submissions by al1swedel, followed 32 minutes later by the first of five from Tanya160; this score was also later reached by a third account, so it is not exclusive;
• 16130.989: first reached twice by Tanya160 on August 25, then first reached by al1swedel about 5 hours 49 minutes later, followed by two more al1swedel submissions the following night; this score is exclusive to the pair.

Their approximately six-megabyte memory range is only corroborative: globally, hundreds of submissions from dozens of handles fall in that broad range. The direction changes between the two plateaus (al1swedel to Tanya160, then Tanya160 to al1swedel). I found no public team, group, organization, or profile link between them. For that reason I classify this as medium-strength behavioral evidence, not as a public identity link.

5. Earlier reports involving IceKylin, candy0014, and Retired_Zhisheng

I reported this group to the organizers during the contest, and it is important to quote both the supporting facts and the exact limited scope of the organizer's response accurately.

Ticket chronology

The ticket text and UI timestamps below were supplied from my organizer correspondence; they are not exposed by the public Codeforces API, and I have not independently verified the UI timezone. The linked submission facts themselves were independently checked.

The profile-name fields cited in #271554

These were profile first-name/last-name fields, not the Codeforces handles themselves. In #271554 I recorded:

• Retired_Zhisheng: oAhWV5MJRCTrRtVCTVfU oAhWV5MJRCTrRtVCTVfU;
• candy0014: JFIdvF5X0UCqmql0rQs9 JFIdvF5X0UCqmql0rQs9.

I considered the anonymized-looking fields and candy0014's later deletion of them suspicious when I reported them. The independently recoverable history is more limited. A Wayback capture from February 25, 2026, about six months before this Challenge, already shows candy0014 using the JFId... string as both first and last name. If my contemporaneous #271554 record is accepted, that string remained visible on August 24 and was cleared sometime afterward; the official API omitted both fields by August 28. Retired_Zhisheng, by contrast, still had the oAh... string in both fields on August 28.

Thus the public evidence supports a later removal of candy0014's fields, not a claim that both accounts first installed random strings after reaching the top. It cannot establish the exact removal time or motive. I still ask the organizers to inspect their own profile-change audit log, which can determine the exact sequence unavailable publicly.

IceKylin and candy0014

There is a historical Codeforces team 166232 named “IceKylin 0339 team.” This is the team's name on Codeforces, not wording I invented for ticket #271544. candy0014 is shown as a member, while IceKylin is shown as a former member. The public page does not prove that their membership periods overlapped during this Challenge, and I found no joint submission under this team.

Ticket #271549 pointed out this pair:

• IceKylin 387862272: 16581.199, 796 ms, 126,976 bytes;
• candy0014 388234286: 16581.492, 796 ms, 65,536 bytes.

The equal runtime and 0.293-point score gap make this more specific than an ordinary nearby-score event, but its strength must not be overstated. Among 604 positive submissions from 32 handles in the post-hoc high-score subset (score >= 16500), there are three cross-account pairs with the same runtime and a score gap no greater than 0.3; the cited IceKylin/candy0014 pair is one of them. For both cited submissions, the other is the closest-scoring foreign submission among all 796 ms submissions. candy0014 has exactly one positive 796 ms submission; IceKylin has two.

The limitations are equally important. Runtime 796 ms occurs in 920 positive submissions from 455 handles. In the contest-wide screen without the high-score cutoff, this pair has one observed same-runtime hit within 0.3 versus 0.01764 expected, raw p 0.01749, but family q 1; 692 of 65,093 directed screens have at least one such hit, 452 have exactly one, and IceKylin has 49 such counterparts. The submissions are also 74 hours 54 minutes apart and their memory differs substantially. Thus this is not evidence of a real-time hand-off or an identical binary. I still consider it a source-review target because the separately established historical-team association makes the coincidence less isolated.

The full history also corrects an overstatement in my initial report: candy0014 did not make only one or two submissions in total. The account made 101 submissions, 87 positive, beginning on August 17. Any review should use that complete chronology rather than the incorrect “one or two submissions” premise.

Retired_Zhisheng and candy0014: what the organizer confirmed, and what remains unchecked

In ticket #271554 I reported the public-profile pattern involving Retired_Zhisheng and candy0014. In the response attached to #271557, the organizer wrote that the accounts were indeed connected. The response then said that their codes were completely dissimilar and that no obvious evidence of joint work was visible. From the ticket order, I understand that answer as referring to the pair reported immediately beforehand.

I do not regard that response as an exoneration. It confirms the non-random part that a public observer cannot normally establish with certainty—the accounts are connected—and reports only that code similarity was not found. The rules prohibit collaboration and sharing solution ideas, not merely submitting textually similar source. Different implementations, independent rewrites, different LLM generations, or code built around shared test-level observations can look completely different while still resulting from prohibited information exchange. The response also does not say whether the organizers compared semantic control flow, constants and policy tables, all intermediate sources, the complete 22-test vectors, compilation artifacts, or account/session logs.

The public strict-best chronology contains several near-synchronized progress episodes. These are not identical final scores, but they are concrete, reproducible jumps rather than a vague impression from the leaderboard:

The clearest pair is therefore not an exact-score collision: it is the August 24 sequence in which candy0014 gained 55.645 points and Retired_Zhisheng gained 49.985 points only 4 minutes 24 seconds later, followed by another 30.444-point improvement from Retired_Zhisheng. Earlier that day, both accounts had already made very large strict-best jumps only 12 minutes 49 seconds apart.

This occurred during a much broader late movement among the prize-position accounts discussed in this post. In the final-score reconstruction, seven of them first crossed 16500 on August 24: candy0014 at 11:34:38 (388234286), paqi at 11:44:35 (388235233), Retired_Zhisheng at 12:38:47 (388241023), mojo__hugo at 16:04:07 (388263099), mahiro_zcy at 18:11:19 (388277401), Zinc-acetate at 18:22:15 (388278759), and Mango2011 at 23:50:59 (388294972). The threshold is descriptive and selected after the fact; deadline pressure and normal parallel progress are possible explanations. Nevertheless, this concentration is exactly why checking only whether one connected pair's source texts look alike is too narrow.

For precision, the pair has no shared positive exact total and no cross-account submissions within 60 seconds; the fixed near-score screen also finds no score-distance-≤5 pair within one hour. Those facts do not negate the table above: the observed strict-best episodes are separated by 4 to 44 minutes and often land at different totals. The relevant organizer-side comparison is whether the same hidden tests, parameters, or scheduling ideas improved at the same moments, which requires the unavailable 22-test vectors and source history.

Because the organizer confirmed the relationship, because the public chronology contains repeated synchronized progress, and because many high-ranking accounts moved sharply in the same late period, I believe this pair should be reopened as part of a cluster-wide investigation. In my view, that combination is affirmative evidence of possible prohibited coordination, not merely neutral social context. “The code is different” answers one narrow question. It does not resolve whether ideas, test discoveries, parameters, prompts, or other prohibited information were shared.

Other all-time leads involving Retired_Zhisheng

The same contest-wide screen found two long-lag same-runtime patterns that I also ask the organizers to inspect:

• Retired_Zhisheng/IceKylin: five event pairs, all distinct, within 0.3 with the same runtime, versus approximately 0.231 expected event pairs from the empirical same-runtime baseline (family q about 0.00436). All five IceKylin records are earlier, and none is within one hour. The closest example is IceKylin 387495808, 15948.596 at 734 ms, followed six days later by Retired_Zhisheng 388195700, 15948.598 at 734 ms. Memory differs.
• Retired_Zhisheng/BitByBit123: 28 same-runtime event pairs within one point, spanning 19 distinct score pairs, versus approximately 0.275 expected (family q about 3.15e-42). Every BitByBit123 record is earlier; none is within one hour. For example, 387787710, 15948.065 at 687 ms, precedes Retired_Zhisheng 388170345, 15947.944 at 687 ms, by more than three days. Memory differs.

These are repeated all-time output/runtime correlations, not evidence of a live hand-off. I found no direct public association between Retired_Zhisheng and IceKylin. A separate endpoint check also found no shared actual party, Codeforces team, group, city, organization, or party/team intermediary for Retired_Zhisheng and BitByBit123. Their consistency and one-way order make them useful source-comparison leads, while the long delays remain a substantial alternative-explanation caveat.

Compact appendix for the accounts I am asking the organizers to review

For completeness, this table gives a strongest additional counterpart for each remaining account discussed in this review, selected by the same fixed, symmetric contest-wide comparison rules. “Events/distinct” prevents a large number of resubmissions at one score from looking like many independent discoveries. Different rows can use different fixed profiles, so the counts are not directly comparable across rows. “Screen-positive” means only that the selected correlation survives the screen's multiple-comparison rule; it does not mean that misconduct was established.

Pair-by-pair reading of the appendix

The compact table is easy to misread, so below is the concrete interpretation of every fixed-profile representative. Expected counts and q-values come from the same fixed comparison families for all handles; they do not incorporate public-team evidence and do not turn a correlation into a verdict.

• Rank 2, IceKylin / mayanktr3900. There are eight events versus 0.0778 expected in the same 15-minute/five-point family, but only two distinct score pairs (q about 6.2e-11). One pair is only 11 seconds apart: IceKylin 388657345 at 15906.799, followed by mayanktr3900 388657369 at 15905.495. Their runtimes and memory differ. The small number of distinct plateaus and final-day burst structure make this much less robust than the event count suggests; no public relationship was found.

• Rank 3, mahiro_zcy / nameyoutube. The all-time same-runtime five-point family has 214 events versus 33.81 expected and 130 distinct score pairs (q about 1.13e-91). A close example is mahiro_zcy 388040701, 16390.156 at 421 ms, and nameyoutube 388088104, 16390.028 at the same 421 ms roughly ten hours later. Memory differs. This is a broad all-time runtime/near-score correlation, not a short-time transfer, and the public team and group records I checked show no shared public object.

• Rank 4, Zinc-acetate / Calto. The corresponding all-time same-runtime family contains 154 events versus 18.26 expected and 99 distinct score pairs (q about 3.87e-82). Calto 387313504 scored 16393.839 at 546 ms; almost a week later, Zinc-acetate 388060975 scored 16393.889 with the same runtime but very different memory. Again, this is a long-lag runtime/near-score correlation. No shared team, party, group, city, organization, or enumerated intermediary was found between the endpoints.

• Rank 6, paqi / Irval. The all-time same-runtime/five-point family has 116 events versus 7.32 expected and 73 distinct score pairs (q about 2.28e-90). paqi 387872660 scored 16382.978 at 609 ms; Irval 388631798 scored 16382.946 at the same runtime approximately 6.7 days later. Memory differs, and the closest examples consistently place Irval later. The long lags and absence of a public relationship make this a source-review lead, distinct from the much tighter paqi/Turritopsis_nutricula cadence.

• Rank 7, Retired_Zhisheng / BitByBit123. This is the 28-versus-0.275, 19-distinct, one-way all-time same-runtime pattern detailed above (q about 3.15e-42). Every matching BitByBit123 record is earlier and every lag exceeds one hour. It deserves semantic source comparison, but not a claim of a live hand-off from timestamps alone.

• Ranks 10 and 14, mojo__hugo / Lao_mang. Their reciprocal all-time same-runtime/five-point screen has 603 event pairs versus 41.77 expected and 178 distinct score pairs (q about 3.3e-296). That broad count, the narrower 23-versus-2.54 result at 0.3 points, the pair-exclusive plateaus, shared Guangzhou University field, and the three-hop team-party path are all described above. The pair has no direct common team.

• Rank 11, youlinaixu / tony1107. There are only two all-time same-runtime/five-point events versus 0.0821 expected, two distinct pairs, and family q about 0.457. For example, tony1107 387769338 scored 15689.969 at 1281 ms, while youlinaixu 388525112 scored 15685.223 at the same runtime almost a week later; memory differs. This does not survive the multiple-comparison correction. The relevant public connection for youlinaixu remains the direct rp++ party with paqi and Hyacinth, not this weak pair.

• Rank 15, al1swedel / Mohamed_Mahmoud11. The 15-minute/five-point screen contains 33 events versus 5.85 expected and 13 distinct score pairs (q about 1.5e-12). al1swedel 388114190 scored 16102.138; Mohamed_Mahmoud11 388114511 scored 16102.728 four minutes 18 seconds later. Their runtime and memory differ. Mohamed_Mahmoud11 belongs to many public groups, but none is shared with al1swedel; no team, city, organization, or party/team intermediary was found. This broad timing result is separate from the more interpretable al1swedel/Tanya160 exact plateaus.

• Rank 16, Jeffrey / Lyoshach. The 15-minute/one-point family has four events versus 0.0176 expected, two distinct score pairs, and q about 2.6e-05. Lyoshach 387991882 scored 16027.901; Jeffrey 387992371 scored 16027.563 five minutes 11 seconds later. Runtime is the same 1937 ms, but memory differs. With only two distinct pairs and no public relationship, I classify this as repeated but fragile.

• Rank 17, Mango2011 / yuangengqi. The 15-minute/five-point family has 19 events versus 0.685 expected but only three distinct score pairs (q about 6.7e-18). yuangengqi 387352967 scored 15917.136, followed 55 seconds later by Mango2011 387353049 at 15915.376; their runtimes and memory differ. Both have public team histories, but their party/team neighborhoods are disjoint. The result is burst-dependent and is not evidence of a shared binary.

Several additional repeated correlations were not each account's single strongest profile but are still worth preserving for source review:

Finally, these are the very short one-off events that I found and preserved even though none survives the fixed multiple-comparison screen:

There is also a second Mango2011/mahiro_zcy observation 453 seconds apart with score gap 0.229, but it remains part of the same weak local episode. I found no exact Codeforces handle or public historical mapping for the label kitaksu. kitekat1__4 is a real account, but equating it with kitaksu would be speculation; no public team/group/profile link to IceKylin was found.

I regard this documented set—especially where repeated score behavior overlaps with a concrete public association—as strong grounds for a source-and-log review. I am not claiming that the one-off rows independently prove cheating; their purpose is to ensure the organizers receive every reproducible lead rather than only the most dramatic examples.

What I am asking the organizers to check

Please do not decide this from rounded totals alone. The organizers can test the hypotheses directly:

1. Compare source code semantically, including control flow, constants, policy tables, class-specific branches, and unusual ordering decisions—not only textual similarity.
2. Compare the complete 22-test score vectors and per-test metric vectors for the cited submissions. A 22-dimensional fingerprint is much more informative than one rounded total.
3. Check whether the paired submissions came from related sessions, devices, networks, or compilation artifacts. This can be done privately; no personal data needs to be published.
4. Review the membership history and verified contestant identities behind the public teams and groups.
5. Because the final judging rule selected the latest strictly positive submission, inspect that selected source for each prize-relevant account as well as the cited probe submissions.
6. If the review finds no violation, publish an anonymized explanation of what was checked. That would be much more convincing than silence and would protect innocent contestants from speculation.

My position

I reached second place during the Challenge and was pushed down to fifteenth place in less than three hours. I ultimately finished eighteenth. Seeing that sudden movement alongside the correlations above is why I investigated instead of silently accepting the result.

Under this handle, I finished within the published prize ranges in three previous Huawei/ICPC Challenges, and the current result is a fourth such placement. I have invested a great deal of time in this format. I am not asking for automatic disqualifications from a spreadsheet; I am asking the organizers to use the evidence above to perform the source-and-log review that only they can perform. My conclusion from the public pattern is nevertheless direct: I believe prohibited collaboration occurred and that it materially affected the prize standings.

If the organizers are not prepared to conduct and explain a source-and-log review when presented with evidence of this kind, then in the LLM era it is time to admit that the Huawei Challenge format no longer has a meaningful competitive future; after four published prize-range placements in these Challenges under this handle, I will no longer recommend that anyone participate in them again.

P.S. I apologize if parts of this post read awkwardly. I used machine translation for some passages because I did not have time to polish the English, and I believe the organizers need to see this information as soon as possible.

Btw, here’s a funny video showing how the cheaters simultaneously pushed me from top 2 to top 15 in less than three hours: https://www.youtube.com/watch?v=yXFaB-dMkpk