Keep “IMD seat 798” live and send these candidates to cross-play. No supplied model clearly combines an improvement against current with 99 percent against Starter and resistance to an effective fresh exploiter. All ten leaders against current score below 0.5 against their own reported exploiters. This recommendation follows from the matchups and protocol limitations; a top-three board finish remains unproved. M1: table, notes and conclusion M2: table and notes

The requester's task snapshot, dated 2026-11-06, places the network 5th of 36: “IMD seat 798” has 53 kills and 7 losses in 60 matches; third has 59 kills and a draw, and the leader has won every match. The brief's fitted rating is -6121.625 + 400 x log10(s / (1 - s)), with score per match s capped at 0.999; its stated target is 59 of 60. These are the requester's historical figures, not a new board measurement. The brief also says that the latest upload replaces the previous entry and that the operator uploads manually. The challenge's model interface is 224 observation numbers and 3 action numbers. Challenge repository

The supplied merges, M1 and M2, describe local experiments against public or locally trained opponents. They cannot identify the flight styles responsible for the seven private-board losses. Raw item JSON, models and replays were not supplied here; these are transcribed reports, without match reruns. Source identities appear below. M1: opening and conclusion M2: provenance; private-opponent limitation: requester brief.

The table ranks candidates by final score against current, descending. Candidate-side scores credit elimination wins 1, HP wins 0.75 and draws 0.5, divided by matches. Starter uses 3,000 random-spawn matches; built-ins aggregates four opponents over 800; current uses 1,000, requested half from each side; exploiter evaluation uses 500. W/L/D means wins/losses/draws; wins can include HP wins. Scores preserve source precision, including 1 versus 1.0. The Starter guard uses reported score because complete outcome counts are generally unavailable. M1: record definitions and table M2: abbreviations and table

R means exactly Abs, Add, Atan, Cast, Concat, Cos, Div, Gather, Greater, Max, Min, Mul, Neg, Sin, Split, Sqrt, Sub, Where; plus signs add operators. Sizes are decoded ONNX bytes. Both merges report size and operator checks; exceptions concern other rows, discussed below. M1: metadata check M2: metadata check and abbreviations

Rank / model / source	What it is	Starter (3000)	Built-ins (800)	Current (1000): score; W/L/D	Own exploiter (500)	Bytes	ONNX operators
1. sp-starter-1 M1	Starter imitation then PPO league; FP16 224-64-64-3	0.9855833333333334	0.989375	0.938; 936/59/5	0.034	38122	Cast, Gemm, Relu
2. sp-curriculum-2 M2	Two-regime rule with evolved 32-ReLU random-feature selector	0.9755	1	0.917; 906/72/22	0.144	23153	R + Gemm, Relu
3. sp-curriculum-4 M2	PPO-trained 8-wide ReLU selector between two tuned rules	0.9928333333333333	1	0.8655; 838/107/55	0.065	36463	R + Gemm, Relu
4. sparring-head M1	Rule plus zero-started residual MLP; 56 observation values + 3 rule actions, 64 hidden	0.9745	0.99875	0.828; 810/154/36	0.269	35530	R + Gemm, Relu, Slice
5. sp-distill-1 M1	Rule inside 150m; learned 47-32-32-3 pursuit controller outside	0.99375	0.99875	0.8195; 802/163/35	0.254	31919	R + Gemm, Relu
6. rule-mutual M1	Rule constants retuned with mutual-kill penalties	0.9966666666666667	0.99875	0.8145; 788/159/53	0.06	18930	R
7. sp-gradient-5 M2	Rule plus PPO 56-64-24-3 residual, clipped to [-2,2]	0.9805	1	0.794; 757/169/74	0.401	40926	R + Clip, Gemm, Relu
8. rule-defence M1	Three-way gated defence, weaving and late fallback	0.9975	1.0	0.7775; 739/184/77	0.273	31253	R
9. rule-constants M1	Existing rule graph, constants tuned against a diverse pool	0.9963333333333333	1.0	0.72; 667/227/106	0.293	18930	R
10. sp-gradient-2 M2	Rule plus weight-averaged PPO 56-48-24-3 actor adjusting speed, lead and firing tolerance	0.998	1	0.713; 622/196/182	0.43	36916	R + Gemm, Relu

The exploiter column is not a common-opponent league. For sp-starter-1, the selected counter was the strongest of three PPO runs with an imitation start and a reported 9 minutes; sparring-head used fresh rule-residual ES for 545 s. Sp-distill-1 reports adaptation lineages and selection lasting 556.4780877099838 seconds; rule-mutual reports concurrent rule/head searches lasting 595.2301568984985 seconds; rule-constants reports 591.6773280350026 seconds. Rule-defence reports 603.5821514129639 seconds for the selected rule search and 626.6040420532227 seconds for its parallel neural search, exceeding a strict ten-minute limit. M2 describes ten-minute searches for sp-curriculum-2, sp-curriculum-4 and sp-gradient-5; sp-gradient-2 selected the strongest of four 600 s searches. These are training budgets; 500 is the evaluation sample, not the training budget. M1: corresponding rows and rule-defence note M2: corresponding item notes and group discussion

Within the top ten, sp-curriculum-4, sp-distill-1, rule-mutual, rule-defence, rule-constants and sp-gradient-2 clear a Starter score of 0.99 and have substantial reported advantages against current. Their exploiter scores are, respectively, 0.065, 0.254, 0.06, 0.273, 0.293 and 0.43. Sp-starter-1, sp-curriculum-2, sparring-head and sp-gradient-5 fall below the Starter guard. Sp-gradient-2 is the most promising of these eligible leaders on its own counter-test, but different counter families and missing shared-opponent comparisons prevent a general robustness ranking. M1: table M2: table and sp-gradient-2

Lower-ranked apparent exceptions need care. Rule-plus-net reports Starter 0.9903333333333333, current 0.5765 and exploiter 0.568, but its ten-minute search failed to improve on the starting current-rule opponent; that is evidence against the retained baseline, not successful resistance to a newly effective counter. Rule-gunnery reports 0.9933333333333333, 0.5345 and 0.503 respectively, but its restricted six-constant counter search lacks a documented ten-minute duration. Its two 5,000-match current holdouts scored only 0.5094 and 0.5103. Both merit cross-play; neither establishes the requested combined improvement. M1: rule-plus-net and rule-gunnery

The five quick seats identified these practical constraints:

• harness-speed: The unchanged rule ran at 41-45 matches a second against Starter under ONNX Runtime, versus 656-671 with native rule evaluation. Inference was the reported bottleneck. The same current rule scored 0.17 against a newly searched rule counter. Use native evaluation to accelerate search, with ONNX validation of finalists and parity checks; speed alone does not improve the policy. M1: harness-speed
• harness-losses: The audit classified 18 of 23 Starter losses as opening geometry and 5 as failed finishing shots. A one-HP dodge intervention cut mirror mutual kills from 200 to 71 but lowered score to 0.4675. The unchanged rule scored 0.259 against its 600-second rule hill-climb exploiter. These are reported local diagnoses, with an unresolved external-source warning; they do not identify the private board's losses. M1: harness-losses and provenance warning
• sparring-head: A residual that changed flying rather than gunnery reached 0.828 against current, while Starter dropped to 0.9745 and the fresh-counter score was 0.269. Its 35530-byte graph adds Gemm, Relu and Slice to R. Evasive pursuit is a useful training target, but this candidate sacrifices the Starter guard. The item attributes the historical 27 percent neural-loss anecdote to the earlier n7 rule, rather than the final current entry. M1: sparring-head
• sparring-small: A 4677-byte 56-16-3 ReLU network, using Gather, Gemm and Relu, learned a ceiling-camping specialist. It scored 0.60375 against current but only 0.20925 against Starter, 0.003125 against built-ins and 0.0 against its own exploiter. A tiny policy can expose a distinct weakness without being a general replacement. M1: sparring-small
• size-and-ops: Compiler/linker failures and SIGSEGV prevented simulator measurements and the size sweep. Its 0.9933333333333333 Starter value was inherited, not measured. The returned model was the unchanged 18930-byte rule; its supplied operator list was wrong, and M1's protobuf inspection found R. Consequently these seats establish no optimal model size or measured operator-cost sweep. M1: size-and-ops and metadata check

On flight speed, the separate rule-energy seat rejected every tested fixed or conditional target-speed override and returned current unchanged. Its comparisons used only 80 to 120 matches per variant, and its 0.838 “exploiter” score was against an untrained full-throttle substitute. This supports retaining the existing throttle policy pending stronger evidence. Rule-defence reduced hits taken from 4.519 to 4.171 on paired seeds, yet remained exploitable; rule-mutual still had 53 mutual kills against current and 46 in its 1000-match mirror. Improving a single failure mode did not establish broad resilience. M1: rule-energy, rule-defence and rule-mutual

For each five-run self-play family, cells give final score against current / final score against own exploiter. “Unchanged” means current was returned; “blocked” has no usable measurements. Short or ineffective searches remain as reported and are qualified below. M1: self-play rows M2: table and blocked-record notes

Family / source; keys are sp-family-run	Run 1	Run 2	Run 3	Run 4	Run 5
starter M1	0.938 / 0.034	0.004 / 0.002	0.0065 / 0.008	0.0 / 0.052	0.0015 / 0.123
distill M1	0.8195 / 0.254	0.518 / 0.51	0.1665 / 0.068	0.516 / 0.074	0.419 / 0.079
league M2	unchanged: 0.5 / 0.5	0.668 / 0.129	0.711 / 0.1355	unchanged: 0.5 / 1	0.7065 / 0.4005
curriculum M2	0.001 / 0.5	0.917 / 0.144	blocked; unchanged	0.8655 / 0.065	0.06 / 0.02949999999999997
evolution M2	0 / 0	0.1625 / 0.002	blocked; unchanged; unverified records excluded	0.001 / 0.228	0.197 / 0.098
gradient M2	0.017 / 0.488	0.713 / 0.43	blocked; no model	0.3265 / 0.296	0.794 / 0.401
response M2	0 / 0.01	0.5785 / 0.395	0.6185 / 0.058	blocked; no model	0.6625 / 0.147

Starter imitation: Run 1's PPO league produced the strongest current-matchup result across the supplied tables, but its Starter score was 0.9855833333333334 and an independently imitation-started counter reduced it to 0.034. Run 2's gated Starter experts failed against current. In run 3, ES improved Starter performance but washed out a narrow current-rule specialist during balanced-league training. Run 4's built-in score was 0.020625; run 5's was 0.0265625, and run 5 omitted current from its training pool. Thus imitation plus PPO produced a valuable adversarial policy here, while these ES/gated variants did not transfer reliably. Run 3's exploiter used 9.5 minutes, run 4's 7 minutes and run 5's 200 s; they do not establish equal-budget method comparisons. M1: sp-starter-1 through sp-starter-5

Rule distillation: Run 1 succeeded as a hybrid: pure cloning underperformed, while preserving the rule inside 150m and learning pursuit outside delivered the strong record in the main table. Run 2's 1% student blend reached only 0.518 against current and Starter 0.989. Its recorded exploiter score of 0.51 uses the zero-bias/current-rule choice even though another confirmation score was 0.49 against the +1 shoot variant; no timed ten-minute search is documented. Run 3's DAgger clone remained weaker than its teacher and rejected damaging ES fine-tuning. Run 4 came close to a current tie but scored 0.9534375 against built-ins, including 93 HP wins; run 5 did not resolve sharp rule-action switches. These runs support preserving exact tactical geometry while learning selected behaviours; they do not demonstrate a faithful, robust standalone distilled policy. Counter durations for runs 3, 4 and 5 were 420 s, 570 s and 560 s. M1: sp-distill-1 through sp-distill-5

League training: Runs 2, 3 and 5 improved the direct matchup by tuning or selecting rules, but all lost to fresh counters. Run 2's 258-opponent minimax league still yielded 0.129 against its exploiter, while current scored 0.81 against that same opponent on the same 500 cases. Run 3 selected over 88 opponents and achieved Starter 1, but its built-in score was 0.89625 and exploiter score 0.1355. Run 5 preserved Starter 0.99516666, but its 480 s head-only counter held it to 0.4005 while current scored 0.879 against that counter. Its exported ONNX was not executed in ONNX Runtime; only onnx.checker passed. Runs 1 and 4 returned current unchanged because their searches found no effective improvement; run 1's exploiter was a clone, and run 4's failed to earn a kill or HP win. Their high resistance figures cannot establish hardening. M2: sp-league-1 through sp-league-5

Curriculum: Run 1's scratch-trained MLP never passed the built-ins gate and used only a 200 s exploiter. Runs 2 and 4 found strong rule selectors; run 2 missed its current-entry curriculum gate and the Starter guard, while run 4 passed the Starter guard but collapsed to its counter. An earlier run-4 checkpoint scored 0.058 against its challenger, whereas current scored 0.64185 against that challenger over 10,000 matches: a direct warning about matchup cycles. Run 3 was blocked by missing OpenSSL headers/pkg-config. Run 5's DAgger policy scored 0.9484166666666667 against Starter and 0.06 against current. Staging opponents helped the rule selectors' direct matchup, but these runs did not complete a robust progression to the goal. M2: sp-curriculum-1 through sp-curriculum-5

Evolution from scratch: Run 1's 4592-byte network reached Starter 0.9926666666666667 yet lost 1,000 of 1,000 to current and scored 0 against its exploiter. Runs 2 and 4 specialised to weaker opponents. Run 5's current result was 0/606/394, with no wins; its replay audit attributed Starter wins to Starter hitting the ground rather than gun hits. Run 3 failed to build a working binary and returned current; its carried-forward figures are not measurements. These runs show that small networks and Starter success can coexist with complete tactical failure against current. They do not isolate whether more compute, different rewards or different capacity would repair scratch evolution. M2: sp-evolution-1 through sp-evolution-5

Policy gradients: Runs 2 and 5 retained the rule and learned adjustments; both gained substantially against current. Run 2 preserved Starter 0.998 and reached the strongest documented own-exploiter score among M2's improved candidates, 0.43. Run 5 fell to Starter 0.9805; its counter is particularly useful for the next league because the counter also scored 0.6355 against current. Standalone clone/PPO runs 1 and 4 remained below current after about 8 minutes and about 20 minutes of PPO, respectively; run 1's counter ran only 150 s. Run 3 was blocked by a missing C linker. These are positive results for learning on top of the rule, with inadequate evidence for dismissing longer standalone PPO. M2: sp-gradient-1 through sp-gradient-5

Best-response chains: Run 1 drifted into policies that beat one another but scored 0 against current, and its counter ran only 240 s. Runs 2 and 3 improved current matchups but exposed selection-set overfitting and a strong neural-head counter respectively. Run 4 was blocked by a missing linker toolchain. Run 5 repaired Starter to 0.999 and built-ins to 0.99875, but still scored only 0.147 against its exploiter. Its chain illustrates forgetting: link 1 scored 0.901 against current and link 6 scored 0.158. Its large final evaluations used the native graph, with ONNX agreement reported on 9,024 decisions. The useful direction is retaining current and previous counters throughout training, then selecting on new opponents and fresh seeds; advancing down a response chain alone did not produce general strength. M2: sp-response-1 through sp-response-5

Most gains retained the rule with learned adjustments. Sp-starter-1 is the standalone PPO exception, although it fails the Starter guard and counter-test. M2's “every gain” conclusion applies only to its own group. Blocked runs and unequal budgets limit conclusions about algorithms. M1: sp-starter-1 and sp-distill-1 M2: table and across-group discussion

Other rule experiments reinforce the same caution. Rule-losses reached Starter 0.9986666666666667 and current 0.5645, but only 0.212 against its selected counter. Rule-harden returned current without completing its exploit-and-patch rounds; its 0.994 counter score came from a weak search. Rule-opening also returned current, with no successful opening patch or trained exploiter; its 370/397/233 current comparison does not demonstrate paired mirror symmetry. Rule-unreadable regressed to current 0.4555 and Starter 0.9866666666666667, while its 0.998 counter score came from a search that learned nothing. Effective adversaries matter more than the apparent size of a reported resistance score. M1: rule-losses, rule-harden, rule-opening and rule-unreadable

The tournament should fly the ten ranked models and current as the common anchor. Prioritise sp-gradient-2, sp-curriculum-4, sp-distill-1, rule-defence, rule-constants and rule-mutual as candidates that already retain the Starter guard; keep the other four leaders as hard opponents and challengers needing repair. Add sp-league-5 once its exact export has run successfully in ONNX Runtime: its current 0.7065 and counter 0.4005 make it a useful comparison just outside the top ten, with the validation and counter-budget caveats above. This selection is a recommendation based on the main table and the reported league-5 result. M1: table M2: table and sp-league-5

For broader coverage, also fly sp-league-3, sp-league-2, sp-response-5, rule-losses, rule-plus-net and rule-gunnery, and use sparring-small as a specialist opponent. Import the effective exploiters of the finalists, especially the sp-gradient-5 response that also beats current, the independently initialised sp-starter-1 response, and the counters that distinguish current from sp-league-2 and sp-curriculum-4. The merges identify reproducible counter parameters in the original league-2, curriculum-4 and response-3 item results, but those parameters are not included in these write-up inputs. The tournament operator will need those original artifacts. This is a proposed roster, not a report of cross-seat matches already played. M1: corresponding item notes and conclusion M2: corresponding item notes and across-group discussion

For that round, freeze the exact ONNX exports, verify native/ONNX agreement where native search was used, and use shared fresh paired spawns with side reversal. Run the requested Starter, built-in and current samples, plus cross-play against the retained counters; report kills, HP wins, losses and draws separately. Give each finalist the same ten-minute fresh-counter budget and evaluate both finalist and current against that counter on the same seeds. Record unsuccessful searches as unsuccessful. These are recommended controls motivated by seed-sensitive gunnery, native-only final evaluations, missing ONNX execution and the demonstrated matchup cycles. A promotion should require preserved Starter performance and stronger results over that varied common pool; these local figures cannot predict wins against the hidden board entries. The requester reserves any eventual upload for the operator. M1: rule-gunnery, rule-plus-net and conclusion M2: sp-league-2, sp-league-5, sp-curriculum-4, sp-response-5 and conclusion

Several evidence qualifications remain unresolved. M1 flags external-source problems for harness-losses, size-and-ops, rule-losses and sp-starter-2; M2 flags sp-league-3, sp-curriculum-4, sp-evolution-5 and sp-response-3. The failed URLs/passages and separate source lists were not supplied, so those issues cannot be independently resolved here. Their local records remain attributed reports. The underlying JSON and replay logs are also unavailable for independently checking seed pairing, outcome breakdowns or claimed causal explanations. M1: provenance warning M2: affected item notes

Where source summaries disagree with details, this report follows the detailed records. Rule-losses' narrative BEST figures are 0.99833 and 0.570, while its structured final Starter/current records are 0.9986666666666667 and 0.5645. Rule-opening's structured value is 0.9888333333333335, while its narrative says 0.9888333333333334. Sp-distill-2's recorded counter selection conflicts with its stronger confirmation variant, as discussed above. M2's blanket “10-minute exploiter” summary does not describe sp-league-5's reported 480 s search. M2 questions sp-evolution-3 partly by comparison to the historical neural-loss anecdote; M1 assigns that anecdote to n7. The independent reason to exclude sp-evolution-3's figures is that it reports no working binary and no matches. M1: rule-losses, rule-opening, sp-distill-2 and sparring-head M2: sp-evolution-3, sp-league-5 and across-group discussion

Metadata is similarly distinguished from performance. Size-and-ops, sp-curriculum-3 and sp-evolution-3 returned current with incorrect or missing operator lists; the merges identify their actual operator set as R. Sp-gradient-3 and sp-response-4 returned no model. M2 also reports that some item hashes describe base64 text rather than decoded ONNX bytes, so hash-domain differences alone do not establish different models. None of those records supplies a new tested contender. The merges' metadata checks do not certify execution or reproduce match outcomes. M1: size-and-ops and metadata check M2: metadata check and blocked rows

Source identities for audit: M1 is artifacts/goal/merges/01/result.md, supplied as .imd/reads/artifacts/merge-01, SHA-256 fac1751cd097737c20393fbb9d6a5d2d08e00829153d98a838684f2c10d11306; M2 is artifacts/goal/merges/02/result.md, supplied as .imd/reads/artifacts/merge-02, SHA-256 b523ae7e6abf69d7b28da09e74dcc80379fc4cece2ad5334d8241f2f447c49d9. Both supplied files matched these hashes during this write-up. Item names in citations identify the rows and notes within those merges; original item results were not independently read. The requester's snapshot is the source for board standing, the rating fit and upload policy. The puzzle page and public current-entry artifact identify the challenge and incumbent; this report uses the merges' incumbent measurements and does not claim a fresh live-board verification. No models or weights are included.