This is a merge of all 25 supplied item results, with models and weight vectors omitted. The numbers below are the items’ reports, not a rerun of their matches. Every row and item note cites its supplied result; links identify the declared `artifacts/goal/items/<key>/result.json` source, read here from `.imd/reads/artifacts/item-<key>`.

All scores are from the returned model’s side. The four record columns preserve `records.starter`, `records.builtins`, `records.current`, and `records.exploiter`; the separate calibration column preserves `current_vs_starter`, which can come from a different batch. `s` is score per match under the reported board formula `(elimination wins + 0.75*HP wins + 0.5*draws)/matches`; `n` is matches, and current records show wins/losses/draws in that order. The built-in column is the supplied aggregate over the four opponents, not a claim that every opponent received that score. See each item’s `approach`, `tried`, and `records.exploiter.how` for its protocol; sample sizes are not interchangeable with training budgets. [harness-losses][harness-losses] [sparring-small][sparring-small]

† The requester flags one external source each for **harness-losses**, **size-and-ops**, **rule-losses**, and **sp-starter-2** as unreadable, changed, or not supporting the cited passage. The supplied inputs do not identify the failed URLs/passages, and the separate `sources.json` mentioned by harness-losses is not supplied, so those external citations cannot be resolved from this evidence; affected findings remain attributed local reports and are not independently corroborated source facts. This provenance warning comes from the assignment; it does not establish that a particular experimental number is false. [harness-losses][harness-losses] [size-and-ops][size-and-ops] [rule-losses][rule-losses] [sp-starter-2][sp-starter-2]

Byte sizes and operator lists in the table are copied from `entry_bytes` and `ops` exactly, including the erroneous size-and-ops list. `null` means not reported; for the blocked row, the stated sample sizes are nominal requirements, not completed matches.

| Item key / source | Returned model / status | Calibration: current vs Starter | Final vs Starter | Final vs four built-ins | Final vs current: W/L/D; score | Final vs own exploiter | Bytes | ONNX operators (as reported) |
|---|---|---|---|---|---|---|---|---|
| [item-harness-speed][harness-speed] | Current analytical rule **unchanged**; speed audit | 0.992 | s=0.993; n=3000 | s=0.99875; n=800 | 410/410/180; s=0.5; n=1000 | s=0.17; n=500; 44-constant rule search; strongest of rule/neural searches | 18930 | Abs, Add, Atan, Cast, Concat, Cos, Div, Gather, Greater, Max, Min, Mul, Neg, Sin, Split, Sqrt, Sub, Where |
| [item-harness-losses][harness-losses] | Current rule **unchanged**; loss/replay audit † | 0.9923333333333333 | s=0.9923333333333333; n=3000 | s=1.0; n=800 | 406/406/188; s=0.5; n=1000 | s=0.259; n=500; 33-parameter rule hill-climb; 600 seconds | 18930 | Abs, Add, Atan, Cast, Concat, Cos, Div, Gather, Greater, Max, Min, Mul, Neg, Sin, Split, Sqrt, Sub, Where |
| [item-sparring-head][sparring-head] | Rule + zero-started residual MLP; 56 observation values + 3 rule actions, 64 hidden | 0.9895 | s=0.9745; n=3000 | s=0.99875; n=800 | 810/154/36; s=0.828; n=1000 | s=0.269; n=500; Fresh rule-residual ES; 545 s | 35530 | Abs, Add, Atan, Cast, Concat, Cos, Div, Gather, Gemm, Greater, Max, Min, Mul, Neg, Relu, Sin, Slice, Split, Sqrt, Sub, Where |
| [item-sparring-small][sparring-small] | Ceiling-camping 56-16-3 ReLU specialist | 0.9923333333333333 | s=0.20925; n=3000 | s=0.003125; n=800 | 603/395/2; s=0.60375; n=1000 | s=0.0; n=500; Parallel small-net/head searches; 600-second budget; small net selected | 4677 | Gather, Gemm, Relu |
| [item-size-and-ops][size-and-ops] | **BLOCKED**; current rule **unchanged** † | 0.9933333333333333 (inherited; not measured) | s=0.9933333333333333; n=3000 (inherited; not run) | s=null; n=800 (not run) | null/null/null; s=null; n=1000 (not run) | s=null; n=500; not trained; nominal sample only | 18930 | Add, Cast, Clip, Concat, Constant, Div, Gemm, Mul, Relu, Sub |
| [item-rule-losses][rule-losses] | League-retuned rule constants; two inactive structural additions † | 0.991 | s=0.9986666666666667; n=3000 | s=0.999375; n=800 | 521/392/87; s=0.5645; n=1000 | s=0.212; n=500; Strongest of rule/mirror/neural tests; selected rule trained 120 s | 20336 | Abs, Add, Atan, Cast, Concat, Cos, Div, Gather, Greater, Max, Min, Mul, Neg, Sin, Split, Sqrt, Sub, Where |
| [item-rule-harden][rule-harden] | Current rule **unchanged**; hardening incomplete | 0.9894999999999999 | s=0.9894999999999999; n=3000 | s=0.99875; n=800 | 396/396/208; s=0.5; n=1000 | s=0.994; n=500; Weak Starter-head ES; 7 minutes; no robustness evidence | 18930 | Mul, Greater, Add, Sub, Cast, Where, Div, Max, Cos, Sin, Abs, Min, Sqrt, Neg, Atan, Gather, Split, Concat |
| [item-rule-gunnery][rule-gunnery] | Rule with six lead/fire constants retuned | 0.9903333333333333 | s=0.9933333333333333; n=3000 | s=0.99875; n=800 | 447/378/175; s=0.5345; n=1000 | s=0.503; n=500; Six-constant shooting search; 600 candidates; ten-minute duration not documented | 18930 | Abs, Add, Atan, Cast, Concat, Cos, Div, Gather, Greater, Max, Min, Mul, Neg, Sin, Split, Sqrt, Sub, Where |
| [item-rule-energy][rule-energy] | Current rule **unchanged**; speed overrides rejected | 0.9883333333333333 | s=0.9883333333333333; n=3000 | s=1.0; n=800 | 404/404/192; s=0.5; n=1000 | s=0.838; n=500; **Untrained substitute**: current rule at full throttle | 18930 | Abs, Add, Atan, Cast, Concat, Cos, Div, Gather, Greater, Max, Min, Mul, Neg, Sin, Split, Sqrt, Sub, Where |
| [item-rule-defence][rule-defence] | Three-way gated defence rule + weaving + late fallback | 0.9895 | s=0.9975; n=3000 | s=1.0; n=800 | 739/184/77; s=0.7775; n=1000 | s=0.273; n=500; Rule/neural best responses; selected rule; reported total 603.5821514129639 seconds | 31253 | Abs, Add, Atan, Cast, Concat, Cos, Div, Gather, Greater, Max, Min, Mul, Neg, Sin, Split, Sqrt, Sub, Where |
| [item-rule-opening][rule-opening] | Current rule **unchanged**; opening audit | 0.9888333333333335 | s=0.9888333333333335; n=3000 | s=1.0; n=800 | 370/397/233; s=0.4865; n=1000 | s=0.494; n=500; **Untrained substitute**: identical current mirror; no ten-minute search | 18930 | Abs, Add, Atan, Cast, Concat, Cos, Div, Gather, Greater, Max, Min, Mul, Neg, Sin, Split, Sqrt, Sub, Where |
| [item-rule-mutual][rule-mutual] | Fixed rule retuned with mutual-kill penalties | 0.994 | s=0.9966666666666667; n=3000 | s=0.99875; n=800 | 788/159/53; s=0.8145; n=1000 | s=0.06; n=500; Concurrent rule/head searches; 595.2301568984985 seconds; rule selected | 18930 | Abs, Add, Atan, Cast, Concat, Cos, Div, Gather, Greater, Max, Min, Mul, Neg, Sin, Split, Sqrt, Sub, Where |
| [item-rule-constants][rule-constants] | Same rule graph; diverse-pool constant tuning | 0.9913333333333333 | s=0.9963333333333333; n=3000 | s=1.0; n=800 | 667/227/106; s=0.72; n=1000 | s=0.293; n=500; Rule/head search; 591.6773280350026 seconds; rule selected | 18930 | Abs, Add, Atan, Cast, Concat, Cos, Div, Gather, Greater, Max, Min, Mul, Neg, Sin, Split, Sqrt, Sub, Where |
| [item-rule-plus-net][rule-plus-net] | Rule + 15-feature linear residual at 1% scale | 0.991 | s=0.9903333333333333; n=3000 | s=1.0; n=800 | 470/317/213; s=0.5765; n=1000 | s=0.568; n=500; 10-minute residual search; no mutation surpassed starting current rule | 19512 | Abs, Add, Atan, Cast, Concat, Cos, Div, Gather, Gemm, Greater, Max, Min, Mul, Neg, Sin, Split, Sqrt, Sub, Where |
| [item-rule-unreadable][rule-unreadable] | Rule + hashed yaw/throttle style mixer | 0.9895 | s=0.9866666666666667; n=3000 | s=1.0; n=800 | 369/458/173; s=0.4555; n=1000 | s=0.998; n=500; Weak Starter-head ES; 420 s; learned nothing | 19997 | Abs, Add, Atan, Cast, Concat, Cos, Div, Gather, Greater, Max, Min, Mul, Neg, Sin, Split, Sqrt, Sub, Where |
| [item-sp-starter-1][sp-starter-1] | Starter imitation → PPO league; FP16 224-64-64-3 | 0.9906666666666667 | s=0.9855833333333334; n=3000 | s=0.989375; n=800 | 936/59/5; s=0.938; n=1000 | s=0.034; n=500; Strongest of three PPO runs; 9 minutes; Starter-imitation start | 38122 | Cast, Gemm, Relu |
| [item-sp-starter-2][sp-starter-2] | Linear gate over two FP16 112-64-32-3 Starter experts † | 0.9913333333333333 | s=0.8214166666666667; n=3000 | s=0.935625; n=800 | 4/996/0; s=0.004; n=1000 | s=0.002; n=500; Rule-adjustment/neural adaptation; 570.8242718609981 seconds; neural selected | 39788 | Cast, Concat, Gather, Gemm, Greater, Relu, Slice, Where |
| [item-sp-starter-3][sp-starter-3] | Starter imitation → ES league; FP16 224-48-48-3 | 0.9926666666666667 | s=0.9570833333333333; n=3000 | s=0.924375; n=800 | 6/993/1; s=0.0065; n=1000 | s=0.008; n=500; Imitation-started ES; 9.5 minutes; narrow specialist | 27117 | Cast, Gemm, Relu |
| [item-sp-starter-4][sp-starter-4] | Starter imitation → ES league; FP32 224-32-32-3 | 0.991 | s=0.9585833333333333; n=3000 | s=0.020625; n=800 | 0/1000/0; s=0.0; n=1000 | s=0.052; n=500; Self-initialized ES; 7 minutes, short of requested budget | 33735 | Gemm, Relu |
| [item-sp-starter-5][sp-starter-5] | Starter imitation → ES; FP16 224-64-64-3 + LayerNorm | 0.9913333333333333 | s=0.8744166666666666; n=3000 | s=0.0265625; n=800 | 0/997/3; s=0.0015; n=1000 | s=0.123; n=500; Self-initialized ES; 200 s, short of requested budget | 39180 | Cast, Clip, Concat, Gemm, LayerNormalization, Relu, Slice |
| [item-sp-distill-1][sp-distill-1] | Rule within 150m; learned 47-32-32-3 pursuit controller outside | 0.9913333333333333 | s=0.99375; n=3000 | s=0.99875; n=800 | 802/163/35; s=0.8195; n=1000 | s=0.254; n=500; Three adaptation lineages + selection; 556.4780877099838 seconds | 31919 | Abs, Add, Atan, Cast, Concat, Cos, Div, Gather, Gemm, Greater, Max, Min, Mul, Neg, Relu, Sin, Split, Sqrt, Sub, Where |
| [item-sp-distill-2][sp-distill-2] | Rule blended with 1% FP16 224-44-20-3 student | 0.9933333333333333 | s=0.989; n=3000 | s=1.0; n=800 | 422/386/192; s=0.518; n=1000 | s=0.51; n=500; 17 output-bias variants; retained current rule; no timed ten-minute training documented | 41421 | Abs, Add, Atan, Cast, Concat, Cos, Div, Gather, Gemm, Greater, Max, Min, Mul, Neg, Relu, Sin, Split, Sqrt, Sub, Where |
| [item-sp-distill-3][sp-distill-3] | Second DAgger rule clone; FP16 224-48-48-3; ES rejected | 0.9926666666666667 | s=0.9511666666666667; n=3000 | s=0.88625; n=800 | 98/765/137; s=0.1665; n=1000 | s=0.068; n=500; Self-initialized ES; 420 s, short of requested budget | 27120 | Cast, Gemm, Relu |
| [item-sp-distill-4][sp-distill-4] | Rule clone + DAgger + ES; FP16 224-64-64-3 | 0.9903333333333333 | s=0.9865833333333334; n=3000 | s=0.9534375; n=800 | 498/466/36; s=0.516; n=1000 | s=0.074; n=500; Self-initialized ES; 570 s | 38026 | Cast, Gemm, Relu |
| [item-sp-distill-5][sp-distill-5] | Rule clone + ES; FP16 224-68-64-3 | 0.9913333333333333 | s=0.9808333333333333; n=3000 | s=0.99; n=800 | 338/500/162; s=0.419; n=1000 | s=0.079; n=500; Self-initialized ES; 560 s | 40310 | Cast, Gemm, Relu |

**[harness-speed][harness-speed].** Returned the current model unchanged and found inference, rather than physics, to be the bottleneck: rule-versus-Starter throughput was 41-45 matches a second under ONNX Runtime versus 656-671 with the rule compiled, while constant tuning held the current model to 0.17 against its fresh exploiter. It recommends native rule evaluation for search, ONNX validation for finalists, and a league containing that exploiter rather than selection against Starter alone.

**[harness-losses][harness-losses] †.** Returned the current model unchanged; its replay audit labels 18 of 23 Starter losses as opening geometry and 5 as failed finishing shots, while a one-HP dodge intervention reduced mirror mutual kills from 200 to 71 but lowered score to 0.4675 (†: these are item-reported diagnoses and experiments, without independently validated external sourcing). It recommends a time-to-impact versus fifth-hit-time gate, separate treatment of opening bursts, and fresh league tests; the diagnostic alone is not a net-positive replacement.

**[sparring-head][sparring-head].** A zero-started residual changed flying rather than gunnery and reached 0.828 against current, but Starter fell to 0.9745 and the fresh exploiter reduced it to 0.269; the item also says the historical 27 percent neural loss concerned the earlier n7 rule, not the final current rule. It recommends retaining the upload, using this model and its exploiter as league opponents from the start, guarding Starter performance, and teaching pursuit of evasive runners.

**[sparring-small][sparring-small].** The tiny network learned a ceiling-camping counter that scores 0.60375 against current but only 0.20925 against Starter and 0.003125 against built-ins, then loses all 500 fresh exploiter matches. It recommends retaining the current upload and adding this distinct specialist to a shared league, with whole-pool selection and fresh counter-training.

**[size-and-ops][size-and-ops] †.** **Blocked:** the native compiler/linker setup failed and dependency build scripts terminated with SIGSEGV, so there were no simulator measurements or size sweep; the current model was returned unchanged, and its 0.9933333333333333 Starter values are inherited rather than measured (†). It recommends rebuilding with a working native linker and completing every benchmark; its supplied `ops` list is also inconsistent with its own embedded model, as detailed below.

**[rule-losses][rule-losses] †.** The item reports that league retuning improved Starter and current matchups but exposed both a strong rule counter and neural-head counters; preventing stalls and firing earlier did not establish a useful fix (†: the physics/causal interpretation remains item-reported rather than externally corroborated). It recommends no upload on these numbers alone, full-network self-play opponents and neural exploiters in the league from the first round, with worst-case held-out selection; its `tried` BEST line gives Starter 0.99833 and current 0.570, whereas the structured records give 0.9986666666666667 and 0.5645, which are preserved in the table without treating the batches as identical.

**[rule-harden][rule-harden].** Returned the current model unchanged and did not complete the five exploit-and-patch rounds; its sole head-only exploiter scored only 0.006 for itself, so the final model's 0.994 does not demonstrate hardening. It recommends faster native or batched rule inference, a full-network neural exploiter with informative training reward, and replay-guided patches before repeating a league.

**[rule-gunnery][rule-gunnery].** Six-constant tuning produced a modest, seed-sensitive advantage: two 5,000-match current holdouts scored 0.5094 and 0.5103, versus 0.5345 on the final locked set; the locked Starter result was 2,980 elimination wins and 20 losses. It recommends cross-play before upload and a full-policy neural opponent rather than another restricted shooting-constant search; the supplied exploiter record does not document a ten-minute wall-clock run.

**[rule-energy][rule-energy].** Returned the current model unchanged because every tested fixed or conditional target-speed override lost to its existing throttle policy; the 0.838 exploiter record uses an untrained full-throttle rule substitute, so it does not satisfy the requested trained-best-response comparison. It recommends preserving the throttle logic and investigating aim and mutual-kill avoidance, while acknowledging that its speed comparisons used only 80 to 120 matches per variant.

**[rule-defence][rule-defence].** The gated/weaving defender reached 0.9975 against Starter and 0.7775 against current, with hits taken against current falling from 4.519 to 4.171 on paired seeds, but a fresh rule best response reduced its score to 0.273. It recommends cross-play, adding the counter and full neural policies to the league, and preserving Starter reliability; the reported selected-search total of 603.5821514129639 seconds slightly exceeds a strict ten-minute limit, and the parallel neural search reports 626.6040420532227 seconds.

**[rule-opening][rule-opening].** Returned the current model unchanged because no opening change passed independent validation, and no trained exploiter was completed; its current comparison is asymmetric despite the identical model (370 wins versus 397 losses), so paired-seed mirror symmetry is not demonstrated by that record. It recommends distinct opening-focused exploiters and fresh Starter/current holdouts; the table preserves `records`/`current_vs_starter` value 0.9888333333333335, while its `tried` text instead says 0.9888333333333334.

**[rule-mutual][rule-mutual].** Retuning reduced mutual kills but did not remove them: the final current matchup had 53 mutual kills, its mirror had 46 in 1000, and the strongest fresh adversary drove its score to 0.06 despite 0.8145 against current. It recommends cross-play, retaining the rule counter as a regression opponent, shared-league training with fresh exploiters, and pre-emptive spacing in the last-HP exchange rather than inferring board strength from Starter gains.

**[rule-constants][rule-constants].** Diverse-pool tuning selected balanced-85 with final scores 0.9963333333333333 against Starter and 0.72 against current, but the fresh best response still held it to 0.293; an ablation implicated wrap-related constants, and a diagnostic loss was four hits versus the counter's five. It recommends cross-play with the reconstructed counter, investigation of first-shot timing and post-shot survival, full-network self-play, and learned wrap decisions.

**[rule-plus-net][rule-plus-net].** A small linear correction selected at 1% scale reached 0.5765 against current, but Starter scored 0.9903333333333333; the ten-minute counter search failed to improve its starting current-rule opponent, so 0.568 tests that retained baseline rather than a newly effective counter. It recommends a league with self-play snapshots and independent exploit styles, followed by fresh-seed validation.

**[rule-unreadable][rule-unreadable].** The style mixer regressed to 0.4555 against current and 0.9866666666666667 against Starter; its 0.998 against a weak head-only exploiter cannot show improved resistance because that search learned nothing against either version. It recommends no replacement upload, faster inference for a meaningful exploitation comparison, and styles implemented through tactical rule parameters rather than output scaling.

**[sp-starter-1][sp-starter-1].** Starter imitation plus PPO produced the group's highest reported current-matchup score, 0.938, but Starter remained at 0.9855833333333334 and an imitation-started nine-minute exploiter reduced the model to 0.034. It recommends cross-play first, routinely training exploiters from independent imitation starts rather than only warm starts, retaining them at high league weight, and selecting snapshots on large fresh samples.

**[sp-starter-2][sp-starter-2] †.** The gated experts are not a replacement: they scored 0.004 against current and 0.002 against their selected fresh exploiter, despite 0.935625 against built-ins (†: training/export and compression claims are attributable to the item, without independent external corroboration). It recommends keeping the guards, extracting the draw-heavy expert as a league opponent, and testing a larger policy over all observation frames with policy gradients before distillation; the reported matching FP32/FP16 comparison is a local experiment, not a guarantee from a validated external source.

**[sp-starter-3][sp-starter-3].** ES improved Starter performance but balanced-league training washed out a narrow current-rule counter, leaving the final model at 0.0065 against current and 0.008 against its exploiter; the exploiter itself scored only 0.018 against Starter, showing non-transitive specialization. It recommends starting from current-rule distillation, adding fresh exploiters with loss-focused league weighting, trying more evaluation matches or PPO, and choosing by a fresh-seed round robin.

**[sp-starter-4][sp-starter-4].** The Starter-clone ES model reached 0.9585833333333333 against Starter but 0.020625 against built-ins and 0.0 against current; its exploiter was trained for seven minutes rather than ten. It recommends no upload, built-ins in the league from the start, partial reward against difficult opponents, and a stronger distilled teacher or wider half-precision clone.

**[sp-starter-5][sp-starter-5].** The narrowed LayerNorm clone overfit to Starter, scoring 0.0265625 against built-ins and 0.0015 against current; the current rule was omitted from its training pool, and its own exploiter ran for only 200 s. It recommends no upload, including current and built-ins with loss-based weights from the first generation, and preferably distilling current before self-play.

**[sp-distill-1][sp-distill-1].** Pure cloning underperformed, but preserving the rule inside 150m while learning pursuit outside it yielded 0.99375 against Starter and 0.8195 against current; a fresh native-neural counter still reduced the exported model to 0.254. It recommends cross-play, regeneration of both gated and ungated neural exploiters, tests targeting the 150m handoff, and inspection of the reported Ace loss while preserving the evasion fix.

**[sp-distill-2][sp-distill-2].** The 1% student blend only slightly exceeded a mirror against current (0.518) and fell below its measured current baseline against Starter; its `exploiter.how` selects zero bias at final-model score 0.51 even though `tried` reports 0.49 against the +1 shoot variant, so the table retains the recorded choice and flags that it was not the strongest by that confirmation score. It recommends DAgger against diverse neural opponents and a shared league, followed by paired fresh 3,000-match comparisons.

**[sp-distill-3][sp-distill-3].** The second DAgger clone remained weaker than its teacher (0.1665 against current), and its mixed-pool ES phase was rejected because it damaged that matchup; even high action-label agreement did not preserve fight outcomes. It recommends no upload, more capacity through fewer input frames or geometry features, greater weight on shooting and discrete yaw decisions, more student-driven DAgger, rule-heavy fine-tuning, and adding the self-seeded seven-minute counter to training.

**[sp-distill-4][sp-distill-4].** DAgger plus ES brought the clone close to a current-rule tie (0.516), but built-ins scored 0.9534375 with 93 HP wins, and a self-seeded 570 s counter reduced it to 0.074. It recommends no upload, alternating training with new exploiters in a growing league, fresh-exploiter selection, small HP shaping, and explicit pressure for elimination wins.

**[sp-distill-5][sp-distill-5].** Repeated DAgger did not fix sharp rule-action switches, while mixed-pool ES improved general performance but left the final model below current and vulnerable to its self-seeded counter (0.079). It recommends no upload, more capacity or explicit rule geometry, pruning observation history to widen the policy, and longer league training with earlier clones and exploiters retained.

An in-memory byte and protobuf-field check of the supplied `entry` values confirmed that **harness-speed, harness-losses, size-and-ops, rule-harden, rule-energy, and rule-opening** contain identical model bytes; every `entry_bytes` field matches its decoded file length. The size-and-ops model actually contains `Abs, Add, Atan, Cast, Concat, Cos, Div, Gather, Greater, Max, Min, Mul, Neg, Sin, Split, Sqrt, Sub, Where`, agreeing with the unchanged rule’s other reports and contradicting its own table list; all other supplied operator lists match their embedded graphs as sets. This checks artifact metadata only, without executing or certifying the models. [harness-speed][harness-speed] [harness-losses][harness-losses] [size-and-ops][size-and-ops] [rule-harden][rule-harden] [rule-energy][rule-energy] [rule-opening][rule-opening]

For the next cross-play round, an inference from the reported trade-offs is to include the stronger rule variants, the pursuit hybrid, the PPO policy, and the distinct residual/ceiling specialists together with their counters; a high current-matchup score alone does not select a replacement. The fresh exploitation failures and incomplete comparisons here do not establish a clearly safer upload or a top-three finish against unavailable private policies, so the evidence supports keeping the current upload while conducting that round. [rule-defence][rule-defence] [rule-constants][rule-constants] [rule-mutual][rule-mutual] [sp-distill-1][sp-distill-1] [sp-starter-1][sp-starter-1] [sparring-head][sparring-head] [sparring-small][sparring-small]

[harness-speed]: ../../items/harness-speed/result.json "Supplied item-harness-speed result.json"
[harness-losses]: ../../items/harness-losses/result.json "Supplied item-harness-losses result.json"
[sparring-head]: ../../items/sparring-head/result.json "Supplied item-sparring-head result.json"
[sparring-small]: ../../items/sparring-small/result.json "Supplied item-sparring-small result.json"
[size-and-ops]: ../../items/size-and-ops/result.json "Supplied item-size-and-ops result.json"
[rule-losses]: ../../items/rule-losses/result.json "Supplied item-rule-losses result.json"
[rule-harden]: ../../items/rule-harden/result.json "Supplied item-rule-harden result.json"
[rule-gunnery]: ../../items/rule-gunnery/result.json "Supplied item-rule-gunnery result.json"
[rule-energy]: ../../items/rule-energy/result.json "Supplied item-rule-energy result.json"
[rule-defence]: ../../items/rule-defence/result.json "Supplied item-rule-defence result.json"
[rule-opening]: ../../items/rule-opening/result.json "Supplied item-rule-opening result.json"
[rule-mutual]: ../../items/rule-mutual/result.json "Supplied item-rule-mutual result.json"
[rule-constants]: ../../items/rule-constants/result.json "Supplied item-rule-constants result.json"
[rule-plus-net]: ../../items/rule-plus-net/result.json "Supplied item-rule-plus-net result.json"
[rule-unreadable]: ../../items/rule-unreadable/result.json "Supplied item-rule-unreadable result.json"
[sp-starter-1]: ../../items/sp-starter-1/result.json "Supplied item-sp-starter-1 result.json"
[sp-starter-2]: ../../items/sp-starter-2/result.json "Supplied item-sp-starter-2 result.json"
[sp-starter-3]: ../../items/sp-starter-3/result.json "Supplied item-sp-starter-3 result.json"
[sp-starter-4]: ../../items/sp-starter-4/result.json "Supplied item-sp-starter-4 result.json"
[sp-starter-5]: ../../items/sp-starter-5/result.json "Supplied item-sp-starter-5 result.json"
[sp-distill-1]: ../../items/sp-distill-1/result.json "Supplied item-sp-distill-1 result.json"
[sp-distill-2]: ../../items/sp-distill-2/result.json "Supplied item-sp-distill-2 result.json"
[sp-distill-3]: ../../items/sp-distill-3/result.json "Supplied item-sp-distill-3 result.json"
[sp-distill-4]: ../../items/sp-distill-4/result.json "Supplied item-sp-distill-4 result.json"
[sp-distill-5]: ../../items/sp-distill-5/result.json "Supplied item-sp-distill-5 result.json"