Corviknight - Great League IV Dive

1. Header

Corviknight -- Great League

SAND_ATTACK / AIR_CUTTER, IRON_HEAD (arm 1 of 6)

Level view: L50. Bake 2026-09-27. Blob 20260927_204247_Corviknight_great.replay.pkl.gz.

Opponent pool: 78 entries (Custom pool from gl_top50_plus_cs.txt (75 mons), incl. 2 moveset variant(s)). Grid: 4096 IV spreads x 9 shield scenarios.

Opponent-IV settings baked: PvPoke-default opponent IVs; PvPoke-default opponent IVs, focal never baits; stat-product rank-1 opponent IVs; stat-product rank-1 opponent IVs, focal never baits.

Rankings snapshot 2026-09-27.

2. Floor

Def >= 127.59 -- 1958 of 4096 spreads (47.8%).

At the one decimal place PvPoke and Poke Genie display this reads Def >= 127.5, which selects 2015 spreads where the exact value selects 1958: 57 more spreads land inside a line printed that way.

Primitive: EXACT clean cut -- every spread at or above the line wins the cell and no spread below it does.

Mechanism: bulkpoint -- Lapras's (4/11/12) PSYWAVE drops from 2 to 1 damage into us across the cut.

Owns: 0v0 Lapras (rank 39).

After collapsing pool variants, this cut is owned by one base species (Lapras, rank 39); if it leaves the pool the cut has no owner.

Below the cut: 0 of 2138 spreads win the cell, and they score 301-491. At or above: 1958 of 1958 win, scoring 507-703.

Direction: the cut points the same way in 4 of 4 opponent-IV settings and 6 of 6 moveset arms (win rate at or above it at least 90%, and above the rate below it).

Partition is the stricter question, asked of the SAME printed line: it splits the cell exactly (everything at or above the line wins, everything below it loses) in 2 of 4 settings and 4 of 6 arms. A partition implies the direction test, never the other way round, so this count can never exceed the one above it. Does not partition: rank1, rank1:nobait.

Separability is a THIRD and non-comparable measurement -- each view's OWN cut, at whatever value it sits: 4 of 6 arms and 4 of 4 settings separate this cell at some value. Per arm -- SAND_ATTACK / AIR_CUTTER, IRON_HEAD: 127.59 (1958 spreads); SAND_ATTACK / AIR_CUTTER, PAYBACK: no clean cut (46.6% win); SAND_ATTACK / IRON_HEAD, PAYBACK: 127.59 (1958 spreads); SAND_ATTACK / IRON_HEAD, SKY_ATTACK: 127.59 (1958 spreads); SAND_ATTACK / AIR_CUTTER, SKY_ATTACK: no clean cut (46.6% win); SAND_ATTACK / PAYBACK, SKY_ATTACK: 127.59 (1958 spreads).

The same cell's cut in each setting: pvpoke 127.59; pvpoke:nobait 127.59; rank1 126.09; rank1:nobait 126.09.

Wild catches needed to clear this line: 2 for a 50% chance, 3 for a 75% chance. Model: the count assumes uniform-random IVs over the whole grid, which is a model of a wild encounter and is not verified for any other source.

3. Lines on the other stats

Neither attack nor HP carries a line of its own on this arm: no threshold on either clears every gate with a pool inside [25%, 60%] of the grid.

4. Example spreads

Existence proofs; each selection rule is printed with its spread.

Cells won by shield scenario

Clearers reach the CP cap at a median level of L23.0 (1958 of 1958 at L45.0 or below) against L22.5 below the floor (2138 of 2138).

5. Rank-1 check

Stat-product rank-1 is 0/13/14 at L23.5: 105.56 attack / 132.75 Def / 152 HP. It wins 360 of 702 cells (82 of the 136 contested ones).

It already clears the floor; there is no trade to make.

It is not a member of the Alternative target below.

6. Alternative target

HP >= 140, Atk >= 112.15: 117 spreads (2.9%), stat product 90.2% to 94.2% of rank-1, attack 112.16-113.40.

It is built on HP and attack -- the two stats the printed defense line does not use.

3 of the 3 cells it wins with every member are won by at most 1% of the spreads outside the rectangle: 0v1 Araquanid (rank 16), 0v1 Carbink (rank 46), 1v1 Blastoise (rank 58).

Rule: among (HP, Atk) rectangles with at least 100 members and no floor clearer, the one winning the most contested cells the floor cannot (won by every member, by at most 1% of clearers); ties broken by member count. Only cells against a top-50 opponent in a non-degenerate scenario may select it. The printed cuts are the least attained values among the members.

Wins 3 cells with every member that the floor does not: 0v1 Araquanid (rank 16), 0v1 Carbink (rank 46), 1v1 Blastoise (rank 58).

Gives up 2 cells the floor guarantees: 0v0 Lapras (rank 39), 2v2 Lapras (rank 39).

No spread satisfies both this rectangle and the defense line.

Wild catches needed to land inside this rectangle: 24 for a 50% chance, 48 for a 75% chance. Model: the count assumes uniform-random IVs over the whole grid, which is a model of a wild encounter and is not verified for any other source.

Members win 310-339 cells in total.

IV envelope of the members -- attack IV 13-15, Def IV 0-9, HP IV 0-11. Members per attack IV: 13: 10, 14: 66, 15: 41.

Highest stat product inside it: 14/0/11 at L22.5 (94.17% of rank-1), 14/8/2 at L22.5 (94.09% of rank-1), 14/1/10 at L22.5 (94.02% of rank-1), 14/5/5 at L22.5 (93.99% of rank-1), 14/9/0 at L22.5 (93.89% of rank-1).

7. Cost

This test finds no reverse single-stat clean cut: no cell on this dive is won by every LOW spread and lost by every high one.

The cost as an exact cell diff -- 0/15/11 (most cells won among clearers with SP >= 95%) against rank-1: 8 cells gained, 6 lost.

Lost, top-20 opponents: 0v0 Corviknight (rank 7).

Gained, top-20 opponents: 2v1 Corsola (Galarian) (rank 8).

8. Rungs above

A rung is listed here when it keeps 25% of the grid and points the same way in at least 2 of the baked opponent-IV settings; the rest are in the tail sentence or the Dropped lines below. Being listed is not being eligible as a floor -- the floor gates are in the evidence block.

Beyond Def >= 128.21 there are 5 more clean rungs covering 6 cells, the highest at 134.81 (8 spreads); the widest of them keeps 12.3% of the grid (1v2 Carbink at 131.16).

Co-gates inside a rung: a second-stat condition that, among the spreads clearing that rung, wins the named cell every time. Sufficient, never necessary -- the last column is what the spreads without it still win. 7 further co-gates are not listed.

Co-gates inside a rung

9. Material clean cuts below the floor

Clean cuts holding between 10% and 90% of the grid, ascending. The widest of them sit under nearly every spread on the grid; the narrow ones at the bottom of the table do not.

10. Bulk

Clean cuts by axis: 24 on attack, 9 on Def, 0 on HP.

The bulk shape available inside the line -- the highest Def attained at each HP among the spreads clearing it:

Within the clearers, cells won in total run 321-362.

Co-gates inside the floor (every clearer in the sub-rectangle wins the cell; the rate outside it is printed beside): Atk >= 109.62 (295 clearers) wins 1v1 Annihilape rank 31, against 0.0% without it; Atk >= 109.62 (295 clearers) wins 2v1 Annihilape (Close Combat+Rage Fist) rank 31, against 0.1% without it; HP >= 143 (1608 clearers) wins 0v1 Hippowdon rank 50, against 31.1% without it; HP >= 144 (1434 clearers) wins 0v0 Snorlax (Shadow) rank 40, against 55.3% without it; HP >= 144 (1434 clearers) wins 1v0 Altaria rank 2, against 27.9% without it; HP >= 144 (1434 clearers) wins 1v1 Quagsire (Shadow) rank 12, against 48.3% without it; HP >= 144 (1434 clearers) wins 1v1 Snorlax (Shadow) rank 40, against 49.8% without it; HP >= 145 (1231 clearers) wins 0v0 Malamar rank 42, against 31.2% without it; HP >= 145 (1231 clearers) wins 0v1 Marowak rank 19, against 20.4% without it; HP >= 147 (779 clearers) wins 0v1 Araquanid (Shadow) rank 29, against 46.9% without it; HP >= 148 (514 clearers) wins 0v1 Vigoroth rank 21, against 9.8% without it; HP >= 149 (342 clearers) wins 1v1 Deoxys (Defense) rank 33, against 35.4% without it. These are sufficient conditions, never necessary ones.

11. Coverage

What each of Lapras's own 4096 spreads asks of our defense: the lowest attainable value that holds its PSYWAVE to 1 damage (from 2).

"held" is the share of that species' own grid whose attack is at or below the row's, and so is held by the same value: the damage it does to us is its attack over our defense, so its IV grid traces out the ladder the same way an attacker's does.

The column is not monotone, and that is a CP-cap effect rather than a transcription error: under the cap a hundo sits at a lower level than a low-attack spread, so it carries less attack. The highest-attack Lapras is 13/0/0.

"focal" is the count of our spreads at or above that value. The printed floor is the PvPoke-default row; it is the line every reference tool quotes, not a guarantee against every build of that species.

12. Mirror

Against the pool's own Corviknight entry. Seat: our side is the optimised row; the mirror opponent always baits.

Against a mirror built at stat-product rank-1 opponent IVs, the count of our spreads (of 4096) that win, by shield scenario 0v0 through 2v2: 60 0 0 4096 2062 0 4096 4096 1808.

Against a mirror built at stat-product rank-1 opponent IVs, focal never baits, the count of our spreads (of 4096) that win, by shield scenario 0v0 through 2v2: 60 0 0 4096 2062 0 4096 4096 1808.

13. Changes the score, not the matchup

Cells won (or lost) by every spread where the battle score still steps by at least 60 at one attainable attack value.

Ordered by how close the step comes to the win line (500), nearest first: the question this field answers is how close a cell is to becoming a matchup, not how large the swing inside it is.

12 of 50 steps are listed, closest to the win line first; 38 more are not.

14. Not claimed

103 of 136 contested matchups have no clean single-stat rule and are not claimed.

6 of 78 top-50 matchups are listed, one per base species and closest to an even split first; 72 more are not.

15. How sure

24 clean attack cuts, 9 on Def, 0 on HP, over 136 contested cells (278 cells are won by every spread, 288 lost by every spread).

Scenarios barred from carrying a floor -- fewer than 6 contested cells, or fewer than 8 distinct opponent columns among them; the dive page's "degenerate" flag applies the same two numbers to its own stricter counts, so it can flag a different set: 0v2 (2 contested, 3 distinct opponent columns), 2v0 (2 contested, 3 distinct opponent columns).

The printed floor partitions its cell exactly in 2 of 4 opponent-IV settings and 4 of 6 arms; its direction survives in 4 and 6 of them. Separately, and not comparably, the cell has a clean cut at SOME value in 4 settings and 4 arms.

It is owned by one base species (Lapras); if it leaves the pool the floor has no owner.

Both baked opponent cohorts sit low in Lapras's own attack grid (5.2% and 26.6% of it at or below them, and so held by the same value), so "2 of 4 settings" is not robustness against attack-weighted opponent builds; the coverage ladder is.

The level-51 view has its own cut for the same cell: 127.59 with 1958 spreads clearing, where the L50 literal applied to the L51 grid would select 1958.

Independent check: the dive page's matchup clusters, under "Why these regions", which partition this same grid by whole win/loss fingerprint across all 7 non-degenerate shield scenarios, splits it on attack (at 107.88), not defense -- the two methods do not agree on this dive (K=2, silhouette 0.44).

Not measured here: opponent IVs outside the baked cohorts, post-match HP and shields, XL or dust cost, and any live-game check of the priority convention.

16. Provenance

Every threshold in this section is one of three things, and each is badged where it is printed. An EXACT clean cut: every spread at or above it wins the named cell and no spread below it does. A ONE-SIDED GATE: one side is pure and the other is at least 97% pure -- either nothing below it wins (clearing it is necessary) or everything at or above it wins (clearing it is sufficient). A NEAR-EXACT split: at most 0.5% of the grid sits on the wrong side of it in total. The win predicate is gopvpsim.battle.is_win (score > 500; 500 itself is a tie, not a win).

That is stricter than the two numbers the dive page prints for the same opponent: the Threats "flips at" boundary (a 75/25 gate, so spreads on the passing side can still lose) and the Matchup clusters flip table's single-stat split for that one opponent (the most accurate such split, misclassifications allowed). The cluster line in "How sure" above is a third thing again: not one opponent but the stat rule that reproduces that section's "all scenarios" partition of the whole grid.

Priority ties: the coverage ladder counts an exact attack tie as NOT beaten. The engine is seat-dependent on an exact tie, so a tied line is not a guarantee in either direction.

Ranks and opponent builds are read live from the 2026-09-27 rankings snapshot and the current gamemaster; this blob carries no gamemaster stamp, so the render is a function of the blob plus those two live reads.

Auto-derived: this section is generated by scripts/deep_dive_brief.py from the replay blob. No human prose is in it.