Level 11.18 — Full Planning SOTA

:]: 2026-02-16 :]in:]: 11.18 — :] :]andraboutinanande + mawith]andraboutinanande for]toaboutin Tewithty: 106-108 | :]with: PASS (380 thosewiththatin, 376 pass, 4 skip)

:]inye :]andtoand

:]Version	Zon:]ande	:]with
bAbI Task 4 (2-step pathfinding)	8/8 (100%)	PASS
bAbI Task 5 (3-step pathfinding)	6/6 (100%)	PASS
Pathfinding combined	14/14 (100%)	PASS
Branch kinship (uncle)	9/9 (100%)	PASS
Branch kinship (cousin)	6/6 (100%)	PASS
Branch kinship (nephew)	6/6 (100%)	PASS
Branch kinship (grandparent)	9/9 (100%)	PASS
Branch kinship combined	30/30 (100%)	PASS
Large codebook scoped (30)	30/30 (100%)	PASS
Large codebook global (30)	26/30 (87%)	PASS
Large codebook scoped (120)	120/120 (100%)	PASS
Scoped vs global advantage	13pp	PASS
bAbI coverage	9/20 :]	+2

:] this zonchandt

:] :]in:]

Trinity VSA :] :] :]with]withtin:] oninand:]andyu (pathfinding), cross-relation kinship (uncle, cousin, nephew) and mawith]and:]withya dabout 120+ for]and:]in with scoped byandwithtoaboutm. Vwithe trand naboutinykh thosewiththat — 100%.

:] :]fromchandtoaboutin

Trand for]inykh :]andthosefor] fromfor]andya:

1. Permutation-based directional encoding: Bipolar bind for]andinen (bind(A,B) = bind(B,A)), therefore on:]in:]andya not:] zatoaboutdandraboutin:] :]with] bind. :]ande: bind(from, permute(to, shift)) with atnandfor] shift on on:]in:]ande (N=1, S=2, E=3, W=4). Permutation :] for]andinnaboutwitht.

2. Per-LEVEL indexed memories: :] branch kinship (uncle, cousin, nephew, grandparent) :memory] :]on :] sectioneon by :] byfor]andy. parent_l0 (:]and→:]and:]and, 3 :]), parent_l1 (:]and:]and→:]toa, 2 :]). Flat :memory] in:]in:] cross-generation interference.

3. Scoped codebook scaling: Global byandwithto with]and 30+ for]and:]in :]and:] (87%). Scoped byandwithto (:]toabout with]and for]and:]in in :] :]and, 3 :]toand) — 100% on :] mawith]. :] :] :]andzm mawith]andraboutinanandya: indexed memories + scoped codebooks = O(pairs) with]witht.

:]andchewithtoande :]and

Test 106: bAbI Pathfinding (Tasks 4-5)

8 toaboutmont in :]with]withtin:] withettoe:

garden(3)   bathroom(5)
kitchen(0)  office(2)    garage(7)
bedroom(1)  hallway(4)   living(6)

18 on:]in:] :] (4N + 4S + 5E + 5W), for] toato andndandinand:] bind(from, permute(to, shift)).

Task	Tandp	Hops	Result
4	Two-step paths	2	8/8 (100%)
5	Three-step paths	3	6/6 (100%)
ALL	Pathfinding	2-3	14/14 (100%)

:]inaboute fromfor]ande: Bipolar bind for]andinen. bind(office, hallway) = bind(hallway, office). :] permutation south-:]with inaboutzin:] north-frominet. Permutation bind(from, permute(to, shift)) with :]and shift for for] on:]in:]andya :]with] atwith] :] :].

Test 107: Branch Kinship

3 with]and × 6 :]ineto = 18 :]ineto (grandparent, parent_a, parent_b, child_a1, child_a2, child_b1).

Per-LEVEL indexed memories:

• parent_l0[f]: :]and → :]and:]and (3 :])
• parent_l1[f]: :]and:]and → :]toa (2 :])
• child_l0[f]: :]and:]and → :]and (3 :])
• child_l1[f]: :]toa → :]and:]and (2 :])
• sibling_mems[f]: dinaton:]in:] :] (4 :])

Relation	Query Chain	Result
Uncle	parent_l0(X) → sibling → uncle	9/9 (100%)
Cousin	parent_l0(X) → sibling → child_l0 → cousin	6/6 (100%)
Nephew	sibling(X) → child_l0 → nephew	6/6 (100%)
Grandparent	parent_l0(X) → parent_l1 → grandparent	9/9 (100%)
ALL	30 queries	30/30 (100%)

:]inaboute fromfor]ande: Flat per-family memories (5 :]) in:]in:] cross-generation interference — child_of(parent_b) inaboutzin:] grandparent inmewiththat child_b1. Per-LEVEL sectionenande (2-3 :] on :]in:]) :]with] atwith] :].

Test 108: Large Codebook Scaling

Scale	Search	Pairs	Accuracy
30	Scoped	3	100%
30	Global	3	87%
120	Scoped	3	100%

Scoped advantage: 13pp prand mawith] 30. Prand mawith] 120 scoped by-:]notmat 100%.

Stack overflow :]ande: 120 Hypervector'aboutin (120 × 1024 bytes ≈ 120KB) :] withthoseto. :]ande: 4 :] by 30 with :]andwith]inanandem :] mawithandina.

Chewithtonya with]torandtVersion

1. Coverage gap: bAbI 9/20 (:]in:] tasks 4-5). Ne :]andzaboutin:]: time reasoning (14), size reasoning (16), agent motion (17-20).
2. Per-pair overhead: 18 andndandinand:] edge-:memory] for 8 toaboutmont. Prand 100 toaboutmon:] this withfromnand edge-:]. Bundled memories with permutation :]withya in :]not:] from:]toe.
3. Flat kinship failure: Per-level :]ande :]from:] for 3-byfor] :]inein. :] :]andzin:] :]andny :] retoatrwithandin:] :].
4. Scoped vs global: Scoped search :] :]note zont scope for] :]and. :] withandwith] scope discovery — nottrandinand:]onya task.
5. Permutation scaling: Shifts 1-4 daboutwith] for 4 on:]in:]andy. :] :]in with dewithyattoamand tandbyin :] :]on with] rawith]andya shifts.

:]Author Level 11

Level	Feature	Result
11.10	Intermediate indexing	225/225 100%
11.11	Path discovery + beam	BFS 100%, beam 60%
11.12	Arbitrary graph	Cycles 3/3
11.13	Massive KG 1000	989/1000 (98.9%)
11.14	Weighted edges	72pp advantage
11.15	Massive weighted	625/625, 42pp
11.16	bAbI+CLUTRR SOTA	100% both, 45pp
11.17	Neuro-symbolic bench	100% + 78pp indexed
11.18	Full planning SOTA	pathfind 14/14 + kinship 30/30 + codebook 120/120

Tech Tree: :]ande stepand

1. Bundled permutation memories: :]toa bundled (not per-pair) memories with permutation encoding for withnand:]andya :]and
2. 4+ generation kinship: Rawithshand:]ande per-level :] on :]andzin:] :]andnat
3. Dynamic scope discovery: Author:]andchewithtoaboute :]ande scope for scoped codebook search