D7 grammar thinking styles + categorical-algebraic inference architecture

.claude/board/INTEGRATION_PLANS.md

@@ -36,6 +36,19 @@
 
 ---
 
+## v1 — Categorical-Algebraic Inference (authored 2026-04-21)
+
+**Author:** main-thread session 2026-04-21
+**Scope:** Meta-architecture document proving that parsing (Kan extension), disambiguation (free-energy minimization), learning (NARS revision), memory (AriGraph commit), and awareness (method-call history) are one algebraic operation — element-wise XOR on role-indexed slices of a 10K binary VSA vector — viewed through five lenses. Grounded in Shaw 2501.05368 (category theory) + 13 supporting papers. Does not replace elegant-herding-rocket — extends it with the categorical foundation.
+**Path:** `.claude/plans/categorical-algebraic-inference-v1.md` (496 lines)
+**Deliverables:** This plan produces no NEW D-ids. It grounds the existing D2/D3/D5/D7/D8/D10 deliverables from elegant-herding-rocket in the categorical-algebraic framework and establishes the five-lens litmus + object-does-the-work test as architectural invariants.
+
+**Status (2026-04-21):** Active. Companion to elegant-herding-rocket-v1, not a replacement.
+
+**Confidence (2026-04-21):** CONJECTURE on the Kan-extension-IS-free-energy equivalence. FINDING on all other claims (grounded in shipped code + paper proofs).
+
+---
+
 ## v1 — Codec Sweep via Lab Infra, JIT-first (authored 2026-04-20)
 
 **Author:** main-thread session 2026-04-20

.claude/board/STATUS_BOARD.md

@@ -117,7 +117,7 @@ early — CausalityFlow extension deferred). Plan path:
 | D2  | DeepNSM emits `FailureTicket` on low coverage | **Queued** | — |
 | D3  | Grammar Triangle wired into DeepNSM via `triangle_bridge.rs` | **Queued** | — |
 | D5  | Markov ±5 SPO+TEKAMOLO bundler with role-indexed VSA | **Queued** | — |
-| D7  | NARS-tested grammar thinking styles (meta-inference policies) | **Queued** | — |
+| D7  | NARS-tested grammar thinking styles + active-inference free-energy + RoleKey-as-operator | **In progress** | branch `claude/teleport-session-setup-wMZfb` — `thinking_styles.rs` (12 tests), `free_energy.rs` (7 tests), `role_keys.rs` bind/unbind/recovery_margin (12 tests incl 5-role lossless superposition), `divergence_from(prior)`, Finnish case patch |
 
 ### Phase 3 — Queued
 

.claude/knowledge/grammar-tiered-routing.md

@@ -364,10 +364,13 @@ patterns. Each case maps directly to a TEKAMOLO slot or an SPO role:
 ```
 Case          Ending     TEKAMOLO slot / Role
 ─────         ──────     ────────────────────
-Nominative    -∅         → Subject (S)
-Genitive      -n         → Possessor / Object of some verbs
-Accusative    -n/-t      → Object (O)
-Partitive     -a/-ä      → Partial object / negated object
+Nominative    -∅         → Subject (S) / Total object (plural)
+Genitive      -n         → Possessor / Total object (singular)
+Partitive     -a/-ä      → Partial / negated object
+Accusative    -n/-t      → Object — PERSONAL PRONOUNS ONLY
+                            (minut / sinut / hänet / meidät /
+                             teidät / heidät). NOT a general
+                             object marker (Latinate transplant).
 
 Inessive      -ssa/-ssä  → LO (in, inside)
 Elative       -sta/-stä  → LO (from inside)

.claude/knowledge/paper-landscape-grammar-parsing.md

@@ -0,0 +1,300 @@
+# Paper Landscape — Grammar Parsing × VSA × Active Inference
+
+> **READ BY:** integration-lead, truth-architect, family-codec-smith,
+> any agent touching deepnsm, grammar, AriGraph, or the free-energy
+> resolution pipeline.
+>
+> **Created:** 2026-04-21
+> **Scope:** Maps 14 recent papers onto the lance-graph grammar stack
+> (DeepNSM + RoleKey VSA + FreeEnergy active inference + AriGraph).
+> Each entry: citation, one-line finding, what it validates/challenges
+> in our architecture, and the specific code cross-reference.
+
+---
+
+## Tier 1 — Foundational (directly validates our algebraic substrate)
+
+### Shaw, Furlong, Anderson & Orchard (2501.05368) — VSA Category Theory Foundation
+
+**Finding:** Right Kan extensions prove that dimension-preserving
+binding/bundling MUST be element-wise operations. Division ring
+structure required for full reversibility. Co-presheaf generalization
+decouples index category (dimensional compression) from value category
+(ring structure).
+
+**Validates:**
+- `RoleKey::bind` (element-wise XOR on contiguous slices) is
+  categorically optimal — not a design choice, a theorem consequence.
+- XOR on GF(2)^d IS a division ring → full reversibility holds.
+- Our slice-addressing scheme (ℐ = disjoint intervals [0..2000),
+  [2000..4000), ...) is an instance of their index category with
+  monoidal product = disjoint union.
+
+**Key equations:**
+- Kan extension: `(Ran_e v⊗̄w)_i = ∫_{jk} ℐ(i,e(j,k)) ⋔ (v_j·w_k)`
+- Simplifies to element-wise: `v⊗w = ∫_i v_i · w_i`
+- Role-filler: `w = (first ⊗ v_1) ⊕ (second ⊗ v_2)` with
+  recovery `v_1 ∼ first ⊘ w` — our RoleKey::bind + unbind.
+- Braiding ρ for sequences: `list(x_1,...,x_n) = x_1 ⊕ ρx_2 ⊕ ρρx_3 ⊕ ...`
+  — this IS `vsa_permute` per position in the Markov bundler (D5).
+- Non-commutative binding needed for hierarchical structure — validates
+  why we use DIFFERENT role keys for S/P/O.
+
+**Cross-ref:** `contract::grammar::role_keys::{RoleKey::bind, unbind, vsa_xor}`.
+
+---
+
+### Kleyko, Davies, Frady, Kanerva et al. (2106.05268) — VSA/HDC Survey Part II
+
+**Finding:** VSA's algebraic structure enables "computing in
+superposition" — efficient solutions to combinatorial search via
+high-dimensional distributed representations. Computational
+universality established.
+
+**Validates:** Our XOR-superposition of N role bindings (tested at
+5 simultaneous roles recovering at margin 1.0) IS computing in
+superposition. The combinatorial search problem they describe =
+our counterfactual hypothesis enumeration in `Resolution::from_ranked`.
+
+**Cross-ref:** `contract::grammar::role_keys::vsa_xor`, `free_energy::Resolution`.
+
+---
+
+### Gallant & Okaywe (1501.07627) — MBAT: Objects, Relations, Sequences
+
+**Finding:** Matrix binding (MBAT) satisfies machine-learning
+constraints for VSA: similar structures → similar vectors. Phrases
+should be binding-sums. Three-stage learning: representation →
+association → inference.
+
+**Validates:** Our three-stage pipeline mirrors theirs:
+1. Representation = RoleKey::bind (content → role-indexed VSA)
+2. Association = Markov ±5 bundling (context accumulation)
+3. Inference = FreeEnergy resolution (hypothesis selection)
+
+Their "phrases as binding-sums" = our SPO triple as
+`SUBJECT_KEY.bind(s) ⊕ PREDICATE_KEY.bind(p) ⊕ OBJECT_KEY.bind(o)`.
+
+**Cross-ref:** Plan D5 `MarkovBundler`, `Trajectory`.
+
+---
+
+## Tier 2 — Empirical validation of the grammar tier
+
+### Graichen, de-Dios-Flores & Boleda (2601.19926) — "Grammar of Transformers" (337-article systematic review)
+
+**Finding:** TLMs handle formal syntax well (agreement >85% BLiMP)
+but show weak, variable performance on syntax-semantics interface
+(<75% on binding, coreference, quantifier scope, island effects).
+Severe English dominance (69%). Mechanistic methods underutilized.
+
+**Validates:** Our tiered routing — DeepNSM handles the >85% formal
+syntax locally; FreeEnergy + counterfactual resolves the <75%
+syntax-semantics interface. Their call for "syntax-semantics interface
+investigation + mechanistic methods" = exactly what our active-
+inference stack provides.
+
+**Cross-ref:** `contract::grammar::ticket::FailureTicket` (escalation
+for the <75% tail), `free_energy::Resolution`.
+
+---
+
+### Jian & Manning (2603.17475 / EACL 2026) — Abstraction-First Language Learning
+
+**Finding:** GPT-2 learns class-level verb behavior BEFORE item-
+specific behavior. Sequential emergence: syntactic subcategorization
+(t<100) → semantic argument structure (t>100) → non-local
+dependencies (t>1000). Count-based exemplar baseline is strictly
+worse.
+
+**Validates:**
+- `GrammarStyleConfig::nars.primary = Deduction` (class-level rules
+  first) IS the abstraction-first policy.
+- Sequential emergence maps to Markov radius scaling: ±1 captures
+  subcategorization, ±3 captures argument structure, ±5 captures
+  non-local. WeightingKernel::MexicanHat emphasizes local first.
+- Their 4 verb classes (to-dative / motion / reciprocal / spray-load)
+  = rows in our 144-verb taxonomy with characteristic TEKAMOLO priors.
+- Exemplar-first baseline fails = Markov bundling without role-key
+  structure is class-blind. Role keys ARE the abstraction mechanism.
+
+**Cross-ref:** `contract::grammar::thinking_styles::NarsPriorityChain`,
+`context_chain::WeightingKernel::MexicanHat`.
+
+---
+
+### Schulz, Mitropolsky & Poggio (2510.02524) — How LMs Learn CFGs
+
+**Finding:** KL divergence over PCFG decomposes as sum over
+subgrammar contributions (Theorem 4.3). Transformers learn all
+subgrammar levels in PARALLEL. Models FAIL on deep recursion
+despite handling long shallow contexts.
+
+**Validates:**
+- Our `FreeEnergy { likelihood, kl_divergence, total }` decomposition
+  mirrors their KL-over-subgrammars. Each role-key slice IS a
+  "subgrammar" in the VSA decomposition.
+- Recursion failure = why we use Markov ±5 contextual coherence
+  instead of recursive parsing. Deep recursion becomes "does this
+  nested structure cohere with ±5 context?" — a flat comparison.
+- Parallel subgrammar learning = our FSM handles all PoS categories
+  simultaneously.
+
+**Cross-ref:** `contract::grammar::free_energy::FreeEnergy`.
+
+---
+
+### Alpay & Senturk (2603.05540) — Grammar-Constrained LLM Decoding
+
+**Finding:** Doob h-transform: `p(v|y<t) = p(v|y<t) · h(y<tv)/h(y<t)`.
+Grammar survival probability modulates base LLM distribution.
+Structural Ambiguity Cost (SAC): right-recursive O(1)/token,
+concatenative Θ(t²)/token. Lower bound: Ω(t²) for parse-preserving
+engines.
+
+**Validates:**
+- Their grammar-conditional is the dual of our free-energy: both
+  are multiplicative modulations of a base distribution by structural
+  constraint.
+- SAC = our counterfactual branch count. Pearl 2³ mask reduces SAC
+  by committing causal bits from morphology.
+- Their Ω(t²) lower bound does NOT apply to us: we don't preserve
+  the full parse forest. Active inference commits to argmin_F and
+  discards (or marks epiphany). We trade parse-preservation for
+  decision speed.
+
+**Cross-ref:** `contract::grammar::free_energy::Resolution` (commit
+discards losers), `EPIPHANY_MARGIN` (preserves runner-up only when
+margin is tight).
+
+---
+
+## Tier 3 — Supporting evidence for specific design choices
+
+### Starace et al. (2310.18696, EMNLP 2023) — Joint Encoding of Linguistic Categories
+
+**Finding:** Related grammatical categories share overlapping
+encodings in LLMs; pattern holds cross-lingually.
+
+**Validates:** Role-key slice adjacency for morphologically-related
+cases (Finnish Adessive and LOKAL_KEY map to overlapping TEKAMOLO
+slots). Cross-lingual bundling works because categories are shared
+at the representational level.
+
+**Cross-ref:** `contract::grammar::role_keys::FINNISH_SLICES`,
+`contract::grammar::role_keys::LOKAL_KEY`.
+
+---
+
+### Tjuatja, Liu, Levin & Neubig (2305.18185) — Agentivity Probe
+
+**Finding:** Optionally transitive verbs test agent-vs-patient role
+assignment. GPT-3 outperforms corpus statistics.
+
+**Validates:** Pearl 2³ bit 0 = agency. Optionally transitive verbs
+= exact Wechsel case ("The door opened" vs "John opened the door").
+Their dataset = potential eval benchmark for `Resolution::resolve`.
+
+**Cross-ref:** `contract::grammar::ticket::CausalAmbiguity::plausible_mask`,
+`contract::grammar::free_energy::Hypothesis::causal_mask`.
+
+---
+
+### Petit, Corro & Yvon (2310.14124) — Supertagging + ILP
+
+**Finding:** Supertagging (per-token category) + integer linear
+program for structural consistency = compositional generalization.
+
+**Validates:** Our PoS tagging (supertag) + `TekamoloPolicy::require_fillable`
+(structural consistency). ILP = our Markov ±5 coherence (both prevent
+locally-plausible but globally-inconsistent parses).
+
+**Cross-ref:** `contract::grammar::tekamolo::TekamoloSlots`,
+`thinking_styles::TekamoloPolicy`.
+
+---
+
+### Sultana & Ahmed (2602.20749) — Grammar–Semantic Feature Fusion
+
+**Finding:** 11 explicit grammar features + frozen BERT = 2-15%
+improvement. Grammar as explicit inductive bias, not learnable module.
+
+**Validates:** Grammar-as-inductive-bias is the right framing. Their
+11 features are a shallow version of our TEKAMOLO slot-filling +
+SPO extraction. Full role-indexed VSA bundling should exceed their
+2-15% improvement substantially.
+
+**Cross-ref:** `contract::grammar::tekamolo`, `role_keys`.
+
+---
+
+### Shaikh, Ziems et al. (2306.02475, ACL 2023) — Cultural Codes
+
+**Finding:** Sociocultural background characteristics significantly
+improve pragmatic reference resolution.
+
+**Validates:** `GrammarStyleAwareness` as per-style empirical prior.
+Different thinking styles resolve the same ambiguity differently
+because their priors over signal-profile frequency differ — exactly
+the cultural-prior effect they measure.
+
+**Cross-ref:** `contract::grammar::thinking_styles::GrammarStyleConfig`.
+
+---
+
+### Perez-Beltrachini et al. (2301.12217) — Conversational Semantic Parsing
+
+**Finding:** Multi-turn QA grounded to SPARQL over large-vocab KGs.
+Challenges: entity grounding, conversation context, generalization.
+
+**Validates:** AriGraph triplet-graph + ContextChain = our equivalent.
+Their "conversation context" = our ±5 Markov chain. We don't need
+SPARQL because SPO triples are queried directly via
+`TripletGraph::nodes_matching`.
+
+**Cross-ref:** `arigraph::triplet_graph`, `grammar::context_chain`.
+
+---
+
+### Hussein (2602.14238) — CFG/GPSG Parser
+
+**Finding:** CFG+GPSG parser producing dependency + constituency
+trees; handles noise; UAS 54.5%.
+
+**Validates:** Our baseline to beat. Their noise tolerance =
+our `PartialParse` + `FailureTicket`. UAS 54.5% should be
+significantly exceeded by adding Markov coherence + role-key binding.
+
+**Cross-ref:** `contract::grammar::ticket::PartialParse`.
+
+---
+
+## The unclaimed intersection
+
+**No paper in this landscape combines:**
+
+1. Structural parsing (rule-based, not neural)
+2. Active-inference ambiguity resolution (free-energy, not attention)
+3. Role-indexed distributed representation (VSA with Kan-extension-
+   justified element-wise ops)
+4. NARS-revised epistemic awareness (per-parse revision, not gradient)
+
+Shaw et al. provide the algebraic foundation (Tier 1). Graichen
+et al. identify the target (syntax-semantics interface, Tier 2).
+Jian & Manning validate the dispatch order (abstraction-first, Tier 2).
+Alpay & Senturk formalize the grammar-conditional dual (Tier 2).
+
+Our stack sits at the intersection. The closest prior art is
+Shaw's category-theoretic VSA + Petit's supertagging+ILP, but
+neither has the active-inference free-energy loop or the NARS-
+revised epistemic awareness layer.
+
+---
+
+## Papers not yet fully retrieved
+
+- **biorxiv 2022.02.22.481380v3** — PDF too large for WebFetch.
+  Likely a neuroscience paper on VSA / neural binding.
+- **ResearchGate VSA-for-CFGs (Mitropolsky?)** — 403 forbidden.
+  This is likely the 2003.05171 paper already cited in the plan
+  (VSA encoding of Chomsky-normal-form CFGs via Fock space).