research(deepnsm-v2): Haspelmath grammar harness on frozen 8:8:8 thought ABI

.arxiv/2026-10_haspelmath-deepnsm-ce64-harness.md

@@ -0,0 +1,171 @@
+# Haspelmath semantic-map → DeepNSM-v2 → CE64 → hydrate.rs
+**Research/executable harness · 15 October 2026 · not a production writer**
+
+## Thesis
+
+The substrate already has stable word identities, a multi-reading grammatical FSM, and
+verse addressing. Use Martin **Haspelmath** (not "Haspelmann") as a falsifiable
+**linguistic-function continuity constraint**, not as a new grammar parser, a semantic
+distance function, or an instruction set. The resulting pipeline is:
+
+```
+existing lexical source (COCA / UD / WordNet / language-aware morphology)
+  -> bind_tokens: (surface, PosSet) -> Reading (u16 WordId; OOV refused)
+  -> parse_readings: certain / alternative / ambiguity / overflow
+  -> [separate attested grammatical-form function masks]
+  -> hydrate(markers, per_verse_certain, basin): SPO + original verse path + order
+  -> CausalEdge64 syntactic operand view (8:8:8 fixed register/ISA contract)
+     + ONE externally pinned, immutable epoch-wide operand dictionary
+     + version-qualified canonical meaning for all readers
+  -> downstream authority/provenance/temporal writer: NOT IMPLEMENTED
+```
+
+This executable probe is in `crates/deepnsm-v2/src/haspelmath.rs`.
+
+**ABI correction (15 Oct):** The original draft mistakenly introduced a
+per-verse palette. That would make the same CE64 operand byte mean different
+things in two verses, defeating the intention of a globally interoperable
+8:8:8 register contract. The implementation now requires a single frozen,
+explicit, version-identified operand dictionary for the execution epoch.
+This is a test fixture only until the canonical ISA/codebook registry supplies
+and verifies that dictionary across independent clients. BF16 and TCP/IP are
+*interoperability analogies*, not claims that a numerical format alone grounds
+semantic truth or that every NAL/Hyperon operator is equivalent.
+It changes no existing grammar decision, hydration behavior, CE64 bit definition,
+or Lance/OGAR writer.
+
+## Primary source, not paraphrased into a stronger claim
+
+Martin Haspelmath, *The Geometry of Grammatical Meaning: Semantic Maps
+and Cross-Linguistic Comparison*, chapter 8 in *The New Psychology of Language,
+Volume 2* (2003), user-provided scan `SemMaps.pdf`, printed pp. 211–242.
+
+- **pp. 212–213 and Fig. 8.1**: "functions" is deliberately neutral between
+  different senses and context-dependent uses. English *to* spans direction,
+  recipient, experiencer, and purpose. The other dative functions in Fig. 8.1
+  are not necessarily uses of English *to*.
+- **p. 216, Fig. 8.3**: an edge indicates adjacency. Its drawing orientation
+  and physical length have no linguistic significance. Do NOT use Euclidean,
+  Morton or word-ID distance as proof of semantic proximity.
+- **pp. 216–217**: candidate maps are induced from contrasts between
+  languages. A multifunctional gram is **hypothesized** to occupy a contiguous
+  region. The arrangement is not mechanically uniquely determined by one
+  language; independent cross-linguistic observations constrain it.
+- **pp. 222–223, Figs. 8.6–8.8**: indefinite-pronoun distributions are a second
+  family of semantic functions, not free additional dative nodes. Different
+  languages carve distinct connected subsets.
+- **pp. 230–231**: linking neighboring maps into a larger map is possible, but
+  the useful extent of any map depends on the question; the method does not
+  require deciding whether each polyfunction is polysemy or vagueness.
+
+**Interpretation discipline:** a disconnected, genuinely *complete* attested
+function set falsifies this particular proposed map. A disconnected *partial*
+sample is unresolved, not a grammar error. A connected set is compatible with
+the proposed map, not proof of meaning, truth, or actual referential correctness.
+
+The example graph is exactly the **eight Fig. 8.1 dative-function nodes**,
+with the illustrated undirected connections. Its bit positions are strictly
+local to this research example, not new classid, CE64 or ABI assignments.
+
+## What is proven by code vs still missing
+
+| Existing code, consumed | Harness addition | Still outside scope |
+|---|---|---|
+| `fsm::Reading / PosSet / parse_readings` | Existing lexical alternatives remain alternatives | German case, full morphological agreement, nested relatives |
+| `spo::Spo` with **u16** IDs | Fixed epoch-wide dictionary (max 256 operands) for CE64's **u8** slots | Production canonical registry, full-vocabulary indirection |
+| `hydrate::hydrate` | Counts/addresses verified; unaddressed triples refused | Index-only hydration replacing the TOC, full throughput claim |
+| `CausalEdge64::ZERO` and S/P/O setters | Grammar-only low 24 bits, *high 40 zero* | Evidence/root persistence and NAL confidence calibration |
+| `tekamolo::read_clause`, `loci`, `wave` | Explicit future join points | TEKAMOLO x Haspelmath adjudication and 24xi4 anaphora |
+| H12 (#1551) semantic-map preregistration | Executable linguistic falsifier, not H12 replacement | Inference-function cross-system H12 comparisons |
+
+**Do not conflate:**
+1. English *to* vs French *à* is a comparison of multifunctionality,
+   not a mathematical mapping from the lexical word "to" to its WordId neighbors.
+2. CE64's three **u8 operand lanes are a versioned ALU/ISA register contract**,
+   not u16 `Spo::pack` and NOT per-verse tokens. A stable, externally pinned
+   operand dictionary at the ingest boundary is essential to make codes mean
+   the same thing for every consumer of the same sealed epoch. Never infer
+   semantic identity from coincidental byte equality between unaligned versions.
+   This harness's tiny dictionary is a compatibility fixture, not a new global
+   codebook authority or a solution for all 65,536 WordIds.
+3. Syntax does not deserve f=1, c=1, a Pearl intervention mark, or
+   `EpistemicState5::Causes`. Bare grammar output has no independent
+   witness-based evidence roots. This probe leaves the CE64 header unset.
+4. `form_evidence.evidence_root` is an external provenance identifier to
+   track, not an independent observation count; never revise confidence twice
+   just because two thought layers reference the same root.
+5. `hydrate.rs` addresses the already parsed triple stream through a TOC.
+   It does not derive SPO triples or parse morphology.
+
+## Reproducible tests / falsifiers
+
+```bash
+cargo test --manifest-path crates/deepnsm-v2/Cargo.toml haspelmath
+cargo test --manifest-path crates/deepnsm-v2/Cargo.toml --all-targets
+cargo clippy --manifest-path crates/deepnsm-v2/Cargo.toml --all-targets -- -D warnings
+cargo fmt --manifest-path crates/deepnsm-v2/Cargo.toml --check
+```
+
+| Gate | Test | Kills |
+|---|---|---|
+| HSM-0 | Fig. 8.1 English *to*, German *zu*: connected | Wrong adjacency topology |
+| HSM-1 | Purpose+judicantis: disjoint | A vacuous "everything is adjacent" map |
+| HSM-2 | Incomplete function evidence stays `Incomplete` | Mistaking missing linguistic data for contradiction |
+| HSM-3 | Single verse FSM → hydrate → CE64 → u16 roundtrip (IDs >255) using ONE frozen ABI | Silent narrowing and ephemeral relabeling |
+| HSM-4 | Parser alternatives preserved separately from certain hydrated givens | Silent arbitrary disambiguation |
+| HSM-5 | Attempt >256 globally admitted operand codes, or an unknown word | Silent overflow, implicit remap or aliasing |
+| HSM-6 | Mismatched per-verse markers refused | Cross-verse identity drift |
+| HSM-7 | Missing evidence source refused | Unsupported evidence annotation |
+| HSM-8 | CE64 upper bits remain zero | Fabricated NARS/Pearl/epistemic certainty |
+
+The code tests are *authored*, not reported as locally run. CI/local Cargo
+verification is required before enabling consumers or calling the gates green.
+
+## Next rung (separate approved PRs)
+
+**Wave A: Annotated token-to-function admission.** Read UD case/morphology,
+function evidence and cross-language correspondences as externally identified
+witnesses. Compare English *to*, German *zu*, French *à* on Haspelmath's
+recorded examples. The grammar parser must not treat these as universal
+decision rules. Study a counterexample and require a FAIL/REVISION path.
+
+**Wave B: Grammatical Sudoku.** Cross-join the admissible function mask with
+SPO/TEKAMOLO, relative-clause argument roles, and 24xi4 local antecedent
+displacements as *separate typed readings* of the immutable ABI.
+Preserve multiplicity, lexical unknowns, scope, ambiguity and signed local
+pointer overflow. Unbound references escalate instead of being guessed.
+
+**Wave C: Hydrated CE64 evidence.** Add a genuinely authorized writer whose
+shared-ABI scope is `(classid, sealed version, register codebook/version)`,
+*not* one codebook per verse. Preserve stable WordIds separately from the
+agreed operand ABI, grammar receipts, original evidence roots, source timestamps,
+and read authorization. The ABI must be identical for all readers of that
+version, with incompatible dictionaries rejected. A full-vocabulary strategy
+requires explicit, global indirection or a wider addressing contract; do not
+silently redefine the 8:8:8 physical register. Require CE64 high-bit interpretations to follow the
+current global header contract; no private reassignments.
+
+**Wave D: Measurement.** Pin the KJV + Luther corpus snapshots, report
+tokenization/PoS/grammar/hydration/CE64 packing separately, with cold/warm
+batches, peak allocation and accuracy against held-out UD gold. A 64K
+addressable corpus and fast hydration are *not* proof that grammar or
+counterfactual inference was solved in the same time. Compare exact candidate
+sets to the existing parser oracle before claiming a speedup.
+
+## Kill / stop conditions
+
+- A complete cross-linguistic attestation disconnected on a frozen map:
+  **map falsified**, not an invitation to modify oracle data to fit.
+- Missing canonical register operand or frozen ABI overflow: **refuse CE64 projection**.
+- Any altered `parse_readings` certain/alternative accounting: **STOP**.
+- Inference from grammatical contiguity to causality or truth confidence:
+  **STOP**.
+- Source-root duplication elevating confidence, stale temporal windows, or
+  unauthorized view composition: **STOP**.
+- A build lacking the relevant production writer: **do not call it wired**.
+- A per-verse or per-thought register meaning remap: **STOP**, violates the
+  interoperability goal (the BF16 / TCP-IP analogy).
+
+Related plans: D-HOLO-H12 in #1551, D-ORACLE-0 in #1549,
+D-HL-5 in #1564, D-MTP-0 in #1555, the DeepNSM-v2 D-LXC experiments
+in #1321, and the source-of-truth register/semantic distinction.

crates/deepnsm-v2/examples/haspelmath_wiring.rs

@@ -0,0 +1,69 @@
+//! Small deterministic CLI fixture for the Haspelmath wiring seam.
+//! Run: cargo run --manifest-path crates/deepnsm-v2/Cargo.toml --example haspelmath_wiring
+//! The witness and grammatical-function assignment are TEST INPUTS, not mined
+//! evidence, and no CE64 truth/causal headers are minted by this example.
+
+use deepnsm_v2::fsm::{Pos, PosSet};
+use deepnsm_v2::haspelmath::{
+    bind_tokens, run_harness, DativeFunction, FormAttestation, FrozenOperandAbi,
+};
+use deepnsm_v2::vocab::PaletteVocab;
+
+fn main() {
+    let mut vocab = PaletteVocab::new();
+    vocab.from_frequency_ranked(["doctor", "gave", "letter", "to"]);
+    let p = |pos| PosSet::single(pos);
+    let tokens = bind_tokens(
+        &vocab,
+        &[
+            ("doctor", p(Pos::Noun)),
+            ("gave", p(Pos::Verb)),
+            ("letter", p(Pos::Noun)),
+            ("to", p(Pos::Other)),
+        ],
+    )
+    .expect("known words");
+
+    // English 'to' covers this four-function region in Fig. 8.1.
+    // "coverage_complete" is a controlled fixture assumption, not
+    // automatically inferred by observing one token occurrence.
+    let attestation = FormAttestation {
+        word: vocab.id("to").expect("known"),
+        functions: DativeFunction::Direction.bit()
+            | DativeFunction::Recipient.bit()
+            | DativeFunction::Purpose.bit()
+            | DativeFunction::Experiencer.bit(),
+        coverage_complete: true,
+        evidence_root: 1,
+    };
+    // Register ABI is version-pinned across the entire execution epoch.
+    // This tiny shared dictionary is a fixture, not the deployed global ISA.
+    let abi = FrozenOperandAbi::new(
+        1,
+        &[
+            vocab.id("doctor").unwrap(),
+            vocab.id("gave").unwrap(),
+            vocab.id("letter").unwrap(),
+            vocab.id("to").unwrap(),
+        ],
+    )
+    .expect("shared dictionary fixture");
+    let out = run_harness(&[(1, 1)], &[tokens], &[vec![attestation]], 3, &abi)
+        .expect("fixture is addressable");
+    for edge in &out.encoded {
+        let decoded = out.decode(edge, &abi).expect("frozen ABI decoding");
+        println!(
+            "verse={:?} order={} SPO=({}, {}, {}) ce64={:016x} map={:?}",
+            edge.path,
+            edge.order,
+            vocab.word(decoded.subject).unwrap_or("OOV"),
+            vocab.word(decoded.predicate).unwrap_or("OOV"),
+            vocab.word(decoded.object).unwrap_or("OOV"),
+            edge.word.0,
+            out.attestations[0].2
+        );
+    }
+    assert_eq!(out.encoded.len(), 1);
+    assert_eq!(out.hydration.unaddressed, 0);
+    assert_eq!(out.encoded[0].word.0 >> 24, 0);
+}