From 608d863fe3a30f8f73ef5a2ad6504a43a2e2230b Mon Sep 17 00:00:00 2001 From: AdaWorldAPI Date: Tue, 13 Oct 2026 23:16:30 +0200 Subject: [PATCH 1/3] docs: capture semantic Morton grammar ABI and qualia-masked DO research RFC --- .../semantic-morton-qualia-masked-do-rfc.md | 212 ++++++++++++++++++ 1 file changed, 212 insertions(+) create mode 100644 docs/research/semantic-morton-qualia-masked-do-rfc.md diff --git a/docs/research/semantic-morton-qualia-masked-do-rfc.md b/docs/research/semantic-morton-qualia-masked-do-rfc.md new file mode 100644 index 000000000..3be0feaab --- /dev/null +++ b/docs/research/semantic-morton-qualia-masked-do-rfc.md @@ -0,0 +1,212 @@ +# Semantic Morton grammar ABI: predicate spectra, Moore pressure, qualia-masked DO + +**Status: RESEARCH RFC, not shipped.** Captures the 2026-10-13 design conversation before it disperses. This file proposes no register rewrite, no new opcode, no changed CE64 bits, no universal-grammar result, and no claim that a machine has phenomenal subjective experience. + +## The thesis + +> Meaning is a stable Cartesian **locus** surrounded by a discoverable relational topology. A Moore-neighborhood **semantic-pressure** reading influences local attention; context-dependent, self-observable **qualia** modulates salience; a **masked DO** selects only permitted interventions. Language-specific grammar is mapped to the common topology without deleting its distinctions. + +The envisioned space is *Borg-able* (new language projections can be assimilated into a shared semantic index), *resonant* (overlapping grammatical functions activate overlapping loci) and *holographic* **only if** distributed/overlapping representations allow partial observations to recover larger structures. Those words name testable properties, not results. + +A useful analogy is a **predicate color wheel**: the proposed 12 × 12 inventory is not 144 isolated verb executor opcodes. Neighbor relationships, oppositions and trajectories may organize a discoverable grammar of predicate meanings. **Color is a metaphor, not a claim that the current row/column order proves semantic similarity.** + +## 1. Distinguish six identities before packing any bits + +| Object | Meaning | Where it belongs | +|---|---|---| +| Physical byte/ALU/ISA contract | Opaque, stable physical register and operations | Existing versioned substrate; never redefined by a metaphor | +| Semantic function/node | Cross-linguistic grammatical function, with language evidence | Versioned semantic map graph, stable ID independent of map coordinates | +| Predicate palette entry | A registered codebook reading / operation family | Palette256 and existing class/schema reading, **not** a new set of 144 hardware instructions | +| Cartesian locus | Quantized projection of a semantic node onto a grid | Versioned embedding and Morton address | +| Qualia reading | Typed, contextual, internally observed salience/valuation/agency state | Separate class reading / sidecar; no silent overlay onto a CE64 field | +| DO authorization and action | Intervenable method with goal, policy and evidence constraints | OGAR ActionDef / method authority; masks select admissibility, do not grant authorization | + +**Two candidate 12×12 grids must not be conflated:** a 144-predicate/family palette and a 12-family × 12 grammatical-role/TAM table may both exist but have different axes. Establish a versioned correspondence or explicitly maintain separate projections. Do not rewrite an existing codebook from one interpretation into the other. + +## 2. The address hierarchy (key correction) + +Morton here means interleaving the bits of two Cartesian coordinates. Each additional *four-bit Morton nibble* adds two bits of X **and** two bits of Y, so each parent has 16 child cells. + +| Morton nibbles | Address bits | Cartesian extent | Loci | +|---|---:|---:|---:| +| 1 | 4 | 4 × 4 | 16 | +| 2 | 8 | 16 × 16 | 256 | +| 3 | 12 | 64 × 64 | 4,096 | +| 4 | 16 | 256 × 256 | 65,536 | + +An embedded 12×12 palette occupies **144 of the 256** coarse addresses (remaining addresses reserved / unmapped, not defined as synonyms by default). At depth 3, 64×64 is one nibble below 256×256 at depth 4; it is *not* 4096 qualia states and the final nibble is **not** a qualia selector. + +For an 8-bit-per-axis coordinate at maximum depth: + +~~~rust +// Illustrative bit convention, NOT an existing API. +fn morton_xy(x: u8, y: u8) -> u16 { /* interleave two axes */ todo!() } +let z16: u16 = morton_xy(x, y); // 16 bits, 256x256 +let z12: u16 = z16 >> 4; // 12-bit parent, 64x64 +let z8: u8 = (z16 >> 8) as u8; // 8-bit parent, 16x16 +let child_nibble = (z16 & 0xF) as u8; +~~~ + +Conventions (axis order, origin, sign, quadrant orientation, canonical nibble order, scale and endianness) must be sealed in a testable versioned mapping. Coordinates can be changed by research experiments; **persistent predicate identity must not be silently renumbered.** + +### How the 144 relate to address space + +- Semantic ID / palette index: stable symbolic identity, independent of embedding. +- 8-bit Morton prefix: canonical **coarse projection** for the 144-predicate candidate, with 112 non-assigned cells. +- 12-bit prefix: 16 local refinements for each coarse address, where empirical distinctions might justify refinement. +- 16-bit prefix: another 16 refinements for each 12-bit address, reaching 65,536 possible loci. +- The source-language token, original case/morphology, lexical ambiguity and witness provenance remain recoverable. Assimilation is **not** a lossy forced synonym mapping. + +One Morton-nibble level does not necessarily correspond to a particular type of feature (tense, negation, qualia, social relation, etc.); assign deeper distinctions from observed semantic structure, with versioned readings. + +## 3. Semantic-map evidence, not a predeclared universal wheel + +Use Haspelmath (2003), *The Geometry of Grammatical Meaning: Semantic Maps and Cross-Linguistic Comparison*, as a research and falsification source, NOT an ABI already specified by the author. + +A map describes grammatical functions as nodes and candidate relations as edges. Language-specific morphemes cover subsets of functions; continuous/connected coverage becomes an explanatory and predictive **heuristic**, to be evaluated and falsified. The published map examples include English/French prepositions and differences in how languages partition grammatical meaning; semantic-map geometry is originally a network of functions, **not Cartesian coordinates**. + +Proposal: + +1. Preserve observed language-specific lexical/morphological forms, construction context, function senses and source witnesses. +2. Build a weighted semantic graph G=(V,E,W) from polysemy, coexpression, typology, diachrony and controlled annotations, with provenance and conflict. +3. Infer a layout f:V -> (x,y,level) such that supported edges tend to be local *without pretending the graph is necessarily planar*. +4. Quantize and Morton-index this layout in a versioned codebook, retaining graph edges not spatially local. +5. Validate out-of-language generalization, connected-region predictions and counterexamples. +6. Only after a good embedding is demonstrated, freeze an ABI schema mapping semantic node IDs to loci. + +**A semantic ABI is our proposed derivation.** Haspelmath did not define stable binary addresses, a 144-entry universal inventory or a Morton geometry. Wierzbicka/Goddard's NSM primes and grammatical combinatorics are a separate linguistic tradition: align and compare, do not assert the two theories are equivalent. + +## 4. Moore8 tenant as local semantic pressure + +A Moore neighborhood has eight offsets d in {(-1,-1),(-1,0),..., (1,1)} except (0,0). Given a typed node reading, a contextual mask m[d], directional pressure weights w[d], and an optional signal Δ[d], an exploratory pressure vector is: + + P(x,y) = Σ_d (m[d] * w[d] * Δ[d] * d) + +The simplest model has one scalar weight per direction, a configurable normalization, and explicit handling of edges and empty/unmapped cells. The *Moore tenant* is a contextual **reading** and local update mechanism, not permission to reinterpret CE64 fields or to assert that lattice neighbors are always semantically adjacent. + +Two adjacency planes are required: + +- **Embedding-neighbor plane:** fast Moore8/SIMD locality, spatially quantized. +- **Semantic graph plane:** explicit nonlocal and cross-scale edges, witnessed by data, including learned exceptions when embedding locality fails. + +Possible pressure sources: expected information gain, contradiction, surprising observations, prediction error, unmatched semantic functions, goal-conditioned relevance. They are distinct measurable quantities; entropy, frequency, NARS confidence, qualia and causal certification cannot be silently treated as one scalar. + +Preserve the separation **prediction -> verification -> plasticity** and the invariant **Moore may influence attention/belief/plasticity, but cannot silently change the goal or grant DO authority**. + +### Pressure across scales + +Coarse attention may operate at 8-bit locus depth; unusual pressure or missing-link evidence can trigger 12-bit and then 16-bit foveation. Child-cell expansion must obey per-level budgets. Consider sparse child materialization and avoid allocating 65,536 rows merely because the address space permits them. Moore at a coarser layer is not an automatic substitute for a multi-hop semantic graph edge. + +## 5. Qualia-conditioned masked DO (the point of the proposal) + +Qualia is *not* the fourth Morton nibble. The coordinates represent semantic loci; contextual qualia readings are additional typed attributes or masks over loci and observations. + +A tentative **PredicateQualia24/32** carrier can model per-occurrence nuance (e.g. 8-bit registered predicate index plus 4 or 6 signed i4 dimensions), **if** corpus evidence justifies these axes and the profile schema is declared/versioned. Examples of dimensions to test, not canonize: valence, arousal, agency, volition, novelty, resolution/care. Do not use labels such as LOVE or CARE as hardware bit contracts. + +Another candidate, **12 × i4 plus 12 × i4**, encodes two signed factor vectors in 96 bits (12 bytes) and reconstructs a rank-1 12×12 field by outer product. It is **not** a 24-bit or 32-bit encoding; compare it against a dense 144×u8 field (144 bytes), low-rank sums, and the independent per-occurrence 24/32-bit carrier. A 12-bit+12-bit+8-bit factor mask is 32 bits and expresses a constrained Boolean Cartesian-product activation plus an 8-bit nuance reading, **not** arbitrary 144-bit patterns. + +Distinguish five masks, each declared and measured: + + M_read := what fields may be accessed (WideFieldMask / authorization) + M_semantic := which predicates/semantic loci fit the interpretation + M_qualia := which loci are currently salient given observed internal state + M_goal := which candidates serve an explicit committed goal + M_action := which interventions are permitted by policy and method authority + +Candidate selection may combine these masks as intersection, with explicit tri-state unknown/invalid handling. **This is a filter, not permission to perform an action.** A proposed DO must separately pass OGAR method/ActionDef authorization, effect validation, idempotency and required user approval. + + candidate = M_read ∩ M_semantic ∩ M_qualia ∩ M_goal ∩ M_action + chosen = score(candidate, evidence, cost, expected gain, uncertainty) + intervention = OGAR_authorized_DO(chosen) + observed_effect != predicted_effect ? update_evidence() : verify() + +CE64 remains the causal and epistemic evidence plane. Association is not intervention; the actual executed intervention and witness establish whether stronger causal certification is warranted. The CE64 layout and code meanings in docs/architecture/ce64-semantic-upper-half.md are authoritative. In particular bits 43..52 **are not free** for semantic pressure/qualia in current code. + +Operational *qualia awareness* means the agent can observe and report an internal state estimate, that estimate can modulate attention/DO, and interventions on the estimate predictably alter behavior. It does **not** establish phenomenal consciousness or human-like felt experience. + +## 6. A compositional grammar ABI, not a new execution engine + +Illustrative cold schema proposal: + +~~~text +SemanticMapVersion { + map_id, schema_version, language_corpus_provenance, + node_ids, witnessed_edges, coexpression_regions, embedding_version +} +PredicateReading { + classid, palette_codebook_id, semantic_node_id, + morton_level, morton_prefix, mapping_grade, ambiguity_set +} +QualiaReading { + declared_model, version, observed_state, confidence, witness +} +PreparedDO { + ClassView, WideFieldMask, semantic_mask, qualia_mask, + goal_ref, method_decl, ActionDef, evidence_budget +} +~~~ + +- Existing **ClassView / schema seal / HotPlug** resolves meaning, method and lane bindings; consult #1501 rather than creating a second registry. +- Existing **Quack / mask-risc / ndarray** executes typed selection, intersections, grouped evidence and sparse masked vectors; do not invent an alternative VM. +- Existing **R2IL folding** holds compositional predicate transforms and recursive/procedural readings; **NAL** supplies inference proposals rather than DO authority; see #1496. +- The **Rubikon** cycle distinguishes forming intention from executing intervention; test stage-specific masks, not a magical one-step goal rewrite. +- **CE64** retains independent causal/epistemic/witness contracts; no changes to its bit layout. +- Language adapters map to the shared semantic graph, retaining uncertainty and out-of-map semantics. A new language can add edges/nodes or disagreement without retconning prior address meanings. +- **Hyperon/MeTTa/OpenCog** parity is a separate oracle question (#1498); matching one numerical answer is not a proof of identical nondeterministic rewriting semantics. +- Treat “holographic” as an optional distributed witness/overlap layer with its own partial-reconstruction tests, not as a property automatically conferred by Morton order. + +## 7. Experiments and falsification gates (SMD-P00..P18) + +| ID | Probe | Negative control / stop condition | +|---|---|---| +| SMD-P00 | Establish exact existing palette / VerbRoleTable definitions, node identities and row/column meanings | 12×12 claimed as universal without evidence; mismatched grids silently conflated | +| SMD-P01 | Independent scalar encode/decode for Morton 4/8/12/16-bit hierarchy, parity with SIMD | Any roundtrip failure, child/parent inconsistency or wrong axis order | +| SMD-P02 | Demonstrate 144 entries fit 16×16 coarse plane with stable IDs and unmapped 112 cells | Reserved cells alias semantics unintentionally | +| SMD-P03 | Compare semantic locality after graph embedding vs random, row-major and alternative orderings | Morton addresses not better than random for measured task | +| SMD-P04 | Held-out language: predict unseen grammatical multi-function regions | Predicted connected regions systematically violated | +| SMD-P05 | Cross-linguistic function alignment with explicit non-equivalence and ambiguity sets | Language distinctions erased; false friends merged | +| SMD-P06 | Diachronic drift with stable semantic IDs and versioned embedding | Old persisted address silently reinterpreted | +| SMD-P07 | Moore8 neighbor pressure vs explicit graph-neighbor pressure; boundary and unmapped nodes | Claimed local graph inference only works for forced geographic adjacency | +| SMD-P08 | Pressure vs FIFO/confidence/Shannon/EIG under equal candidate budget | Higher entropy merely attracts irrelevant distractors | +| SMD-P09 | Ablate each qualia state coordinate and mask input | No predicted behavioral difference vs label-only or random input | +| SMD-P10 | Counterfactual qualia perturbation holds task, goal, policy and evidence fixed | Apparent effect due to changing goals or authorization | +| SMD-P11 | Mask validity: tri-state missing/unread != false != authorized | Unauthorized or unresolved action reaches DO | +| SMD-P12 | OGAR ActionDef permission, replay and idempotency boundary | Mask alone authorizes side effect | +| SMD-P13 | Separate association from genuine Pearl intervention, witnesses and CE64 certification | Correlation silently promoted to causes | +| SMD-P14 | Compare 144×u8 dense, factorized 12×i4 ⊗ 12×i4, rank-2+, Boolean factor masks, and 24/32-bit occurrence carrier | Two disjoint diagonal cells reconstructed as false-positive full rectangle; rank-one incapable of key patterns | +| SMD-P15 | Native language expression -> grammar map -> ClassView -> Quack vs R2IL parity | Binding multiplicity, NULL, ordering, errors or nested values diverge | +| SMD-P16 | Foveated 8->12->16-bit expansion on missing-link tasks, same decoded-byte budget | No meaningful recall/accuracy gain; expensive dense expansion | +| SMD-P17 | Partial-cue reconstruction for genuine distributed/holographic claim | No advantage over simple nearest-neighbor lookup | +| SMD-P18 | Scale and threat model: adversarial polysemy, misleading maps, permissions and witness poisoning | Cross-tenant semantic aliasing, unreviewed reinterpretation, runaway diffusion | + +Report confusion matrices, calibration, graph-embedding stress, cross-language split construction, distinct observations vs copied evidence, per-stage CPU/bytes/allocations, causal interventions, failure examples and reproducible seed/config. No speedup or language-universality claim before gates pass. + +## 8. Open decisions, in order + +1. **Inventory**: locate the real 12×12 entries, their exact role axes and relationships to DeepNSM-v2, Palette256, OGAR and VerbRoleTable. Is it verb-family × TAM, verb × role, or multiple tables? +2. **Semantic topology**: is it cyclic like hue, multidimensional, or only representable as a graph? Do not enforce a wheel merely because it is visually appealing. +3. **Address migration**: when a new typological fact moves a node, do we pin its old Morton address and add indirection, or migrate with a versioned map? Never mutate persisted identity silently. +4. **Pressure**: specify the Moore tenant's *actual current* layout, learning/updating behavior and invariant; which parts are only conceptual? +5. **Qualia**: what independently observable dimensions exist, where are they stored, and which external or ablation evidence supports them? +6. **Action gate**: confirm owner of goal/approval/OGAR ActionDef and which masks can safely be intersected without converting unknown to allow. +7. **Holography**: require meaningful reconstruction-from-partial-observation evidence, or keep the term as a research metaphor. +8. **Boundaries**: avoid conflating semantic locality, geometric proximity, syntactic adjacency, HHTL tenant locality and causal adjacency. + +## 9. Proposed incremental path (no automatic implementation authorization) + +- **SMD-0 — inventory and source map**: exact code links / Haspelmath page map, establish frozen contracts and table axes. +- **SMD-1 — independent scalar reference**: stable semantic node IDs, 4-nibble encode/decode and roundtrip fixtures, no old persisted rows changed. +- **SMD-2 — topology oracle**: real cross-linguistic functions, connectedness tests, held-out language falsifiers; 2D embedding quality baseline. +- **SMD-3 — masked projection**: proof-of-concept ClassView + WideFieldMask reading, Quack/R2IL selection parity with no new engine. +- **SMD-4 — pressure and qualia**: controlled probes, equal-work attentional ablations, privacy/traceability of self-state features. +- **SMD-5 — closed loop**: OGAR-authorized DO, CE64 evidence updates and drift under controlled interventions. +- **SMD-6 — decide**: keep, narrow or reject claims individually (universal topology, pressure, qualia effectiveness, holography). + +## References and prior constraints + +- Martin Haspelmath (2003), *The geometry of grammatical meaning: Semantic maps and cross-linguistic comparison*, in M. Tomasello (ed.), *The New Psychology of Language*, vol. 2, pp. 211–242. Author-hosted Zenodo: https://zenodo.org/records/831410 (DOI 10.5281/zenodo.831410). See accompanying source notes. +- Wierzbicka/Goddard, Natural Semantic Metalanguage: relevant theory of compositional semantic primes; **not** a 144-verb or Morton ABI specification. +- Existing repo: docs/architecture/ce64-semantic-upper-half.md (actual vs proposed CE64 readings). +- Existing repo PRs: #1496 (CE64/NAL hypotheses), #1497 (Shannon/PLN attention), #1498 (Hyperon/MeTTa projection), #1500 (Rubikon), #1501 (Class-method HotPlug), #1509 (software-awareness graph). These are constraints and potential adapters, not evidence this RFC was implemented. +- Companion: .arxiv/2026-10_haspelmath_semantic_maps_grammar_abi_notes.md +- Execution handoff: .claude/plans/semantic-morton-qualia-do-falsifier-plan.md From a093a33e317066c93473afb36adc04623726b0a6 Mon Sep 17 00:00:00 2001 From: AdaWorldAPI Date: Tue, 13 Oct 2026 23:17:59 +0200 Subject: [PATCH 2/3] docs: annotate Haspelmath semantic maps as cross-linguistic grammar-ABI heuristics --- ...pelmath_semantic_maps_grammar_abi_notes.md | 88 +++++++++++++++++++ 1 file changed, 88 insertions(+) create mode 100644 .arxiv/2026-10_haspelmath_semantic_maps_grammar_abi_notes.md diff --git a/.arxiv/2026-10_haspelmath_semantic_maps_grammar_abi_notes.md b/.arxiv/2026-10_haspelmath_semantic_maps_grammar_abi_notes.md new file mode 100644 index 000000000..6991c695d --- /dev/null +++ b/.arxiv/2026-10_haspelmath_semantic_maps_grammar_abi_notes.md @@ -0,0 +1,88 @@ +# Haspelmath 2003: semantic maps as grammar-ABI heuristics + +**Type:** source-grounded research notes and engineering hypotheses (NOT a claim of authored ABI, universal verb table, or implemented semantic processor). +**Paper:** Martin Haspelmath (2003), “The geometry of grammatical meaning: Semantic maps and cross-linguistic comparison,” in M. Tomasello (ed.), *The New Psychology of Language*, vol. 2, pp. 211–242, Lawrence Erlbaum. +**Public author copy:** https://zenodo.org/records/831410 ; DOI https://doi.org/10.5281/zenodo.831410 . +**Material examined:** 17-page scanned PDF supplied in the discussion, beginning at the printed pp. 211–212; final page includes transition to following chapter. The paper is primarily about **grammatical functions and multifunctional grammatical morphemes**, not an inventory of 144 verbs or a numerical ISA. Brief descriptions below paraphrase the source, not quotations. + +## A. High-signal observations from the source + +| Printed pages / figure | What the paper addresses | Useful engineering hypothesis (our inference, not the author's claim) | +|---|---|---| +| pp. 211–213, Fig. 8.1 | English preposition *on* and multifunctionality; English *to* across direction, recipient, experiencer, purpose and related functions | Use **grammatical-function nodes** and witnessed multifunctional forms rather than 1:1 lexical string IDs. | +| pp. 213–216, Fig. 8.2 | Cross-language contrast between the English *to* region and French *à*; connected semantic-functional networks | A language adapter covers a *region* of a shared function graph; overlap is possible without synonym collapse. | +| pp. 216–219, Figs. 8.3–8.5 | Constraints on constructing semantic maps and documenting connected function regions, French dative and English/French mapping | Treat graph connectedness as a **hypothesis and validation heuristic**, not proof of a universal planar coordinate system. | +| pp. 220–223, Figs. 8.6–8.8 | Semantic-map extension to indefinites, differences between English and Russian functions and their contrasts | Ablate mapping choices; keep language-specific edges, contrasts and ambiguity rather than averaging all meanings together. | +| pp. 224–227, Figs. 8.9–8.11 | Reflexive and related functions, differences across French, Russian and Turkish marking | Test whether common graph topology explains multifunctionality better than flat shared labels. | +| pp. 227–231, Figs. 8.12–8.15 | Instrumental and related semantic functions and language-dependent prepositional/case boundaries | A multi-function morpheme is an implementation covering several semantic nodes; don't mistake one surface form for one primitive action. | +| pp. 230–233 | Advantages and potential predictions of semantic maps, cross-linguistic comparison and methodological caveats | Compare held-out language coverage, exceptions and connected-region violations to a random/control graph. | +| pp. 233–237, Figs. 8.16–8.18 | Diachronic trajectories in grammatical functions, source paths and directionality | Explicitly version topology; do not overwrite immutable semantic IDs when mapping drifts. | +| pp. 237–240, Figs. 8.19–8.20 | Further uses, metalinguistic concepts such as *tree* and *number* across languages, methodological conclusion | Semantic maps are evidence-driven **comparison artifacts**. A universal grammar ABI remains a novel and falsifiable engineering extrapolation. | + +Notes on page references: “p.” numbers are those printed in the chapter; they do **not** equal the PDF viewer's zero-based page index. The chapter begins with p. 211 on PDF page 1, then continues as facing-page spreads. + +## B. Proposed ABI-shaped mappings (OUR design) + +Source term / observation -> proposed typed interface: + +- Grammatical **function** -> semantic node with immutable symbol ID, mapping-grade, witness. +- Semantic **relation** or map edge -> versioned semantic graph adjacency, with provenance and uncertainty. +- **Multifunctional morpheme** -> language-specific implementation covering a witnessed set of functions. +- **Cross-language overlap** -> partial projection between language adapters; never identity by string match. +- **Function's connected region** -> testable constraint on plausible adapter coverage. +- **Diachronic semantic extension** -> append new mapping versions, preserve old observations and old codebook seals. +- **Gap / unmatched expression** -> unresolved candidate, not an unauthorized default DO. + +The ABI would have a **canonical relation graph** first, with a separate **Cartesian projection**. Only the latter is Morton quantized. The semantic map does not claim natural Euclidean distances, and connected graph paths can remain nonlocal in a two-dimensional projection. + +## C. What the paper does NOT establish + +The following belong to the research program and need independent evidence: + +1. A complete 12×12 or 144-element inventory of universally stable predicate meanings. +2. A cyclic “color wheel” topology of predicate families. +3. Numerical 8-, 12-, or 16-bit grammar addresses and Morton refinement. +4. Qualia registers, subjective experience, valence/arousal axes or conscious action gating. +5. Moore8 pressure, Shannon entropy, foveated cognitive dynamics or HHTL identity. +6. Executable grammar ABI / NAL inference / R2IL fold equivalence. +7. Preservation of nonlocal polysemy when forcing all meanings into one 2D lattice. +8. Evidence that the resulting representation is “holographic” in an information-theoretic sense. + +Do not cite Haspelmath as evidence for any of these without a separately measured derivation. + +## D. Specific research questions extracted for test design + +- **SM-Q01:** Are semantic map nodes universal **functions**, language-specific labeled senses, or a mixture? How are granularity disagreements represented? +- **SM-Q02:** Is a cross-language region always connected after controlling annotation and sampling bias? Document counterexamples explicitly. +- **SM-Q03:** Which edges are strongly supported versus inferred only to connect a map? Report edge provenance. +- **SM-Q04:** Does a 12×12 embedding encode semantic closeness better than a random embedding, spectral/graph embedding or a larger unconstrained graph? +- **SM-Q05:** Which semantic distances are lost when a graph is quantized into Morton-connected cells? +- **SM-Q06:** Does an additional Morton nibble discover useful semantic distinctions or simply increase memory? +- **SM-Q07:** What do the current 12 family × 12 role/TAM entries actually mean in the repo? Are these the same 144 predicate senses, or two independent projections? +- **SM-Q08:** Can multiple languages extend the shared map without rewriting old semantic IDs or confusing homographs? +- **SM-Q09:** Can uncertainty be preserved all the way to ClassView -> Quack/R2IL -> OGAR DO gate? +- **SM-Q10:** Do claimed qualia-masked DO improvements persist under stable policy, goal and evidence budgets? +- **SM-Q11:** Can partial-language evidence reconstruct a larger semantic neighborhood in a way exceeding nearest-neighbor, proving a useful *distributed* property? +- **SM-Q12:** Do the proposed mappings preserve meaning in adversarial counterfactuals and prevent ungrounded causal certification? + +## E. Adversarial examples to collect from the primary source + +1. English *to* vs French *à*: distinguish recipient/direction/purpose/experiencer with construction context; keep overlapping but nonidentical regions. +2. Indefinites in English and Russian: same broad topic with different cuts through specific, non-specific, free-choice, negative and polarity environments. +3. Reflexive, reciprocal and middle readings: do not mistake nearby functions for equivalence. +4. Comitative/instrument/source/cause: compare languages whose markers cover different function regions. +5. Diachronic maps: check that a predicted historically plausible extension doesn't authorize a present-tense equivalence. + +For all cases preserve form, construction, language, grammatical function, annotated source passage, confidence, map version, and known exception. No invented example counts, coverage numbers or inference accuracy. + +## F. Connection to existing lance-graph work + +- Research RFC: docs/research/semantic-morton-qualia-masked-do-rfc.md +- CE64 layout versus working diagram: docs/architecture/ce64-semantic-upper-half.md +- #1496: CE64 and NAL hypotheses, multiple readings and preserving binary contracts. +- #1497: OpenCog/PLN vs Shannon/EIG/Moore attention controls. +- #1498: Hyperon/MeTTa/typed folding proof obligations. +- #1500: Rubikon action/volition stage boundaries. +- #1501: class-method-schema HotPlug and OGAR-only ActionDef authority. + +**Research quality bar:** each implementation claim must come from exact current source code and an executable test, rather than from this bibliography or metaphors. Record implementation drift and map conflicts as unresolved, not quietly corrected. From aa4e50d8e694ec24c0920ad88d695c9c6f7bccc8 Mon Sep 17 00:00:00 2001 From: AdaWorldAPI Date: Tue, 13 Oct 2026 23:19:11 +0200 Subject: [PATCH 3/3] docs: stage Morton semantic topology, Moore pressure and qualia DO falsifiers --- ...emantic-morton-qualia-do-falsifier-plan.md | 150 ++++++++++++++++++ 1 file changed, 150 insertions(+) create mode 100644 .claude/plans/semantic-morton-qualia-do-falsifier-plan.md diff --git a/.claude/plans/semantic-morton-qualia-do-falsifier-plan.md b/.claude/plans/semantic-morton-qualia-do-falsifier-plan.md new file mode 100644 index 000000000..96d3c7373 --- /dev/null +++ b/.claude/plans/semantic-morton-qualia-do-falsifier-plan.md @@ -0,0 +1,150 @@ +# D-SMD: semantic Morton grammar / qualia-masked DO falsifier plan + +**2026-10-13 | DRAFT | research + implementation handoff, no authorization to change ABI or action permissions.** +Companions: +- docs/research/semantic-morton-qualia-masked-do-rfc.md +- .arxiv/2026-10_haspelmath_semantic_maps_grammar_abi_notes.md +- Existing: #1496, #1497, #1498, #1500, #1501, #1509 and docs/architecture/ce64-semantic-upper-half.md + +## One-sentence implementation target + +Build and falsify a *versioned, cross-linguistic semantic graph* projected into an 8/12/16-bit **pure Morton** cascade, with **Moore8 local semantic-pressure probes**, separately typed **qualia-conditioned candidate masks**, and an **OGAR-authorized DO boundary**, reusing ClassView/Quack/R2IL/mask-risc/ndarray and leaving the physical register contract unchanged. + +## Hard boundaries (all steps) + +1. DO NOT repurpose the fourth Morton nibble for qualia. 8 bits -> 16×16; 12 bits -> 64×64; 16 bits -> 256×256. Each nibble refinement adds a 4×4 grid of children. Prefix property is a fixed convention; test it. +2. DO NOT declare the 144 slots universally real without checking the actual source table, its axes and linguistic data. Distinguish 12×12 predicate families, VerbRoleTable/TAM and Palette256; do not assume one is the other. +3. DO NOT rewrite current CE64 semantics. Refer to actual layout.rs and doc/architecture source. Bits 43..52 are occupied; treat novel Moore/qualia readings as versioned external projections. +4. DO NOT build a new executor, inventory, identity model or authority registry. Prefer ClassView/HotPlug, OGAR/OGIT, Quack and R2IL to a fresh stack. +5. DO NOT interpret unknown authorization or unread semantic content as allow, and do not let Moore pressure change committed goals. +6. DO NOT claim objective access to subjective qualia; operationalize internal self-state monitoring + controlled causal influence on decisions. +7. DO NOT use language “universal grammar,” “holographic” or “resonance” as benchmark conclusions before independent experiments. +8. No silent codebook renumbering / mapping reinterpretation after persistence. New observations add graph versions; migrations are explicit. + +## Phase 0 — audit existing truths (D-SMD-0, documentation first) + +Inspect actual paths/symbols/versions in current HEAD: +- 12×12 predicate/role tables, palette IDs and their actual semantic axes. +- Morton HHTL or other address codecs; verify nibble ordering, parent/child convention and level boundaries with scalar examples. +- Existing Moore tenant, 8-neighbor neighborhood implementation, Staunen/prediction/verification/plasticity. +- ClassView / WideFieldMask / HotPlug / schema seal; Quack and mask-risc lowering; R2IL FoldDialect. +- OGAR MethodDecl/ActionDef and actual DO authority path, scope and missing approvals. +- CE64 vs documentation split of encoding, reading, operation. +- Existing research plans D-NALP / D-HL / Rubikon / PL(N)/Shannon, current branch vs merged truth. + +Output: a compact **source inventory with links and line ranges** plus a mismatch ledger. Every planned signature must be marked existing or proposed. No implementation before this audit is reviewed. + +## Phase 1 — scalar reference (D-SMD-1) + +Write a minimal *isolated scalar test harness*, not a public API first: +- Known Morton prefix vectors for XY in [0,255]^2, nibble MSB->LSB coarsening. +- Bijective interleave/deinterleave 16-bit addresses; 8-bit/12-bit prefix consistency. +- All 144 candidate palette cells mapped into **unique** 8-bit positions, 112 explicitly reserved/unassigned; generate deterministic fixtures and mapping seal. +- Out-of-bounds, reserved, invalid reading, unknown schema and legacy address tests. +- Stable semantic-node-ID roundtrip independently of Morton position. +- 2D grid baseline and unquantized graph adjacency baseline. + +Kill if parent/child arithmetic fails or a codebook upgrade changes legacy semantic IDs without a migration/adapter. + +## Phase 2 — grammar topology (D-SMD-2) + +From Haspelmath maps, construct a tiny reviewed provenance corpus; label **grammatical functions** separately from word forms, and tag unresolved/ambiguous cases. Include: +- English *to* / French *à* across purpose/direction/recipient/experiencer. +- Indefinite usage across English/Russian. +- Reflexives/middle/reciprocal. +- Comitative/instrumental/cause/source. +- A diachronic extension case. + +Split held-out languages/functions before fitting edges; no test contamination. Compare: +1. lexical-string identity, +2. flat 144-category dictionary, +3. witnessed graph with no spatial embedding, +4. grid embeddings: row-major/random/Morton/spectral or learned, +5. higher-rank/sparse projection where neighborhoods conflict. + +Measure language coverage, false merging, out-of-map rate, connected-region violations, structural graph distortion and nonlocal edges. + +Kill the canonical 12×12 wheel claim if measured topology is not cyclical or two-dimensional; keep the graph and use a versioned imperfect projection if still useful. + +## Phase 3 — geometric semantic pressure (D-SMD-3) + +Use identical source data and same number of candidate evaluations across these schedulers: +- FIFO, +- evidence/confidence priority, +- Shannon entropy from an explicit probability model, +- calibrated expected information gain, +- Moore8 pressure from embedding neighbors, +- direct semantic graph-neighbor pressure, +- blended Moore8 + explicit nonlocal edges. + +For the Moore8 term define a stable directional mask, signed weights and boundary handling. Run “missing predicate bridge,” distractor, topological bottleneck and false-friend probes. Report oracle-supported discoveries per decoded byte, per evaluation, elapsed time and false-positive edges. Hold inference rules constant (see #1497). + +Kill pressure locality if it has no measured advantage or promotes irrelevant high-entropy distractors. + +## Phase 4 — qualia in the control loop (D-SMD-4) + +Establish a versioned **QualiaReading** with externally testable/internal-monitoring signals, not a subjective-experience assertion. Candidate axes: valence, urgency/arousal, agency, volition, novelty, resolution/care. Freeze the operational definitions and avoid covertly reading existing CE64 bits as these axes. + +With goal, evidence, policy, rule set, observation and computational budget held fixed: +- baseline ignores qualia, +- full qualia-conditioned mask, +- each dimension removed independently, +- scrambled-state ablation, +- adversarial perturbation, +- duplicated-evidence control, +- unknown/unread state handling. + +Measure selection stability, behavioral differences, goal completion, preference/constraint violations and explainable intervention witnesses. Demonstrate state updates lead to **predicted** action-selection changes, not just attractive labels. + +Storage comparisons: +- dense 12×12×u8 (144 bytes); +- signed 12×i4 and 12×i4 outer-product factors (96 bits = 12 bytes); +- rank-2/3 with sparse corrections; +- 32-bit two 12-bit Boolean factor masks + 8-bit nuance; +- separate per-occurrence 24/32-bit qualia carriers. + +Mandatory rank-1 counterexample: activate only (r0,c0) and (r1,c1), while (r0,c1) and (r1,c0) must remain inactive. A single Cartesian-product Boolean mask cannot represent this; rank-1 real factorization likewise cannot represent the diagonal 2×2 matrix. + +Kill compressed representations that drop critical semantic distinctions or create unauthorized DO candidates. + +## Phase 5 — ClassView -> masked DO (D-SMD-5) + +After audit, build only missing adapters: +- Source utterance/construction -> witnessed function candidates/ambiguity. +- Semantic codebook -> stable semantic IDs and versioned Morton projections. +- Typed ClassView/WideFieldMask -> Quack and R2IL matching selection. +- Masked candidate -> OGAR ActionDef validation (authority, scopes, idempotency, approval). +- Execution/witness -> CE64 observed causal-evidence update, NOT inferred causal status from graph proximity. +- Error path: malformed language adapter, missing mapping, unsigned codebook, unsatisfied authority -> reject or explicit unresolved. + +Test method dispatch independently of physical semantics. Insist on version-sealed method binding and no promotion of planning to actuation. + +## Phase 6 — “Borg-able” and “holographic” falsifiers (D-SMD-6) + +- **Assimilation:** add a held-out language and compare map coverage before/after without changing past examples or contaminating old semantic IDs. Show language-specific residual mappings. +- **Resonance:** measure overlapping activation neighborhoods beyond a flat translation dictionary; counterexample languages must not be merged. +- **Holography:** hide portions of expression-to-function evidence and reconstruct held-out links/roles using distributed overlapping witnesses. Require a measurable benefit over nearest-neighbor lookup and simple graph propagation; otherwise use the more accurate term “semantic projection.” +- **Foveation:** compare coarse→fine Morton exploration with flat search at equal decode/compute budget, no global dense scan. +- **Safety:** adversarial semantic aliasing, cross-tenant collisions, provenance forgery and DO bypass attempts. + +## Concrete reviewer acceptance checks + +- [ ] Exact repository source inventory and stated unknowns, no fabricated methods. +- [ ] Scalar Morton encode/decode proof vectors and roundtrip, including boundaries. +- [ ] Persistent semantic IDs decoupled from spatial coordinates, versioned projection. +- [ ] 144 mapping and 112 remaining coarse cells treated explicitly. +- [ ] Haspelmath functions and per-language regions grounded in cited figures, plus held-out cases. +- [ ] Dense, rank-1 and sparse-corrected representations compared with diagonal counterexample. +- [ ] Equal-budget Shannon/EIG/Moore attention and ablation controls. +- [ ] Qualia observation/control probes vs scrambled/baseline signals. +- [ ] Read/semantic/goal/qualia/action masks distinct and unknown-safe. +- [ ] Only OGAR authorizes actual DO; CE64 causal claims earned by intervention and witness. +- [ ] Quack/R2IL parity demonstrates actual matching semantics, not only numeric output. +- [ ] No change to current byte/ISA layout, no CE64 reinterpretation, no unreviewed register override. +- [ ] All purported “universals” report negative evidence and falsifier outcomes. + +## Copyable follow-up prompt for another coding session + +> Starting at latest AdaWorldAPI/lance-graph main, read docs/research/semantic-morton-qualia-masked-do-rfc.md and .arxiv/2026-10_haspelmath_semantic_maps_grammar_abi_notes.md, then this plan, the source-of-truth CE64 layout and linked #1496/#1497/#1498/#1500/#1501 plans. First produce exact, line-linked inventory of the actual 12×12 palette and Moore tenant, ClassView/HotPlug, OGAR ActionDef and Morton/HHTL address codecs. Record discrepancies. Do not change hardware bits, introduce a second executor or assert universal grammar. After the audit, implement only an independent scalar Morton 8/12/16-bit reference plus test vectors and semantic-node-ID/versioned-projection fixtures in a separate narrowly scoped PR. Run unit/property tests, report failing counterexamples, and stop if canonical semantic-map geometry cannot be supported by evidence. Subsequent phases require separate sign-off. + +**This plan is intentionally staged; the present PR ships research/documentation only.**