docs(.arxiv): research shelf for MUL, Hyperon, dialectics, joins and semantic maps

.arxiv/2210.07278_meta_uncertainty_bayesian_model_comparison_vs_ce64_mul.md

@@ -0,0 +1,46 @@
+# Meta-uncertainty in Bayesian model comparison: distinct from CE64 MUL
+
+**Source:** Marvin Schmitt, Stefan T. Radev, Paul-Christian Bürkner, *Meta-Uncertainty in Bayesian Model Comparison*, AISTATS 2023 / PMLR 206.  
+**Paper:** https://arxiv.org/abs/2210.07278 · [PDF](https://arxiv.org/pdf/2210.07278) (arXiv v3, 2023-02-21).  
+**Read status:** Focused read of abstract, §§2–3.5, Proposition 2, minimum example and the overview of limitations; NOT an exhaustive reproducibility review.  
+**Date:** 2026-10-19.  
+**Repo owner:** `AdaWorldAPI/lance-graph` research notes. **No implementation claimed.**
+
+## Source-derived claim
+
+Bayesian model comparison gives posterior model probabilities (PMPs) `π_j = p(M_j | y)` over a **specified candidate model set**. A PMP close to 1 is not by itself a guarantee that repeating the experiment will lead to the same model preference. The authors introduce a distribution *over possible future PMP vectors* by simulating data under candidate models, fitting meta-models to the model-implied PMP distributions, and combining their posterior predictives into a data-informed mixture. Its purpose is to assess replicability and mitigate spuriously strong model preference under the assumptions of the candidate models.
+
+The paper distinguishes **three levels**, not three grades of subjective confidence:
+
+1. **Level 1:** uncertainty in the PMP vector for a finite observed dataset.
+2. **Level 2:** variability of model-implied PMP vectors across replicated/simulated datasets under the same generating model.
+3. **Level 3:** uncertainty about the fitted meta-model's distribution of PMPs, owing to a finite number of simulations `K`.
+
+The key construction in §3.3, Eq. (8), is a posterior predictive **mixture distribution for future PMP vectors**, weighted by observed PMPs; it is not simply a new direct posterior for model truth. Proposition 2 states non-vanishing predictive variance under the stated model-mixture conditions even when observed PMPs concentrate on one model.
+
+**Source example (§3.5):** For an atypical sample in a two-model Beta–Binomial comparison, an observed PMP is roughly `0.999`; the paper reports a predictive-mixture **mean** of roughly `0.55`. This is an illustration of model-comparison instability, **not** a universal confidence correction or a claim that `p(M_1|y)=0.55`.
+
+**Caveat:** Closed candidate sets and prior-predictive simulation can be misspecified. The paper's guarantees do not prove that every true explanation lies in the modeled family or that simulations match real experience.
+
+## Existing Lance-graph contract boundary
+
+- `crates/lance-graph-contract/src/epistemic_state5.rs`: CE64 bits **59..63** remain ONE declared product field, **Topology2 × Certification3**, 24 meaningful states and 8 reserved. Never repurpose for calibration probabilities.
+- `crates/lance-graph-contract/src/band_reading.rs`: legacy Meta-Uncertainty-Layer `TrustTexture` names (Crystalline/Solid/Fuzzy/Murky) are code-identical aliases for **topology** (Direct/IndirectKnown/IndirectUnknown/Unknown), **not source reliability estimates**.
+- NAL `F/C` are evidential frequency/confidence with their own semantics, **not** Bayesian PMPs.
+- See `.claude/knowledge/epistemic-frontier-capstone.md`, `.claude/knowledge/ce64-reasoning-substrate-plateau.md`, `.claude/plans/sag-wisdom-frontier-hard-problems-v1.md` for register ownership, provenance and model-misspecification constraints.
+
+## Potential research transfer (HYPOTHESIS, NOT THE PAPER'S RESULT)
+
+An **offline, versioned metacognitive evaluation** can compare a finite set of rival inference policies, reasoning styles or causal models, subject to a declared likelihood/generative model. Store a separate audit receipt keyed by a stable thought version, model-family version and independently grounded data provenance:
+
+- Observed distribution of support for competing models or strategies.
+- Distribution of outputs expected from replication on synthetic/held-out cases.
+- Simulation/sample counts and uncertainty in the replication estimate.
+- Agreement/disagreement with independent outcomes; out-of-family discrepancy gate.
+- Choice of next observation or abstention, with explicit limitations.
+
+Do **not** treat 64 thought slots as independent evidence, the CE64 `Certification3` ordinal as a probability, or duplicated Narsese/MeTTa/Datalog derivations as repeated independent model observations.
+
+**First falsification probe:** A model reports near-certain preference in one small dataset, but repeated held-out datasets reverse it. Compare baseline certainty-only selection against explicit predictive-variability assessment at matched compute. Include (a) identical duplicated evidence, (b) an out-of-model generating process, (c) large-sample convergence, and (d) variance due only to estimator simulation count. A good guard must expose instability without claiming the candidate set is complete.
+
+**Scope:** Research note only; no bitfield, source file, recipe or production ABI changed.

.arxiv/2310.18318v1_opencog_hyperon_100_page_reading_bookmark.md

@@ -0,0 +1,30 @@
+# READ LATER: OpenCog Hyperon (100-page background and introduction)
+
+**Title:** *OpenCog Hyperon: A Framework for AGI at the Human Level and Beyond – High-Level Background & Introduction*  
+**Primary author:** Ben Goertzel; with Vitaly Bogdanov, Michael Duncan, Deborah Duong, Zarathustra Goertzel, Jan Horlings, Matthew Ikle', Lucius Greg Meredith, Alexey Potapov, Andre Luiz de Senna, Hedra Seid, Andres Suarez, Adam Vandervorst and Robert Werko.  
+**Source:** [arXiv:2310.18318v1](https://arxiv.org/abs/2310.18318) · [specific v1 PDF](https://arxiv.org/pdf/2310.18318v1).  
+**Version:** v1, 2023. **Length:** 100 PDF pages.  
+**Status:** **READ LATER / BOOKMARK ONLY** (2026-10-19). Do not characterize sections as validated design, implemented code, or conclusions until read.  
+**Owner:** `AdaWorldAPI/lance-graph/.arxiv`. This note does **not** embed or reproduce the full paper.
+
+## Why queued
+
+Primary-source architectural reference for Hyperon/MeTTa/Atomspace, prospective integration with OpenCog systems, metagraph reasoning, cognitive models and reflective self-modification. Related to the Lance-graph research on Narsese/NAL/ONA, RelAlg/Datalog/R2IL, canonical thought identities, 64 active thoughts and semantic Git. **Relevance is prospective, not demonstrated architectural parity.**
+
+## Reading itinerary (contents headings in PDF, not analysis)
+
+- §§2–3, pp. 5–42: conception of intelligence; Atomspace and MeTTa; specialized spaces, distribution and architectural infrastructure.
+- §4, pp. 43–80: CogPrime cognitive model; episodic, working and procedural memory; reasoning, reinforcement learning, language, multimodal perception, goal refinement, reflexive self-understanding and other minds.
+- Later sections: inspect with attention to future roadmap, limitations, self-modification and capability claims.
+
+## Questions to check when resumed
+
+1. Which MeTTa/Atomspace constructs are precisely defined versus roadmap statements?
+2. Which edges support typed target ports, edge-to-edge or whole-edge references, versus plain SPO?
+3. Which rewrite, fold, temporal, evidence and control semantics can be preserved losslessly through the canonical Lance-graph ABI?
+4. Which memory, attention, goal and introspection descriptions are genuinely executable rather than conceptual?
+5. What parity tests, counterexamples and failure-to-map cases are necessary before an R2IL/Datalog/MeTTa adapter can claim equivalence?
+
+**Related existing notes:** `.arxiv/2026-10_hyperon_metta_to_lance_graph_parity_map.md` and `.arxiv/2026-10_hyperon_metagraph_folds_chronomorphisms_primary_excerpts.md`. Check those first when continuing to avoid duplicating work.
+
+**Repository action:** additive bookmark only. No production change, no existing note overwritten or deleted.

.arxiv/2409.20212_graph_attributes_structure_matching_gasm_bookmark.md

@@ -0,0 +1,22 @@
+# READ LATER: Graph Attributes and Structure Matching (GASM)
+
+**Author:** Raphaël Candelier.  
+**Paper:** *Graph matching based on similarities in structure and attributes*.  
+**Primary source:** [arXiv:2409.20212v1](https://arxiv.org/abs/2409.20212v1) | [PDF](https://arxiv.org/pdf/2409.20212v1), submitted 2024-09-30.  
+**Status:** ABSTRACT REVIEWED / FULL ALGORITHM DEFERRED. **Archived:** 2026-10-19.
+
+## Source abstract
+
+The work proposes **GASM (Graph Attributes and Structure Matching)**, incorporating connectivity and vertex/edge attributes when searching for vertex correspondences between different graphs. It introduces parameters to account for attribute reliability and compares matching quality/timing to baseline graph-matching methods. The abstract reports performance competitive with state-of-the-art methods, including on structure-only problems.
+
+No proof of a perfect or generally lossless match is claimed here. Empirical results have not been independently replicated in this note.
+
+## Relevance to the canonical thought address decision
+
+**Do not apply graph matching where none is needed.** All 64 Lance-graph thoughts share the same 65,536 canonical row addresses in a sealed version: Thought[a][i] and Thought[b][i] already agree on the canonical coordinate. Presence mask intersection and positional folds do not need GASM or a new matching index.
+
+Potential *external* uses: aligning independently minted ontologies, imported entity graphs, schema families or external datasets whose identities cannot be assumed equal. Attribute reliability may help generate candidates, but exact canonical identity still requires an explicit verified binding and preservation of alternatives.
+
+**Probe for later:** Two distinct graphs with intentionally ambiguous structure and conflicting attributes; include false-friend identity, partial isomorphism, missing edges, duplicated labels, and low-reliability attributes. Compare to an independent oracle. Never silently merge NodeGuid identities based solely on matching score.
+
+**Boundary:** Research bookmark only, not evidence that GASM is used or needed in the common-coordinate hot path.

.arxiv/2412.01423_semantic_maps_top_down_crosslinguistic_graph_bookmark.md

@@ -0,0 +1,24 @@
+# READ LATER: Top-down crosslinguistic semantic map graphs
+
+**Authors:** Zhu Liu, Cunliang Kong, Ying Liu, Maosong Sun.  
+**Paper:** *A Top-down Graph-based Tool for Modeling Classical Semantic Maps: A Crosslinguistic Case Study of Supplementary Adverbs*.  
+**Primary source:** [arXiv:2412.01423v3](https://arxiv.org/abs/2412.01423v3) | [PDF](https://arxiv.org/pdf/2412.01423v3); v3 2025-03-29, NAACL 2025.  
+**Status:** ABSTRACT REVIEWED / DETAILED READING DEFERRED. **Archived:** 2026-10-19.
+
+## Source abstract
+
+Classical semantic map models represent conceptual relatedness across languages under a **connectivity hypothesis**. The authors propose constructing conceptual spaces from a dense graph and then pruning to **maximum spanning trees**, with intrinsic/extrinsic evaluation criteria that balance structural coverage and precision. Their case study concerns crosslinguistic supplementary adverbs.
+
+The authors position the method against manually constructed semantic maps. This is not a general formal proof that all meanings in natural language are tree-shaped, nor proof that a tree entails correct grammatical inference.
+
+## Future Haspelmath / DeepNSM-v2 relevance (hypotheses)
+
+- Examine whether semantic-map connectivity can provide a **candidate-space constraint** for independently attested role/semantic relations across languages.
+- Keep language-specific readings, referential/anaphoric uncertainty, and attestation separate from shared canonical claim identity.
+- A spanning tree can lose alternative paths or polysemy; preserve such alternatives before any graph reduction.
+- Test multiple grammatical orientations over a four-sentence anaphoric ring against the **same justified canonical conclusions**, not merely aesthetically similar layouts.
+- Do not confuse the paper's crosslinguistic semantic maps with an executable RelAlg/Datalog AST, CE64 role, Haspelmath grammar ABI, or a verified semantic crosswalk.
+
+Relevant archived work: [Haspelmath DeepNSM harness](2026-10_haspelmath-deepnsm-ce64-harness.md) and [semantic maps ABI notes](2026-10_haspelmath_semantic_maps_grammar_abi_notes.md).
+
+**Read later:** Verify graph-construction formulae, pruning invariants, semantic connectedness, treatment of polysemy, and held-out crosslinguistic evaluation before designing executable transfers.

.arxiv/2502.20233_selective_yannakakis_smash_mask_first_thought_coordinates.md

@@ -0,0 +1,47 @@
+# Selective Yannakakis / SMASH: choose when to reduce before joining
+
+**Paper:** Daniela Böhm, Georg Gottlob, Matthias Lanzinger, Davide Longo, Cem Okulmus, Reinhard Pichler, Alexander Selzer, *Selective Use of Yannakakis' Algorithm to Improve Query Performance: Machine Learning to the Rescue*.  
+**Source:** [arXiv:2502.20233v2](https://arxiv.org/abs/2502.20233v2) | [PDF](https://arxiv.org/pdf/2502.20233v2) | [HTML](https://arxiv.org/html/2502.20233v2) (revised 2025-06-20).  
+**Read status:** Source-grounded selective review of §§I–II, IV–VI, not independently replicated. **Date:** 2026-10-19.  
+**Implementation status:** Research only. No code, compiler, CE64 or existing documentation modified.
+
+## Paper: claims and scope
+
+The authors study the **algorithm-selection** problem: apply Yannakakis-style semijoin reduction only where it helps query runtime. For a join-tree **acyclic** conjunctive query, the classical algorithm performs a bottom-up semijoin pass, a top-down semijoin pass and then joins reduced relations.
+
+For eligible **0MA** (zero-materialization-answerable) aggregate queries, *guardedness* and *set-safety* allow the answer to come from the guard after only the bottom-up semijoins. Examples of set-safe aggregates include MIN, MAX and COUNT DISTINCT; ordinary COUNT(*) and bag-sensitive aggregates are **not** automatically equivalent.
+
+Motivating §I SQL example: PostgreSQL original **3.38 s** vs rewritten **0.11 s**; changing one predicate switches the preferred plan, original **0.05 s** vs rewritten **0.09 s**. A faster theoretical operator may lose on physical overhead.
+
+The authors' MEAMBench contains **4,677** queries after augmentation. Their supervised selector **SMASH** is evaluated with PostgreSQL, DuckDB and Spark SQL. §VI-D reports minor slowdowns on **over half** of test queries when always rewriting, versus about **2%** with SMASH on its evaluated test workload. The authors emphasize avoiding **false positives** in applying the optimization.
+
+The paper does **not** benchmark Lance-graph, Lance 13, R2IL, mask-risc, or a shared Cartesian thought address space. Its times and speedup factors must not be presented as our measurements.
+
+## Lance-graph: directly relevant existing evidence (independent of the paper)
+
+All 64 active thoughts share **65,536 canonical row coordinates** in one sealed version. For selected aligned thought populations, eliminating nonparticipating rows is direct mask algebra, not foreign-key lookup or trie search.
+
+~~~text
+R = Aperture AND Presence[A] AND Presence[B] AND NOT Exclusion[C]
+for each address i in R:
+    fold(Thought[A][i], Thought[B][i])
+~~~
+
+COUNT/EXISTS may terminate on the mask (popcount/any) without a payload read. A complete terminal needs values only at surviving addresses. The identical canonical address is the join witness; equal arbitrary numeric offsets across unrelated domains are not.
+
+- [#1383](https://github.com/AdaWorldAPI/lance-graph/pull/1383): existing gated semijoin through MaskOp::Gather under a gate. **Not** a recommendation to resurrect forbidden foreign-value gather designs.
+- [#1573](https://github.com/AdaWorldAPI/lance-graph/pull/1573): thought-first positional reads, Hamming presence and Lance 13 masked I/O (probe).
+- [#1575](https://github.com/AdaWorldAPI/lance-graph/pull/1575): RelAlg normalize -> address R -> Lance 13 RowAddrMask -> terminal (measured probe).
+- [#1587](https://github.com/AdaWorldAPI/lance-graph/pull/1587): GYO α-acyclicity test plus a cyclic case where semijoin reduction keeps all rows although the full join is empty.
+- [#1590](https://github.com/AdaWorldAPI/lance-graph/pull/1590): shared-coordinate pair is positional; foreign-index/value-addressed gather proposals are marked VIOLATION.
+- [#1599](https://github.com/AdaWorldAPI/lance-graph/pull/1599), [#1605](https://github.com/AdaWorldAPI/lance-graph/pull/1605): partial thought squash / Git-like lifecycle remain design proposals, **not** existing physical join execution.
+
+Lance 13 Scanner::with_row_addr_prefilter accepts irregular RowAddrMask allow-lists, not just contiguous ranges. However, #1575 finds byte-read ratios of **1.00× at uniform 1%**, **0.19× at clustered 1%**, **0.99× at clustered 10%**. Arbitrary logical masks do not guarantee proportional page I/O reductions. Account separately for conversion from R to Lance row IDs, which was outside the timed scanner call.
+
+## Research-transfer hypothesis, not a paper conclusion
+
+A query plan selector could choose among (1) common-coordinate positional mask fold, (2) semijoin reduction over **proven** separator projections for real multi-variable joins, and (3) direct execution. A 0MA shortcut requires demonstrated guardedness and set-safety, including exact bag-multiplicity handling when applicable.
+
+**Falsification suite:** Compare all three at identical answers under (a) aligned rows, (b) unmatched unrelated row universes with the same integer offsets (must refuse), (c) valid acyclic join tree, (d) cyclic #1587 counterexample, (e) duplicate/bag cases, and (f) clustered vs scattered Lance addresses with stale-version mapping. Measure end-to-end wall time, source bytes, conversion, page locality, read amplification and optimizer regret. Compare simple deterministic selection before adding a learned selector or MUL calibration.
+
+**Boundary:** A masked positional intersection is already available for canonical aligned thought populations; general Yannakakis variable-separator propagation is not proven equivalent to plain AND between unrelated relation domains. No new interpreter or production code authorized.