Add XOR-Merkle tree, semiring algebra, and NARS truth gates to SPO store#167
Add XOR-Merkle tree, semiring algebra, and NARS truth gates to SPO store#167AdaWorldAPI wants to merge 3 commits intomainfrom
Conversation
… trajectory Implements the Stack Hardening Guide for graph::spo: New modules: - merkle.rs: XOR-Merkle tree with O(1) insert, authenticated queries, inclusion proofs, epoch snapshots, and truth trajectory diff (22 tests) - semiring.rs: SpoSemiring trait with XorBundle, HammingMin, SimilarityMax, Reachability implementations for pluggable graph walks (6 tests) - case_insensitive.rs: CaseInsensitiveLookup trait for HashMap/BTreeMap with exact-first, case-insensitive fallback (5 tests) store.rs additions: - Belichtung 7-point prefilter (rejects ~90% before full Hamming) - TruthGate epistemic filtering (OPEN/STRONG/CERTAIN presets) - S×P→O, S×O→P, P×O→S NARS-gated axis projections - walk_chain_semiring() for pluggable traversal algebra - Authenticated query wrappers with Merkle proofs - Merkle tree wired into insert/remove sparse.rs: #[track_caller] error enrichment (IntegrityViolation, Located) cognitive_udfs.rs: hamming_bytes() delegates to rustynum SIMD Also: wire graph::spo into build, fix builder.rs invalid hex literal https://claude.ai/code/session_016SeGMg1pgf1MqK8YWkedvV
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 80d5f894dc
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| let wb = b.get_word(idx); | ||
| sample_diff += (wa ^ wb).count_ones(); | ||
| } | ||
| (sample_diff * 18) > threshold |
There was a problem hiding this comment.
The rejection check uses sample_diff * 18 as an estimate, not a mathematical lower bound, so it can reject true matches when differences are concentrated in sampled words (e.g., full Hamming distance is within threshold but sampled words overrepresent the errors). Because the query paths immediately continue on this predicate, query_forward/query_reverse/query_relation (and their truth-gated variants) can now return incomplete results for exact-radius searches.
Useful? React with 👍 / 👎.
| fn flip_up(&mut self, dn: u64, hash: &MerkleHash) { | ||
| let acc = self.children_xor.entry(dn).or_insert(ZERO_HASH); | ||
| *acc = xor_hash(acc, hash); |
There was a problem hiding this comment.
flip_up only mutates children_xor for the immediate node and never walks parent pointers, so changes under non-root parents do not update root_hash(). In a nested tree (for example, inserting a child under a non-root DN), subtree mutations are omitted from the root commitment, which breaks the integrity guarantees of root-hash comparison and authenticated proofs.
Useful? React with 👍 / 👎.
| kids.remove(&dn); | ||
| } | ||
| } | ||
| self.flip_up(parent_dn, &old_leaf); |
There was a problem hiding this comment.
When a DN already exists, the code removes the child from old_parent but XORs out the old hash from parent_dn (the new parent) instead of old_parent. If a leaf is reparented, this leaves stale hash contribution in the old branch and corrupts the Merkle accumulators/root for that update.
Useful? React with 👍 / 👎.
Module 2 (clam_path.rs): IntegrityResult enum + verify_integrity()
- roots_consistent(): compare MerkleRoots between parent/child
- verify_integrity(): walk ClamPath::parent() chain checking word[0]
- verify_batch(): fail-fast batch verification
- 8 integrity tests
Module 4 (store.rs): ClamSpoIndex + sxp2o_clam()
- ClamSpoIndex::build(): builds ClamTree over X-axis sparse data
- sxp2o_clam(): triangle inequality pruned forward query
Phase 1: rho_nn prunes clusters where δ⁻ > ρ (O(log n))
Phase 2: hamming_dense_vs_sparse checks Y-axis on candidates only
- hamming_dense_vs_sparse(): zero-alloc dense×sparse Hamming (Gate 3)
- sparse_to_bytes(): avoid to_dense() which panics on 256-word Container
- Wired graph::spo module into graph/mod.rs
- Fixed invalid hex literal 0xCHA1_D15C in builder.rs
- 8 CLAM index tests + 3 hamming function tests
https://claude.ai/code/session_016SeGMg1pgf1MqK8YWkedvV
…egrity verification Core fixes from PR #167/#168 review, rewritten from spec: - sparse.rs: Widen bitmap [u64;2] → [u64;BITMAP_WORDS=4] to cover all 256 positions - builder.rs: Fix invalid hex literal (0xCHA1_D15C → 0xC4A1_D15C), rewrite label_fp() to produce ~11% density sparse fingerprints via xorshift64 PRNG - store.rs: Add TruthGate epistemic filter (OPEN/STRONG/CERTAIN), fix scent/Hamming scale mismatch with separate scent_radius parameter, add gated query methods - geometry.rs: Add ContainerGeometry::Spo = 6 variant - mod.rs: Wire pub mod spo (was missing — ~2000 lines of dead code) - bind_space.rs: Add clam_merkle field to BindNode, stamp ClamPath + MerkleRoot on write_dn_node()/write_dn_path(), add verify_lineage() + Epoch snapshot All 61 tests pass (24 SPO + 32 ClamPath + 5 bind_space integrity). https://claude.ai/code/session_013JU8MRtxRRTtuc5217r6s5
Task 1 — Error Type Unification: - Create src/query/error.rs with snafu-based QueryError enum (ParseError, PlanError, ExecutionError, TranspileError, UnsupportedFeature, InvalidPattern, Arrow, DataFusion, SpoError) - Add #[track_caller] helper functions + convenience macros - Remove old thiserror QueryError from mod.rs, wire new module - Fix two .unwrap() calls in cypher.rs tokenizer (lines 487, 489) Task 2 — Parser Extraction: - Copy 6 lance-graph parser files into src/query/lance_parser/ (ast, parser, semantic, config, case_insensitive, parameter_substitution) - Rewrite imports: GraphError → QueryError, crate:: → super:: - Add nom + snafu dependencies to Cargo.toml - All 107 tests pass (6 error + 101 lance_parser) https://claude.ai/code/session_016SeGMg1pgf1MqK8YWkedvV
2 participants
Summary
This PR introduces three major features to the SPO graph store: (1) an XOR-Merkle tree for authenticated query results, (2) pluggable semiring algebra for flexible graph traversal, and (3) NARS truth-value gating for epistemic filtering of edges.
Key Changes
XOR-Merkle Tree (
src/graph/spo/merkle.rs)AuthenticatedResultwraps query hits with Merkle proofs for wire-level verificationSemiring Algebra (
src/graph/spo/semiring.rs)SpoSemiringfor pluggable graph traversal algebra (⊗ for edge steps, ⊕ for path accumulation)XorBundle: Default HDC path composition (XOR bind + majority-vote bundle)HammingMin: Shortest semantic path (Hamming cost, min accumulation)SimilarityMax: Best match scoring (similarity product, max accumulation)Reachability: Boolean reachability (AND/OR logic)walk_chain_semiring()method inSpoStorefor flexible traversalNARS Truth Gates (
src/graph/spo/store.rs)TruthGatestruct: Epistemic filtering withmin_freqandmin_confthresholdsOPEN(no filtering),STRONG(f≥0.7, c≥0.5),CERTAIN(f≥0.9, c≥0.8)sxp2o(),spo2x(),xpo2s()variants that apply truth gates before distance rankingSpoHit: Enriched query results with NARS truth values attachedBelichtung Prefilter (
src/graph/spo/store.rs)Store Integration
SpoStore: automatic updates on insert/removeverify_integrity()checks record against Merkle treeremove()method with Merkle cleanupUtility Modules
case_insensitive.rs: Extension trait for forgiving label/verb matching (exact match fast path, case-insensitive fallback)sparse.rsenhancements: Location-enriched error types via#[track_caller]Notable Implementation Details
https://claude.ai/code/session_016SeGMg1pgf1MqK8YWkedvV