From a22ac8606140c937688e8797614cbda89a3f7dde Mon Sep 17 00:00:00 2001 From: AdaWorldAPI Date: Wed, 14 Oct 2026 18:30:12 +0200 Subject: [PATCH 1/3] docs(research): preregister original-oracle fidelity and 144 thought fixtures --- .claude/plans/d-oracle-fidelity-matrix-v1.md | 144 +++++++++++++++++++ 1 file changed, 144 insertions(+) create mode 100644 .claude/plans/d-oracle-fidelity-matrix-v1.md diff --git a/.claude/plans/d-oracle-fidelity-matrix-v1.md b/.claude/plans/d-oracle-fidelity-matrix-v1.md new file mode 100644 index 000000000..8c0627ea3 --- /dev/null +++ b/.claude/plans/d-oracle-fidelity-matrix-v1.md @@ -0,0 +1,144 @@ +# D-ORACLE-0 — Original-oracle fidelity, auditable thoughts, and algebraic normalization + +**Date:** 2026-10-14 +**Status:** RESEARCH PROTOCOL / OPEN ENDS. No oracle parity or end-to-end speedup is claimed. +**Boundary:** research documentation only; does not modify CE64, DeepNSM-v2, NARS, parser, mask-risc, or ABI. + +## 0. Operator intent (governing) + +The goal is **not** one blended NAL/ONA/Hyperon/Pearl/Gadamer reasoning model, and not a fresh implementation masquerading as an oracle. Run and freeze the **original implementations one-to-one as independent reference oracles**; keep their native outputs, traces, and non-equivalent semantics distinct. Compare mathematical operations, logic, statistics, arithmetics, abstractions, and operational decisions across roughly **100–150 pre-registered fixtures**, preserving alternative hypotheses in up to **64 auditable Thought alpha channels per evaluation window**. + +A separate experiment then asks whether the existing, semantics-agnostic register/ISA, Quack/Datalog, class methods and mask-risc can represent and evaluate each original's *eligible operations* without changing results. Only after fidelity gates pass can H11 test end-to-end solution materialization and scaling. Hyperon-experimental is **already Rust**; first evaluate a native-Rust adapter, never assume a port or replacement is needed. + +The three continuing architecture lines: +1. **Logic**: Datalog Logic Objects, reusable Class/Schema/Methods over mathematical primitives; algebraically transform $var **before** $var.think(). Do not resurrect the former object-per-row, pointer-chasing/session scatter-gather execution path; the removal of that path was **not** a rejection of Class Methods. +2. **Intrinsic addressing**: query[$address], Morton coordinates, composed aperture/prefix restrictions, zipper-fold joins. Intrinsic coordinates are not merely a new database index. Do not materialize intermediate extensional relations unless the queried consumer requires them. +3. **Normalization boundary**: many original semantics lower to a small mathematical instruction vocabulary **where mathematically valid**. It is operational-shape normalization, **not** necessarily reduced Shannon entropy of the evidence, and not a license to erase distinctions. + +## 1. Original-reference ledger: versions MUST be frozen before measurements + +| Arm | Original implementation | Runtime | Native output / primary question | Avoid | +|---|---|---|---|---| +| JAVA-NARS | https://github.com/opennars/opennars | Java | NAL term inference, belief revision, stamps, tasks, expectation | Equating every NARS historical version, or counting stamps only after the fact | +| ONA | https://github.com/opennars/OpenNARS-for-Applications | C | application-oriented NARS, operations, temporal events, procedural predictions | Treating ONA as merely the Java code ported to C | +| OPENPSI | https://github.com/opencog/openpsi | OpenCog/GNU Scheme integration; verify precise runner | needs/drives, activation, action selection; Rubikon comparison is an **explicit hypothesis** | Pretending OpenPsi is a NAL or Heckhausen identity | +| HYPERON | https://github.com/trueagi-io/hyperon-experimental | **Rust** core, MeTTa/AtomSpace | unification, bindings, metagraph rewriting, evaluation | Inventing a universal Hyperon truth or NARS operator | +| LG-CE64 | https://github.com/AdaWorldAPI/lance-graph/tree/main/crates/causal-edge | Rust | existing CE64 truth operators and field contracts | Calling packed Pearl projections proof of intervention | +| LG-DEEPNSM | https://github.com/AdaWorldAPI/lance-graph/tree/main/crates/deepnsm-v2 | Rust | competing grammatical readings, inference pointers, BeliefArena, TEKAMOLO | Quoting PR #1314 exclusions as permanent current state | +| INDEPENDENT-MATH | independent reference formulas and small executable models | language independent | Boolean algebra, interval closure, distinct-source counting, trace equality | Using an implementation under test as its own only oracle | + +Capture per oracle: repository URL, immutable commit SHA, tag (if any), license/reuse constraints, compiler/interpreter version, dependency lockfile, hardware, invocation, configuration, seed, full stdout/event trace, structured oracle-specific output, and checksum. Confirm OpenPsi dependencies and output observability before scheduling its fixtures. Do not vendor third-party code without checking licensing. + +## 2. One fixture schema; distinct model-native semantics + +Minimal conceptual metadata, **not** a mandated serialized ABI: + +- Fixture: id, family, task, corpus/synthetic seed, input events with timestamps, context and operation bindings, expected native observables, numerical tolerances. +- Oracle run: implementation + frozen SHA, adapter revision, configuration, deterministic seed, resource budget, original full trace, normalized **comparison view** with an explicit mapping contract, abstention/unsupported reason. +- Auditable thought: slot 0..63 **plus generation/epoch**, hypothesis identity, observation-event IDs, evidence-source attribution IDs, direct/derived distinction, premises and source transitive closure, inference rule + version, original f/c or native truth values, separate interpretative admission, causal proof/witness level, state transitions, address/prefix, time. +- Provenance must distinguish **source** (paper/sensor/submitter), **individual observation receipt** (one dated event), **derivation path**, **shared-root dependence**, **reader-relative echo**, and **interpretive admission**. A source-bit stamp is only a bounded fast filter. Its modulo collisions must conservatively refuse pooling rather than silently assert independence. +- 64 channels = a bounded *concurrent hypothesis window*, **not** only 64 total tests, not 64 universal truth functions, not a proof that distinct evidence roots fit in six bits. Compare more hypotheses through separate windows with stable IDs and no aliasing. +- Keep the original native output alongside any normalized view; normalize only when a declared homomorphism/correctness mapping supports it. Unsupported/non-comparable is a valid recorded result, not a failure to hide. + +## 3. Non-interchangeable operations and semantic firewalls + +| Concept | Correct boundary | Specific falsifier | +|---|---|---| +| Literal F/C counting | f = w+/w, c = w/(w+k) subject to evidence definition and k | Repeat one source through a second inference path | +| NAL deduction/induction/abduction/revision | Specific Pei Wang inference formula, evidence dependency, stamps, choice | Double-counting or loss of premise lineage | +| ONA native task/action processing | ONA output/state transitions under same native inputs | Treat ONA runtime scheduling as identical to Java NARS | +| Shannon entropy | H = -sum p log2 p of **declared random variable** | Substituting c or f for unspecified Shannon distribution | +| Gadamer-inspired interpretation | Reader/context-specific admission or changed horizon; local engineering policy, **not** an original numeric Gadamer theorem | Rewriting original evidence weight when only interpretation changed | +| OpenPsi motivation | Needs, utility/control and action selection | Calling a Rubikon stage an existing OpenPsi API without mapping evidence | +| Hyperon rewriting/unification | Original MeTTa binding/unification/evaluation semantics | Replacing rewrite multiplicity or bindings with CE64 f/c | +| Pearl causal ladder | Association vs intervention vs counterfactual require distinct **causal identification and proof conditions** | Label SO/PO/SPO bits alone as L1/L2/L3 proof | +| Grammar case / TEKAMOLO | Linguistic role and direction (LOCATION/GOAL), not causal intervention by syntax alone | Akkusativ automatically dispatches do() | +| Class method / primitive | Method can compose $var symbolically, before .think(), and lower to existing ISA | Pointer-chasing object-method runtime quietly reinstated | +| ISA carrier / contextual reading | Stable bytes, explicit law/reader, physical kernel semantics-neutral | Changed meaning silently changes opcode or violates RegisterLaw | + +Distinct comparison kinds: **exact formula parity**, **logical conclusion equivalence**, **trace/operational parity** where meaningful, and **structural expressibility**. Never average heterogeneous results into a single accuracy number. + +## 4. Current local evidence and urgent loose ends (grounded 2026-10-14) + +1. **X17 unresolved in visible code path, regression needed (BLOCKING).** + - PR #1319 measured derived A→C from sources 0 and 1 at c=0.6561 and **empty stamp**; offering source 0 again produced c=0.9160 with insufficient provenance. + - Current crates/deepnsm-v2/src/belief.rs has an appropriate disjoint-stamp guard for *nonempty* stamps, but close_transitive still constructs pure-derived entries with Stamp::default() while recording premises. Thus a direct observation can be pooled with already-used underlying evidence. Test red first, follow derived provenance through premises, distinguish empty/unknown from proven disjoint, and do not modify NARS without an authorized later code PR. + - Study planner copy, ndarray nars_revision, deepnsm BeliefArena separately. +2. **CE64 saturation defect (BLOCKING for affected truth parity).** crates/causal-edge/src/isa.rs truth::evidence_weight saturates at f32::MAX; combining two saturated revision operands can overflow to infinity and pack NaN as zero. Existing code comments explicitly pin the defect. Pre-register the reference expected result and avoid silent clamp equivalence claims. +3. **Grammar-vs-Pearl mask confusion (BLOCKING).** crates/deepnsm-v2/src/fsm.rs answered_questions returns causal_edge::pearl::CausalMask to represent syntactic S/P/O question completion; crates/causal-edge/src/pearl.rs names SO/PO/SPO projections as levels of causal reasoning. Shared 3-bit **shape** does not grant an executed Pearl intervention or counterfactual. Require distinct typed readings and proof-witness gate; CE64 intervention/counterfactual intentionally refuse. +4. **Sim transitivity (PROBE).** belief.rs permits Copula::Sim blanket transitive closure; generic similarity is not transitive. Counterexample A~B, B~C, A not~C, then check the **declared** weaker notion actually intended. Arbitrary Rel verbs appropriately do not auto-transit. +5. **DeepNSM status freshness (FIX DOCUMENTATION ASSUMPTIONS).** PR #1314 explicitly excluded morph/TEKAMOLO/NARS at that 2026-10-03 wave. CURRENT main already exports belief, reason, tekamolo, hydrate etc. tekamolo::read_clause holds LaneHypothesis to right corner but ClauseRead exposes committed lanes plus counters, not a complete durable hypothesis ledger. Differentiate built, prototype, contract-only, candidate and absent. +6. **Reasoning trace/provenance shape (PROBE).** reason.rs carries premises and rungs; hydrate.rs preserves addressed triples and non-deduplicated revisit scanpaths. Determine whether each trace can round-trip to auditable Thought identity across transformations and method calls without mixing observation roots. +7. **144-verb palette (HOLD).** PR #1227 records 12x12 roots/tenses, 12x12 slot priors, and 6x24 actual DnVerb addresses with incompatible membership. Predicate palette still missing in #1294 (P=0). Resolve authority by explicit compatibility mapping, not silent remap; 12x12 factors only represent rank-one 144-cell fields unless enlarged. +8. **Register laws / anaphora (PROBE).** #1410 has signed i4 and RelativeOffset vs AxisPosition vs Support law binding, refuses overflow; i4 local window is not a confidence value or global graph pointer. Hold i4 anaphora and predicate factors as separate typed readings, compatible shared arithmetic kernels where legal. +9. **Dialectic engine and old class-method runtime (AUDIT).** .claude/plans/dialectic-engine-v1.md has S1: Boolean reachability first, NARS truth second pass; S2 unique-statement dedup; S4 stamps. Verify these exact invariants with original oracles. Removing prior session pointer-chasing scatter/gather class executor (#1524 lineage) did not abolish class methods as an abstraction; method-on-$var should be symbolic and lower through existing Quack/R2IL/mask-risc. +10. **H8/H9/H11 sequencing.** H8 semantics-neutral shared primitives, H9 shared-coordinate zipper-fold, H11 end-to-end materialization with logic, addresses and normalization. None of these benchmarks proves oracle semantics. H11 needs oracle/fidelity gates first and all costs (preprocessing, maintained inverses, updates, proof witnesses, scratch masks, terminal payload) charged. + +Primary workspace references: #1314, #1319, #1321, #1407, #1408, #1410, #1227, #1294, #981; above paths are evidence **to be rechecked at frozen SHA**. Old PR-body findings are not proof of current absence. + +## 5. Proposed 144-fixture matrix: candidate allocation, NOT completed tests + +| Group | Count | Gold oracle / invariant | Adversarial emphasis | +|---|---:|---|---| +| A. NAL truth / syllogisms | 26 | OpenNARS Java + independent formula oracle | deduction, induction, abduction, choice, revision, boundary f/c, quantization | +| B. Evidence roots / audit | 20 | native stamped traces + independent provenance oracle | X17, diamond dependence, duplicate citation, modulo-64 collision, derived→observed | +| C. ONA operational reasoning | 18 | ONA native execution | task scheduling, events, procedural steps, prediction under budget | +| D. OpenPsi / Rubikon mapping | 16 | OpenPsi native traces + separately specified Heckhausen/Gollwitzer mapping | need conflict, action gating, volition, feedback, no false equivalence | +| E. Hyperon/MeTTa semantics | 20 | hyperon-experimental Rust interpreter | bindings, matching, rewrite order, nondeterminism/alternatives, multiplicities | +| F. Grammar / interpretive hypotheses | 16 | DeepNSM-v2 + independently annotated linguistic cases | PosSet alternatives, TEKAMOLO abstention, i4 window, German case, no automatic do() | +| G. Pearl vs native projections | 16 | independently specified causal models / formal counterexamples | confounding, intervention, counterfactual, observational equivalence | +| H. Class, Datalog, address algebra | 12 | reference relational interpreter + equivalence proof/test | $var symbolic composition, joins, interval closure, thought channel independence | +| **TOTAL** | **144** | | | + +Each row needs actual input, expected output, exact source implementation SHA, executable invocation, provenance/output trace, tolerance, and a kill condition. These counts are proposed budgets, not fabricated executed cases. If an original system cannot express a cross-system fixture natively, record NON-COMPARABLE and put it in a genuine native-domain fixture instead. Oracle alignment is not forced. + +## 6. Pre-registered kill probes + +- **F1/X17**: same source enters as premise and fresh event; c MUST NOT increase by reusing that evidential root. Empty stamp != independently disjoint. Contrast genuinely new independent source. +- **F2**: two structurally different derivations share one observation; no double count. Record ancestry even if fast stamp is folded. +- **F3**: source IDs 0 and 64 collide in the 64-bit filter; may conservatively refuse, never falsely pool. +- **F4**: high c1,c2 CE64 revision under near-1 confidence produces no NaN-to-zero; if it does, mark FAIL and halt parity claims. +- **F5**: complete syntax S/P/O does not infer do(X) or counterfactual. Causal counterexample with confounder. +- **F6**: similarity nontransitivity counterexample must not become a generally valid rule. +- **F7**: syntactic alternatives all survive until an admitted criterion resolves them; ledger shows rejected and abstained alternatives as well as winner. +- **F8**: reproducing a Hyperon binding/rewrite MUST preserve native result multiplicity and ordering when observable; no forced NAL truth object. +- **F9**: ONA and Java NARS differences reported by version and event trace; non-parity does not automatically mean one is wrong. +- **F10**: Rubikon mapping does not declare motivation equivalent to F/C; test counterexample with high-confidence knowledge but low need/drive. +- **F11**: symbolic $var restrict/intersect/zip can precede think without object-per-row allocations; trace result independent of physical representation; account for any auxiliary indexes. +- **F12**: 64 simultaneous thought slots independent with versioned stable IDs; channel reuse cannot conflate provenance. +- **F13**: preservation of NAL f/c through normalization depends on ordering and evidence ancestry, not blanket associative mxm; check S1 two-pass rule. +- **F14**: end-to-end output boundary: many valid answers must incur corresponding serialization/payload cost; sparse answer case should show avoidable intermediate materialization. + +## 7. Sequencing and explicit non-goals + +**Gate O0:** Inventory and frozen originals; reproduce each original's own upstream smoke tests; record external build/install/license costs, skip with reasons where unavailable. Do not silently substitute clones. + +**Gate O1:** Produce fixture ledger and independent gold outcomes in native oracles. Start with F1, F4, F5, F6, F8, F10; keep critical failures visible. + +**Gate O2:** Map native semantics through typed adapters to existing Class/Schema/Methods/Datalog/Quack/CE64. Lower only legally equivalent operations, record refused or unmappable semantics. No new "universal truth value" or model-specific opcode inserted into the immutable byte ISA. + +**Gate O3:** 64 parallel auditable Thought channels with observable event/provenance identities; test collisions, repeated sources, branch differences and replay. All 144 fixture outcomes reproducible with single commanded run documented (multiple batched windows fine). + +**Gate O4:** Only then the H8/H9/H11 optimizations and an 8-arm 2^3 E2E factorial ablation: logic/methods, intrinsic address math, normalization. Include independent row-at-time reference, baseline engines where practical, parse/ingest, normalization, seal, maintained inverses, scratch masks, updates and final materialization. Report both latency and physical work, as well as semantic correctness. A speedup without fidelity is FAIL. + +**Not authorized in this PR:** implementing/fixing X17, altering CE64 truth/ISA/PEARL bits, changing the 512-byte row envelope, adding a new class method executor, claiming human-level intelligence, or claiming universal entropy/performance invariance. + +## 8. Research questions / open decisions + +- Which OpenNARS Java commit best represents Pei Wang's chosen canonical truth rules, versus separate revision/version variants? Freeze each compared lineage. +- Which ONA C runtime flags, event scheduler, NAL level and reproducibility seed can be pinned? +- Can OpenPsi's OpenCog/Scheme environment be reproduced with an archived dependency image? Which actions/needs are externally observable? +- Which subset of MeTTa evaluation/bindings is stable enough for native Rust FFI/crate linkage versus process protocol? License/cross-version checks first. +- What exact audit identity persists when a derived belief becomes directly observed? Keep original trace while preserving uniqueness of conclusion storage. +- Is a 64-bit source mask enough only as fast conservative gating while exact provenance lives in evidence receipts/graph? How do we track source vs event vs shared underlying root? +- Which operations and outputs are genuinely comparable cross-oracle? Where does abstention/non-comparable become mandatory? +- Is Gadamer here an *admission policy* or genuinely a *revision of contextual understanding*? Treat separately from numeric evidence and specify falsifiers. +- How to prevent syntactic answered_questions() and Pearl causal proofs from sharing an implicit semantic enum? +- Can Datalog represent method logic and retained competing alternatives with $var manipulated symbolically before $var.think(), lowering without new pointer-chasing executor? +- Does algebraic normalization preserve model observables under high/low entropy distributions, without collapsing model-specific information? +- On what narrow conditions can the 12×12 predicate field be compact factorization vs requiring arbitrary higher rank? +- What is the correct source of truth for OpenPsi / Rubikon correspondence: program behavior, published models, or a new explicit adapter? + +## 9. Handoff + +See the accompanying .claude/knowledge/oracle-fidelity-reference-ledger-v1.md for source URLs/status and .claude/prompts/d-oracle-0-claude-handoff.md for the executable read-first prompt. This PR only registers the research and unresolved decisions; no measurements were performed by committing this document. From 4370b844d297cc3b5e0dfb26ffdb055b59b34d57 Mon Sep 17 00:00:00 2001 From: AdaWorldAPI Date: Wed, 14 Oct 2026 18:31:48 +0200 Subject: [PATCH 2/3] docs(research): link original-oracle sources and current-code evidence --- .../oracle-fidelity-reference-ledger-v1.md | 51 +++++++++++++++++++ 1 file changed, 51 insertions(+) create mode 100644 .claude/knowledge/oracle-fidelity-reference-ledger-v1.md diff --git a/.claude/knowledge/oracle-fidelity-reference-ledger-v1.md b/.claude/knowledge/oracle-fidelity-reference-ledger-v1.md new file mode 100644 index 000000000..74b86ad45 --- /dev/null +++ b/.claude/knowledge/oracle-fidelity-reference-ledger-v1.md @@ -0,0 +1,51 @@ +# D-ORACLE-0 source and status ledger (2026-10-14) + +**Scope:** provenance map for a research proposal, not an external-oracle run. PR metadata and source snippets were inspected via connected GitHub. Every version must be frozen to immutable commit SHA for actual experiments. + +## Original independent oracles + +| Oracle | Source | Freeze/read before use | Status as of this research PR | +|---|---|---|---| +| OpenNARS Java | https://github.com/opennars/opennars | main source, inference implementations, stamps, truth functions, native test harness and versioning | Reference URL identified; NOT installed, run or evaluated | +| ONA C | https://github.com/opennars/OpenNARS-for-Applications | compiler flags, inference, event cycle, task inputs, choice/revision and native fixture syntax | Reference URL identified; NOT installed, run or evaluated | +| OpenPsi | https://github.com/opencog/openpsi | actual runtime/dependencies, needs/action traces, architecture of OpenCog integration | Reference URL identified; NOT installed, run or mapped to Rubikon | +| Hyperon Rust | https://github.com/trueagi-io/hyperon-experimental | Rust Cargo components, MeTTa test runner, bindings/rewrite trace, pin dependency SHA | Repository identified; NOT compiled, integrated or benchmarked | +| Pearl | Judea Pearl causal hierarchy, explicit SCM gold models (not merely SPO bitmasks) | structural causal model, confounding examples, intervention and counterfactual identification assumptions | External mathematical oracle specification must be written | +| Gadamer | Philosophical hermeneutics; no canonical numeric algorithm | formalize *our* policy with explicit name, admitted/denied evidence and context records | Engineering hypothesis; NOT claim of original executable Gadamer oracle | + +Do **not** require native Hyperon, NARS and OpenPsi to return the same kind of truth object. The independent oracle may be a property or formal causal model where no directly runnable reference exists. Native-language and cross-model comparisons are different tables. + +## Workspace evidence inspected on main + +| Anchor | Grounded statement | Status / next falsifier | +|---|---|---| +| https://github.com/AdaWorldAPI/lance-graph/pull/1314 | multi-reading FSM with certain/alternative; was scoped to lexical grammar wave; current source may have more | Historical baseline, NOT comprehensive current-crate inventory | +| https://github.com/AdaWorldAPI/lance-graph/pull/1321 | position-vs-frequency homograph tests and language probing | Metrics are benchmarks on particular corpora, not ground truth for all grammar | +| https://github.com/AdaWorldAPI/lance-graph/blob/main/crates/deepnsm-v2/src/lib.rs | currently exports belief, reason, tekamolo, hydrate, lexical, fsm etc. | Current built module presence verified by source read, integration/production wiring NOT verified | +| https://github.com/AdaWorldAPI/lance-graph/blob/main/crates/deepnsm-v2/src/fsm.rs | parse_readings retains readings; answered_questions returns Pearl CausalMask for syntactic SPO coverage | Distinct grammar mask from causal proof semantics; F5 | +| https://github.com/AdaWorldAPI/lance-graph/blob/main/crates/deepnsm-v2/src/tekamolo.rs | read_clause holds LaneHypothesis to right corner, then commits/abstains based on corpus evidence | Record each candidate rather than only committed ClauseRead output; F7 | +| https://github.com/AdaWorldAPI/lance-graph/blob/main/crates/deepnsm-v2/src/reason.rs | derivation pointers, rung and fixed-point gates | Check identity/premise round-trip through engine | +| https://github.com/AdaWorldAPI/lance-graph/blob/main/crates/deepnsm-v2/src/hydrate.rs | AddressedTriple + non-global-deduplicating scanpath preserving revisits | Preserve trajectory distinct from set coverage | +| https://github.com/AdaWorldAPI/lance-graph/pull/1319 | X17 was measured: derived empty stamp allowed pooling source already in ancestry | Re-run red-first on frozen main; F1/F2 | +| https://github.com/AdaWorldAPI/lance-graph/blob/main/crates/deepnsm-v2/src/belief.rs | observe/revise_at has disjoint stamp guard, but close_transitive inserts derived entries with Stamp::default(); source id % 64 folding | Must not conflate derived/no direct receipt with provably disjoint; F1/F3 | +| https://github.com/AdaWorldAPI/lance-graph/pull/1407 | CE64 strict ISA and field contracts, revision truth-only, intervention and counterfactual refuse | Unsupported must stay refusal; test boundary/saturation | +| https://github.com/AdaWorldAPI/lance-graph/blob/main/crates/causal-edge/src/isa.rs | truth::evidence_weight saturates; code comment records overflow/NaN-to-zero known defect | F4; math oracle and expected behavior first | +| https://github.com/AdaWorldAPI/lance-graph/blob/main/crates/causal-edge/src/pearl.rs | SO/PO/SPO bits are described as Pearl ladder projections | A projection is not the required causal identification proof; F5 | +| https://github.com/AdaWorldAPI/lance-graph/pull/1410 | signed register law types, RelativeOffset/AxisPosition/Support, refused overflow | Separate law-authorized readings; no confidence/anaphora reinterpretation | +| https://github.com/AdaWorldAPI/lance-graph/pull/1408 | Nars16x8, MoorePalettePairs, MooreNars16 tenant contract | No automatic production consumer; check before assuming | +| https://github.com/AdaWorldAPI/lance-graph/pull/1227 | three incompatible 144-predicate design lineages | Do not choose a new authoritative codebook without full mapping | +| https://github.com/AdaWorldAPI/lance-graph/pull/1294 | P=0 where predicate palette absent; previously wrong SUPPORTS/CONTRADICTS inversion | Check mapping and writer before claims about actual CE64 predicate semantics | +| https://github.com/AdaWorldAPI/lance-graph/blob/main/.claude/plans/dialectic-engine-v1.md | prior S1-S12 design decisions: reachability vs NARS truth and stamp decisions | Reconcile original-oracle parity against prior decisions, never overwrite silently | +| https://github.com/AdaWorldAPI/lance-graph/pull/981 | contextual AttentionFocusFacet reading; axes are semantic neutral; prefix restrictions not address bitwise-OR | Same bytes can have different readings without changing ALU | + +**Known unknowns to keep explicit:** PR #495/#497/#850 historical claims need direct local reading; correct sources for 144 palettes; OpenPsi runtime compatibility; native Hyperon trace API; oracle seed/config schema; full 144 input/expected fixture contents; whether X17 has been fixed on another worktree or unmerged branch; correctness of blanket Sim transitivity. Do not mark any unknown as DONE merely because the feature appears in a PR description. + +## Naming and accountability + +- Original = runnable upstream reference at recorded commit. +- Independent gold = explicit small mathematical/formal reference, constructed separately from implementation under test. +- LG candidate = current or experimental AdaWorldAPI implementation. +- Model parity = exact native behavior under comparable inputs and declared tolerances. +- Abstraction equivalence = preservation of model-specific observables under documented mapping. +- Performance = measured physical cost **after** parity, including hidden preparation/maintenance. +- Non-comparable = outcome preserved as such; no implicit score of zero or success. From 53131901a5a6a6b3c2c8d7144b5f8fe3f918bdeb Mon Sep 17 00:00:00 2001 From: AdaWorldAPI Date: Wed, 14 Oct 2026 18:31:51 +0200 Subject: [PATCH 3/3] docs(research): hand off oracle fixtures, falsifiers and non-goals --- .claude/prompts/d-oracle-0-claude-handoff.md | 45 ++++++++++++++++++++ 1 file changed, 45 insertions(+) create mode 100644 .claude/prompts/d-oracle-0-claude-handoff.md diff --git a/.claude/prompts/d-oracle-0-claude-handoff.md b/.claude/prompts/d-oracle-0-claude-handoff.md new file mode 100644 index 000000000..c0af077c0 --- /dev/null +++ b/.claude/prompts/d-oracle-0-claude-handoff.md @@ -0,0 +1,45 @@ +# Claude Code handoff: D-ORACLE-0, oracle fidelity before H11 + +You are operating in AdaWorldAPI/lance-graph. This is a **research-first, read-first** handoff. Operator intent and gates are recorded in .claude/plans/d-oracle-fidelity-matrix-v1.md; provenance anchors in .claude/knowledge/oracle-fidelity-reference-ledger-v1.md. Read both entirely before action. Rebase your working branch on current main where appropriate, without damaging uncommitted operator work. + +## Non-negotiable principles + +1. **Keep original oracles independent.** Read and version-freeze OpenNARS Java (Pei Wang NAL), ONA C, OpenPsi, trueagi-io/hyperon-experimental Rust/MeTTa. Never rewrite their rules as Ada approximations and then claim oracle fidelity. Run their own test suites/CLIs where feasible; if unavailable report blocked with cause. +2. **Do not blend meaning.** NAL S->P, ONA runtime inference, OpenPsi motivational/action behavior, Hyperon unification, Pearl causal proof, Shannon entropy, Gadamer interpretative admission and grammar SPO have distinct contracts. Shared byte format != shared semantics. Explicitly test that syntax-complete SPO does not automatically count as Pearl counterfactual/intervention. +3. **Run 144 pre-registered candidate fixtures** distributed 26/20/18/16/20/16/16/12 across NAL, evidence, ONA, OpenPsi, Hyperon, grammar, Pearl, and symbolic class/method/substrate. They are PLANNED, NOT ALREADY IMPLEMENTED. First create a fixture ledger with input, original SHA, native invocation/output, expected gold, tolerance, and kill criteria. No invented test outcomes. Track excluded/non-comparable honestly. +4. **64 Auditable Thought alpha channels** hold independent model-specific hypotheses + source/event/premise lineage + operation + interpretation context + time + stable ID/epoch. 64 are simultaneous slots, not 64 total experiments or unlimited provenance compressed into u6. Verify channel reuse cannot conflate roots. +5. **Class/Schema/Methods and Datalog are existing logical abstractions.** $var must be algebraically composable before $var.think(). Do not resurrect the old pointer-chasing/session scatter-gather class executor; that implementation's removal did NOT disqualify Class Methods. Use existing Quack/R2IL/mask-risc and immutable contextual ISA; add no model-specific hardcoded opcodes. +6. **No premature performance conclusions.** D-HOLO H1/H5/H6 tested kernels/ranges, not full solutions. H8/H9 logic/math falsifiers precede H11. End-to-end comparisons must count oracle adapter, normalization, seal-time preprocessing, maintained inverses, scratch masks, update work, final payload materialization and replay/audit overhead. Shannon entropy of declared variables != NARS confidence; normalization means repeated algebraic form, not guaranteed lossless entropy compression. + +## Work order + +A. Inventory and source: read the above two documents, PRs #1314/#1319/#1321/#1407/#1408/#1410/#1227/#1294, current DeepNSM-v2 belief/reason/fsm/tekamolo/hydrate and CE64 isa/pearl. Reconcile historical PR scopes against current code. Inspect original oracle licenses and native build instructions. Write exact SHA ledger. Read #495/#497/#850 locally if accessible and report true contents, never infer from snippets. + +B. Register RED-first falsifiers before changes: +- X17: derived A->C from roots 0+1 then source-0 comes back. Empty stamp may mean missing direct event, NEVER proven disjoint. Preserve source-1 lineage. Control: genuinely new source. +- F/C near c=1: CE64 saturated-weight revision must never silently become NaN-to-zero. +- Grammar-Pearl: syntax S/P/O coverage must not upgrade causal proof. +- Generic Sim nontransitivity; arbitrary relational verb does not transit. +- Original-model agreement/disagreement traces, native ONA task operations and Hyperon binding alternatives. +- OpenPsi/Rubikon mapping must demonstrate correspondences rather than asserting identity. +- Source versus observation event versus derivation root, modulo stamp collisions and 64-slot epoch reuse. +- Interpretation gate changes admissibility without silently modifying raw independent evidence. +Log native expected values before looking at Ada results where possible. + +C. Build *test adapters only after* native output schemas are understood. Retain original outputs and explicit normalized comparison views. A refusal/unsupported/non-comparable is valid. Compare models only under documented matching assumptions. For each defect report code anchor, reproduction, expected result, false-positive controls, remedy candidates and cost. Do NOT fix or refactor code in this research PR. + +D. Produce a truthful closure table: built / measured / failed / blocked / proposed for each module and each fixture group. List exact commands and artifact hashes; calculate parity per comparable subgroup, not one blended oracle score. Prioritize source dependencies and 6 falsifiers before adding breadth. Return design decisions requiring operator authorization for follow-on code PRs. + +E. Only when fidelity and audit gates are met, sketch H11 2^3 end-to-end ablation; keep kernels, full-solution materialization and semantic fidelity separate. + +## Expected deliverables + +- Versioned/oracle source ledger with native run commands and licenses. +- 144-fixture test matrix with independent Gold/native transcripts, unrun explicitly marked. +- Auditable Thought trace schema proposal, NOT an ABI migration. +- Failure-first reproduction notes for X17, causal-ladder semantics, CE64 saturation, Sim transitivity and hypothesis loss. +- Implementation status and executable-gap list. +- Follow-up PR split by ownership (DeepNSM, CE64, adapters, orchestration, H11). No single fusion PR. +- A short operator decision memo: ADOPT / PROBE / PARK / REJECT with costs and counterexamples. + +**Stop condition:** If a planned oracle cannot run, record why, supply a minimal still-useful independent mathematical oracle, and NEVER quietly substitute a new Ada implementation as the "original." Do not merge or modify production reasoning in this task.