diff --git a/.arxiv/2026-10_openpsi_rubicon_care_wisdom_frontier_source_excerpts.md b/.arxiv/2026-10_openpsi_rubicon_care_wisdom_frontier_source_excerpts.md new file mode 100644 index 000000000..a9637a62b --- /dev/null +++ b/.arxiv/2026-10_openpsi_rubicon_care_wisdom_frontier_source_excerpts.md @@ -0,0 +1,152 @@ +# D-SAG-3 source excerpts: OpenPsi, Rubikon, Shannon/EIG, CARE, ONA and Hyperon + +**Date:** 2026-10-12. **Status:** ANNOTATED RESEARCH, not a replication, formal proof, or benchmark. +**Parent work:** merged [#1509 software-awareness graph](https://github.com/AdaWorldAPI/lance-graph/pull/1509); [D-SAG-3 plan](../.claude/plans/sag-wisdom-frontier-ona-hyperon-v1.md). + +**Editorial rule:** These are **short paraphrased excerpts / takeaways**, not verbatim reproductions. Every entry states (a) what is supported by its linked source, (b) the proposed transfer to lance-graph, and (c) a disconfirming experiment. This directory is a literature shelf; OpenPsi READMEs, GitHub PRs and peer-reviewed articles are **not** being misrepresented as arXiv papers. The full text of each work must be consulted before claiming algorithmic equivalence. No source is evidence that a D-SAG-3 probe has already run. + +## S01 — CARE reference-conditioned executable programs (2026) + +**Reference:** Guanyu Liu, Weiyi Kong, Chao Tang, Zeyu Wang, Boer Zhang, Baiqing Li, Peiyu Zhang, Tianyu Shi, *CARE: Context-Aware Ranking Evolution with Executable Scoring Programs for Budgeted Reaction Optimization*, [arXiv:2606.14581v5](https://arxiv.org/abs/2606.14581), revised 2026-08-31. [Local detailed annotation](2606.14581_synthesize_programs_gate_commit.md). + +**Paraphrased excerpt:** A language model synthesizes reusable candidate-scoring logic rather than selecting the next experiment directly. A non-LLM reference policy supplies a competing selection. An explicit gate validates the challenger and chooses before the experiment outcome is known; subsequent results can retain, revise or regenerate the program. Its reported empirical successes concern budgeted **reaction optimization**, not ONA, Hyperon or a general causal reasoner. + +**Transfer hypothesis:** Use a persistent, hash-identified Quack/R2IL-compatible ranking recipe to propose the next SAG introspection or frontier investigation; score **all candidates** and compare with a deterministic reference at a separate pre-outcome gate. Commit intent only after that, and authorize effects independently. + +**Disconfirming fixture:** Candidate permutation changes the selected result; challenger only scores a subset; an invalid/nonfinite challenger still wins; or the scorer reads a post-selection receipt. Each must cause refusal/reversion to reference. + +**Crucial separation:** CARE-as-*paper* is a policy-control pattern. **CARE-as-causal-testing** (anchor deletion, Pearl intervention, falsification, and Gadamer-style reinterpretation) is our independently proposed evaluation layer; do not attribute these mechanisms to the CARE experiment paper. + +## S02 — OpenCog Hyperon metagraph architecture (2023) + +**Reference:** Ben Goertzel et al., *OpenCog Hyperon: A Framework for AGI at the Human Level and Beyond*, [arXiv:2310.18318](https://arxiv.org/abs/2310.18318), 2023. [Reference interpreter](https://github.com/trueagi-io/hyperon-experimental). + +**Paraphrased excerpt:** Hyperon redesigns OpenCog around diverse interconnected representational and reasoning components, including graph-like representations, programmatic symbolic operations and aspirations toward reflective change. It is an architectural overview, **not** an experimentally established human-level AGI result. + +**Transfer hypothesis:** D-SAG-1 already projects software methods, dependencies, compiler refusals and measured costs into ordinary graph rows. Explore typed fold/rewrite parity **on that existing substrate**, without a duplicate AtomSpace, interpreter or query engine. + +**Falsifier:** Nested structural terms or variable-binding multiplicity cannot be round-tripped as ordinary rows; expose the mismatch and stop instead of claiming general MeTTa parity. Refer to [#1498](https://github.com/AdaWorldAPI/lance-graph/pull/1498). + +## S03 — Folding / unfolding typed metagraphs (2020) + +**Reference:** Ben Goertzel, *Folding and Unfolding on Metagraphs*, [arXiv:2012.01759](https://arxiv.org/abs/2012.01759), 2020. + +**Paraphrased excerpt:** Typed metagraph operations can be formulated through recursive morphisms, including catamorphisms and anamorphisms, with treatment of higher-order link targets and recursive structures. This is a mathematical proposal, not proof that mask-risc implements these morphisms. + +**Transfer hypothesis:** Compare a finite scalar fold/unfold reference with restricted ClassView/Quack projections. The fractal-like wisdom frontier is a **repeated discovery process**, not an assertion of exact metagraph morphism equivalence. + +**Falsifier:** Same sum but different nested structure, link-to-link reference or alternative bindings. Numerical equality alone is not semantic parity. + +## S04 — Shannon's uncertainty measure (1948) + +**Reference:** Claude E. Shannon, *A Mathematical Theory of Communication*, *Bell System Technical Journal* 27 (1948), pp. 379–423 and 623–656. [Part I DOI](https://doi.org/10.1002/j.1538-7305.1948.tb01338.x), [Part II DOI](https://doi.org/10.1002/j.1538-7305.1948.tb00917.x). + +**Paraphrased excerpt:** Entropy describes uncertainty for an explicitly defined source probability distribution. It does **not** define agent interest, explanatory causality or trust in an edge. + +**Transfer hypothesis:** EIG must specify the random variable Theta (rival explanations), calibrated posterior p(Theta|D), observation model and cost horizon. No automatic reinterpretation of NARS f/c, PLN strength or CE64 code as p. + +**Falsifier:** An entropy ranker is attracted indefinitely to an irreducibly noisy observation while a lower-entropy discriminating experiment is ignored. + +## S05 — Intrinsic motivation and learning progress (2007) + +**Reference:** Pierre-Yves Oudeyer and Frédéric Kaplan, *What is intrinsic motivation? A typology of computational approaches*, *Frontiers in Neurorobotics* 1 (2007). [DOI 10.3389/neuro.12.006.2007](https://doi.org/10.3389/neuro.12.006.2007). + +**Paraphrased excerpt:** Curiosity and exploration can be made computational through distinguishable intrinsic-motivation mechanisms; psychological accounts do not trivially map to one reinforcement-learning reward law. Novelty, uncertainty, surprise and progress are not interchangeable. + +**Transfer hypothesis:** Stage-7 astonishment is not itself an unlimited reward. OpenPsi-like demands should prioritize learning progress under finite budget and separate that from goal commitment. + +**Falsifier:** High surprise with no predictable learning consumes all exploration; also compare an EIG estimator against an independently defined learning-progress baseline. + +## S06 — Bayesian information-gain intervention choice (2021) + +**Reference:** Michele Zemplenyi and Jeffrey W. Miller, *Bayesian Optimal Experimental Design for Inferring Causal Structure*, [arXiv:2103.15229](https://arxiv.org/abs/2103.15229), 2021. + +**Paraphrased excerpt:** Interventions can be selected to rapidly reduce expected posterior uncertainty about causal structures; the paper proposes tractable summaries rather than repeated full hypothetical posterior fitting. Its evaluations include simulated and protein-signaling networks. + +**Transfer hypothesis:** Use an action/observation model for rival SAG or IAM causal explanations and compare expected posterior entropy before acting. The result is a proposed **investigation**, not a permit to change state. + +**Falsifier:** Estimated EIG derived from future outcomes; uncertainty reduction with no accurate causal identification; or gains disappear on held-out interventions. + +## S07 — Interventions with target and value selection (2022) + +**Reference:** Panagiotis Tigas et al., *Interventions, Where and How? Experimental Design for Causal Models at Scale*, [arXiv:2203.02016](https://arxiv.org/abs/2203.02016), 2022. + +**Paraphrased excerpt:** Bayesian experimental design for causal structure may need to choose both *which variable* and *what intervention value* under nonlinear models, not merely select a high-entropy observable. Results are reported on synthetic and in-silico networks. + +**Transfer hypothesis:** A frontier question must distinguish observational lookup from intervention, include target/value and expected cost, and pass an independent permission gate if the intervention changes the world. + +**Falsifier:** A policy maximizing observational surprises chooses a confounded edge that cannot distinguish rival causal mechanisms. + +## S08 — NARS self and internal operations (2018) + +**Reference:** Pei Wang et al., *Self in NARS, an AGI System*, *Frontiers in Robotics and AI* (2018). [DOI 10.3389/frobt.2018.00020](https://doi.org/10.3389/frobt.2018.00020/full). + +**Paraphrased excerpt:** Self-related information and internal operations have an explicit place in a non-axiomatic reasoner under insufficient knowledge and resources. A proposal to represent software self-knowledge as edges is compatible as a *research analogy*, not automatically the NARS theory's identical mechanism. + +**Transfer hypothesis:** D-SAG-1's compiler evidence can be queried as data, while executing a resulting proposal still needs verified methods and a distinct authority path. + +**Falsifier:** Internal derived facts are recycled as independent evidence or a claim about a method is treated as proof that it ran. + +## S09 — ONA reference evaluation and bridge (implementation source) + +**References:** [OpenNARS-for-Applications](https://github.com/opennars/OpenNARS-for-Applications), [evaluation.py](https://github.com/opennars/OpenNARS-for-Applications/blob/master/evaluation.py), [ONA v0.9.3 release](https://github.com/opennars/OpenNARS-for-Applications/releases/tag/v0.9.3) (2025-04-22), and its [MeTTa-NARS declarative example](https://github.com/patham9/nartech_ros/blob/master/plugins/narsplugintests/declarative.metta), [procedural example](https://github.com/patham9/nartech_ros/blob/master/plugins/narsplugintests/procedural.metta). + +**Observed from upstream descriptions:** ONA's evaluation.py runs C tests and Narsese/English integration examples. The maintainer documents a MeTTa interface using commands AddBeliefEvent, AddBeliefEternal and AddGoalEvent with a frequency/confidence pair; ONA, bridge and native MeTTa are not thereby shown to implement the same evidence algebra. **No official CARE evaluation suite has been identified in the cited material.** + +**Transfer hypothesis:** First capture native ONA and bridge observable traces, then normalize identity, source roots, operation, time, output alternatives and costs. Compare *observable cases* before attempting a cross-engine improvement claim. + +**Falsifier:** Lost event time, modified NARS truth values, missing alternative bindings, silent duplicate evidence, or differing input budget. + +## S10 — Hyperon OpenPsi (implementation source) + +**Reference:** [iCog-Labs-Dev/hyperon-openpsi](https://github.com/iCog-Labs-Dev/hyperon-openpsi), [test suite](https://github.com/iCog-Labs-Dev/hyperon-openpsi/tree/main/test), [curious-agent use case](https://github.com/iCog-Labs-Dev/hyperon-openpsi/tree/main/use-cases/curious-agent). + +**Observed from upstream README:** Experimental MeTTa modules include demand updater, feeling updater, modulator updater, perception monitor and hill-climbing action planner. It is an experiment, not proof of a final general motivation theory. + +**Transfer hypothesis:** Run demand arbitration externally as a typed input to a pure candidate ranking policy. Do not put motivations into evidence or imply action permits. + +**Falsifier:** Same ranked demand maps to an unrelated goal, iteration-order tie changes the outcome, or the chosen goal never affects real action-selection traces. + +## S11 — OpenPsi demand arbitration regression (2026 upstream PR, not merged evidence) + +**Reference:** [hyperon-openpsi PR #117](https://github.com/iCog-Labs-Dev/hyperon-openpsi/pull/117), opened 2026-09-04, **OPEN when checked 2026-10-12**. + +**Paraphrased excerpt from PR description, not established merge:** Multiple demands in the normal interval share satisfaction score 0.84, making implicit tie decisions order-sensitive. The current selector previously mapped affiliation, competence and energy to the same mission goal. The PR proposes explicit deterministic demand order and separate goal destinations. + +**Transfer hypothesis:** Include ties, reversed row order and meaningful goal-to-action mapping in the D-SAG-3 evaluation even if the upstream fix remains unmerged. + +**Falsifier:** Same semantics, different iteration order produces a different winner; or changing dominant demand leaves goal/action unchanged. + +## S12 — Rubikon and implementation intentions (psychology, not an engine) + +**References:** [Prior literature review](2026-10_rubicon_gollwitzer_volition_source_excerpts.md), [D-RV mapping](../.claude/plans/rubicon-volition-v1.md). Consult the original Heckhausen/Gollwitzer works linked there before attributing quantitative results. + +**Paraphrased excerpt:** Goal desirability and feasibility evaluation precede goal commitment; later planning can specify cue-triggered behavior, while evidence of impossibility can justify reconsideration or disengagement. These are psychological constructs, not bits in a NARS belief state. + +**Transfer hypothesis:** Distinguish predecisional investigation, revocable goal commitment, plan-ready cue matching, and a second *execution authorization* gate. Kanban records the lifecycle rather than replacing those distinctions. + +**Falsifier:** A committed-but-unauthorized goal executes; a revoked cue fires; genuine impossibility never triggers disengagement; harmless contradictory evidence destroys every commitment. + +## S13 — Modern PLN vs ONA (separate oracles) + +**References:** [trueagi-io/PLN](https://github.com/trueagi-io/PLN), [executable library](https://github.com/trueagi-io/PLN/blob/main/lib_pln.metta), [test harness](https://github.com/trueagi-io/PLN/blob/main/test.sh), [PeTTa](https://github.com/trueagi-io/PeTTa), [#1497 source map](2026-10_opencog_pln_shannon_attention_research_map.md). + +**Observed:** Modern PLN ships MeTTa truth-rule and test surfaces; ONA is an independently designed C NARS implementation. Neither one's results are a substitute for the other one's inference semantics. + +**Transfer hypothesis:** First vary scheduler with **one fixed inference engine**, then freeze scheduler and vary engines. If a rule is inequivalent, record a mismatch instead of silently coercing it. + +**Falsifier:** Scores change because inference rules or candidate population changed rather than attention policy. + +## Source-to-experiment crosswalk + +| Evidence class | How to use it | How *not* to use it | +| --- | --- | --- | +| arXiv/publisher paper | Formulate conjecture, baseline and explicit limitations | Claim implemented/hot-path parity or arbitrary speedups | +| GitHub README/release | Select actual oracle and CLI/test entrypoints | Claim two runtimes share truth laws | +| GitHub open PR | Construct an adversarial fixture | Treat unmerged fixes as shipped | +| Our merged #1494/#1509 | Ground existing test-local mechanisms and one-hop SAG | Assume a mature production motivational engine exists | +| Our proposed D-SAG-3 | Run pre-registered falsifiers | Claim discoveries merely because graph patterns are visually fractal | + +### Citation/replication rules + +Capture exact commit SHAs for each external runtime and interpreter/submodule; install only after license/compatibility review. Verify tests on pinned snapshots with stdout/exit code and provenance before any publication-quality comparison. Follow source-specific licenses and quote limits. Preserve contradictory results. Every measurement must identify an input fixture, version, budget and independent oracle. diff --git a/.arxiv/2026-10_wisdom_frontier_active_inference_delta_and_misspecification.md b/.arxiv/2026-10_wisdom_frontier_active_inference_delta_and_misspecification.md new file mode 100644 index 000000000..476cfb0d8 --- /dev/null +++ b/.arxiv/2026-10_wisdom_frontier_active_inference_delta_and_misspecification.md @@ -0,0 +1,77 @@ +# Wisdom frontier: active inference, robust curiosity and incremental reasoning + +**Date:** 2026-10-12. **Status:** Research-only annotated literature shelf (second wave for [D-SAG-3 / PR #1511](https://github.com/AdaWorldAPI/lance-graph/pull/1511)). These are **original paraphrased research excerpts** and engineering transfer hypotheses, not verbatim copyrighted extracts, reproduction of proofs, run logs or benchmark results. This directory includes publisher and project research references as well as arXiv sources; the directory name does not assert that every source is an arXiv preprint. + +Read with [D-SAG-3 architecture](../.claude/plans/sag-wisdom-frontier-ona-hyperon-v1.md), [deferred hard problems](../.claude/plans/sag-wisdom-frontier-hard-problems-v1.md), and [first-wave literature](2026-10_openpsi_rubicon_care_wisdom_frontier_source_excerpts.md). + +## E01. Friston et al. (2026): active inference and artificial reasoning + +- **Reference:** Karl Friston, Lancelot Da Costa, Alexander Tschantz, Conor Heins, Christopher L. Buckley, Tim Verbelen, Thomas Parr, *Active inference and artificial reasoning*, [arXiv:2512.21129](https://arxiv.org/abs/2512.21129), initially December 2025; [Nature Communications 17, 10125 (published 8 September 2026)](https://doi.org/10.1038/s41467-026-77209-5). Journal version takes precedence for publication details. +- **Paraphrased excerpt:** The authors frame a policy as a sequence of actions selected partly for its expected effect on beliefs over *alternative underlying world models*. Information gain can be formulated through differences between predictive and updated posteriors, while model reduction supports selecting informative tests. They illustrate structure learning in a partially observed three-ball problem associated with synthetic insight. +- **Interesting transfer:** This is a more precise analogue to our stage-0 wisdom anchor than undirected novelty bonuses: an established explanation is useful when it generates **discriminating predictions** that lead to better identification of hidden structure. +- **Crucial nonclaim:** The publication does not validate CE64's eight-stage model, OpenPsi goalSelector, the Rubikon controller or a popcount frontier. Nor does it establish machine awareness in lance-graph. +- **Probe:** Compare model-targeted EIG against raw entropy and popcount on a three-ball-style independently specified hidden-rule fixture. A correct outcome must predict unseen trials, not simply assign every node a smaller debt. + +## E02. Wei (2024): value of information is not arbitrary intrinsic reward + +- **Reference:** Ran Wei, *Value of Information and Reward Specification in Active Inference and POMDPs*, [arXiv:2408.06542](https://arxiv.org/abs/2408.06542) (2024). +- **Paraphrased excerpt:** Examines a connection between expected free energy (EFE), belief-state planning and reward-oriented decisions, including how an epistemic value term relates to information value in partially observable decision processes. +- **Transfer:** Compare EIG-only curiosity with an explicit pragmatic objective and costs, not one unexplained "cognitive pressure" scalar. Stage 0 is maturation; a demand is preference; an authorized action is still a separate event. +- **Falsifier:** A curiosity agent collects maximally informative but operationally irrelevant samples while a cost-aware policy finishes the task. Document assumptions before importing claims of equivalence. + +## E03. Catanach and Das (2023): information gain under a wrong model + +- **Reference:** Tommie A. Catanach and Niladri Das, *Metrics for Bayesian Optimal Experiment Design under Model Misspecification*, [arXiv:2304.07949](https://arxiv.org/abs/2304.07949) (2023). +- **Paraphrased excerpt:** Standard Bayesian experiment design optimizes information gain using the selected inference model's predictions. The paper motivates and examines additional criteria, including expected generalized information gain and expected discriminatory information, to account for model discrepancy. +- **Transfer:** Make the original D-SAG-3 probability-calibration requirement stricter: holding a normalized posterior over a *wrong set of explanations* is insufficient. Maintain model criticism or a rival hypothesis family outside the trusted candidate set. +- **Falsifier:** Every candidate causal graph is false, yet a closed-world EIG controller becomes more certain; an independent discrepancy test must challenge this self-certainty. Do not import source formulas without their assumptions and quantities. + +## E04. Pathak et al. (2017): action-relevant curiosity + +- **Reference:** Deepak Pathak, Pulkit Agrawal, Alexei A. Efros, Trevor Darrell, *Curiosity-driven Exploration by Self-supervised Prediction*, [arXiv:1705.05363](https://arxiv.org/abs/1705.05363), [ICML proceedings](https://proceedings.mlr.press/v70/pathak17a.html) (2017). +- **Paraphrased excerpt:** The intrinsic signal is prediction error in a learned representation connected to the consequences of the agent's actions, rather than raw pixel surprise. The paper emphasizes filtering aspects of the environment the agent cannot influence. +- **Transfer:** Compare whether a frontier question is actually answerable/learnable under available observations and permissions before spending budgets on surprise alone. +- **Falsifier:** A random, unlearnable high-entropy stream dominates selection forever despite yielding no prediction improvement and no verified discovery. **Do not** claim a model-free method automatically identifies true causal structure. + +## E05. Burda et al. (2018): exploratory bonuses have their own biases + +- **Reference:** Yuri Burda, Harrison Edwards, Amos Storkey, Oleg Klimov, *Exploration by Random Network Distillation*, [arXiv:1810.12894](https://arxiv.org/abs/1810.12894) (2018), [original research code](https://github.com/openai/random-network-distillation) (historical, archived). +- **Paraphrased excerpt:** A prediction-error bonus relative to a fixed random target network can supply intrinsic reward; the authors also study combining exploration and external reward, reporting Atari exploration results. +- **Transfer:** Add RND-like novelty to the **comparison ledger only if a real implementation is pinned**; do not silently conflate a stage transition, residual graph debt, and novelty bonus. +- **Falsifier:** Exploration score remains high on repeated observations without improved task outcome, or dominates task utility because a coefficient changed post hoc. + +## E06. McSherry et al. (2013): differential dataflow, iterative deltas + +- **Reference:** Frank McSherry, Derek Murray, Rebecca Isaacs, Michael Isard, *Differential Dataflow*, [Microsoft Research publication / CIDR 2013](https://www.microsoft.com/en-us/research/publication/differential-dataflow/) and [Naiad technical report (2012)](https://www.microsoft.com/en-us/research/publication/composable-incremental-and-iterative-data-parallel-computation-with-naiad/). +- **Paraphrased excerpt:** Maintains changing results through differences and supports nested iteration; it motivates incremental graph computations that do not necessarily recompute all derived state after small input changes. +- **Transfer:** Our stage-0 seeded frontier needs insertion/retraction semantics and a full-recomputation oracle. Extract *correctness questions* about deltas and fixed points, not a demand to adopt Differential Dataflow or a new runtime. +- **Falsifier:** Removing a support root leaves a cached descendant "known", reintroduces a mutually supporting cycle, or wrongly discards a conclusion that has an independent alternate support. Compare full scalar recomputation after each event. + +## E07. Beckers (2021): actual cause differs from derivation dependence + +- **Reference:** Sander Beckers, *Causal Sufficiency and Actual Causation*, [Journal of Philosophical Logic 50 (2021)](https://doi.org/10.1007/s10992-021-09601-z); [published correction](https://doi.org/10.1007/s10992-021-09632-6). +- **Paraphrased excerpt:** The literature offers multiple formally different criteria for actual causation; simple necessary/sufficient reasoning and counterfactual intuition do not uniquely settle causal judgments. +- **Transfer:** Our isolated stage-0 graph ablation establishes **dependence under current inference rules**, not real-world causal sufficiency. An interventional oracle needs explicit assumptions, outcome controls and identifiability handling. +- **Falsifier:** A parent relation whose deletion makes a derived conclusion disappear is mistakenly certified as the cause of the external observed event. Respect inability to identify a cause. + +## E08. Existing internal research that constrains the transfer + +- [#1494 / Maslov Pyramid of Cognition](https://github.com/AdaWorldAPI/lance-graph/pull/1494): stage-0 wisdom is *demonstrated dependence* in test-local scaffolding; production CE64 bits 50..52 **still mean plasticity**. It is not posterior certainty. +- [#1497 / PLN, Shannon and Staunen](https://github.com/AdaWorldAPI/lance-graph/pull/1497): inference semantics, scheduler and physical execution are independent axes; NARS frequency/confidence are not calibrated probabilities. +- [#1500 / Rubikon volition](https://github.com/AdaWorldAPI/lance-graph/pull/1500): goal commitment is separate from conditional plan and permission to perform an effect. +- [#1509 / SAG](https://github.com/AdaWorldAPI/lance-graph/pull/1509): one-concept, one-hop projection is demonstrated using real Quack; multi-hop and persistent dataset are open, H1's SIMD terminal is not built. +- [CARE arXiv:2606.14581](https://arxiv.org/abs/2606.14581): reference-vs-challenger executable program validation and **pre-outcome** gate are the proposed transfer, not proof of causal intervention or symbolic truth maintenance. + +## E09. Three competing research hypotheses (explicitly unresolved) + +**H_A: Popcount suffices as an attention scheduler.** Likely cheap on suitable masks, but can be defeated by high-cost single missing links, false prerequisites and important deep investigations. + +**H_B: EIG suffices as intrinsic motivation.** More principled on calibrated distributions, but may become confidently wrong under model misspecification, or overexplore when pragmatic utility/authorization is neglected. + +**H_C: Causally gated learning progress + Rubikon commitments improves useful work.** Plausible engineering proposal, but its own overhead may exceed any discovery advantage; causal-gating might block safe observational learning, and commitment may make the system slow to revise when evidence changes. + +**Required disambiguation:** exact same candidate pool, evidence roots, external oracle, budgets and replay under A/B/C plus FIFO/reference; full costs and negative controls; report cases where H_A beats H_C as honestly as reverse cases. + +## Notes on verification + +Publication metadata and high-level abstracts were checked on 2026-10-12 against arXiv/publisher pages; source-specific evaluation code, downstream reproductions, licenses and current external commit SHAs were **not** executed or pinned in this document. Never convert these literature comparisons into "implemented" or "benchmarked" rows merely because the sources exist. A future experimental PR must pin artifacts and execute the named oracles. diff --git a/.claude/plans/sag-wisdom-frontier-hard-problems-v1.md b/.claude/plans/sag-wisdom-frontier-hard-problems-v1.md new file mode 100644 index 000000000..c075e1c8a --- /dev/null +++ b/.claude/plans/sag-wisdom-frontier-hard-problems-v1.md @@ -0,0 +1,143 @@ +# D-SAG-3.1: deferred hard problems in the wisdom frontier + +**Status:** RESEARCH ONLY. Extension of [D-SAG-3](sag-wisdom-frontier-ona-hyperon-v1.md) on PR #1511, based on merged #1509. Nothing here is a shipped feature, performed experiment, authorization, or production ABI decision. [Source excerpts](../../.arxiv/2026-10_wisdom_frontier_active_inference_delta_and_misspecification.md). + +## H1. Four distinct epistemic quantities + +Prerequisite debt, probabilistic uncertainty, causal identifiability and intrinsic motivational urgency are NOT the same numerical object. + +For a finite, declared prerequisite universe U, define masks R (required), K (established), F (refuted), and X (unread/inaccessible). Then: + + conflict = (K & F) | (K & X) | (F & X) + missing = R & ~(K | F | X) + blocked = R & F + unavailable = R & X + debt = popcount(missing) + +Reject or surface conflict rather than resolving its overlap by bitwise precedence. A zero debt means ready to independently validate, not true. R outside the declared U is a typed refusal. X is NOT negative evidence, especially for IAM security. + +Unknown prerequisites can have alternative sufficient sets. Popcount of their union overestimates how many questions must be answered; compare each eligible alternative, accounting for shared roots without exponential unbounded DNF expansion. + +A one-bit missing source that requires expensive, unauthorized remote work can cost more than a three-bit set answered from a local index. Popcount is a cheap structural feature, never a proxy for time, expected information gain or causal effect. + +## H2. Fractal-like recursion is not mathematical self-similarity + +A stage-0 anchor generates predictions, whose contradicted relations can begin their own seven-to-zero maturation cycle. A real graph is not necessarily a tree: it has joins, cycles, repeated ancestry, high-degree hubs, alternative derivations and versioned scopes. Do not claim a fractal dimension or infinite recursion from a recursive diagram. + +Seed the frontier with verified stage-0 anchors only under explicit experimental provenance. Carry typed identifiers (ClassId, schema seal, reading, source revision, candidate identity, evidence epoch, scope) and a causal record for why a candidate was enumerated. + +Each bounded exploration epoch has a stable snapshot, seed list, budgets for work and source access, a candidate dedup key, hop/row limits, cancellation token, per-anchor cap, and a declared deterministic ordering. Budget-exhausted is INCOMPLETE, not a fixed point. A real fixed point requires a finite, fully explored closure. + +## H3. Incremental retraction and truth maintenance + +Merged #1494 measured small test-local hydration and anchor ablation. That is not yet a general incremental truth-maintenance algorithm. + +Fixture: independent root r0 grounds A; A supports B; B supports C; independent root r1 also supports B. Remove r0. A must reopen, but B and C may still be grounded via r1. Remove r1 next. B and C reopen. Never blindly erase every descendant and never retain a stale descendant after its only support disappears. + +Cycle: A -> B -> C -> A with no external root must stay unresolved. With one external root it can ground according to declared rules. Removing that root must not re-ground the component using cached children. Newly adding a *different* independent root later may ground it again, but cannot inherit a phantom root from the first cycle. + +**Independent oracle:** after EVERY delta, replay the entire immutable evidence-event history into a fresh scalar full-recomputation oracle. Require equality of resolved/open sets, independent root sets, maturation, refusal reasons, and eligible proposed actions. Only then evaluate delta update work. Differential Dataflow and seminaive evaluation provide research analogies, not a demand for another engine. + +Events include insert, retract, independent-source correction, schema/readings change, rule revision, permit revision and explicit source inaccessible. An unread source must never silently become empty. Test every event ordering that should commute and declare where real-time observations cannot commute. + +## H4. Multiple layers of identity and freshness + +Test-local proposal record: + + (source_revision, schema_seal, class_id, reading_id, + relation_identity, evidence_epoch, original_root_ids, + derivation_ids, computation_version, authorization_scope) + +Test-local auditable decision record: + + (snapshot, program_identity, compiler_version, candidate_set_hash, + evidence_cutoff, reference_rank, challenger_rank, gate_features, + tie_break, budget, intention_id, permit_id, outcome_receipt_id) + +This is NOT an ABI and is NOT a cryptographically signed receipt. #1509 has a within-build FNV/Debug-derived digest that cannot be assumed stable or collision-resistant across builds. A cryptographic audit claim would require canonical serialization, keys, signature scheme and a threat model. Do not retrofit one by assertion. + +A result cannot remain verified after its schema, class reading, provenance source or compiled program changes without a new check. Use source and data snapshot sealing to prevent temporal look-ahead. + +## H5. Motivation, Rubikon and controller stability + +OpenPsi-like demands select *what to investigate*, not permission to act. Keep four distinct records: epistemic statement, goal desirability, committed intention, and permitted effect. + +Proposed cognitive work states: + + Candidate -> Deliberating -> GoalCommitted -> PlanReady + -> PermitPending -> Executing -> Evaluating + -> Done | Replan | Hold | Disengaged | Revoked + +Checks: +- Equal demands plus reversed iteration order produce the same choice, and different winning demands map to meaningfully different goals (OpenPsi PR #117 is an unmerged upstream proposal at time of initial audit). +- Demand fluctuation every tick must not discard each previously committed plan. Measure goal-switching hysteresis against fast response to genuine contradictions. +- High-entropy curiosity may not starve a low-interest but necessary task forever. Report fairness/age quota as an explicit policy, not a hidden CE64 meaning. +- Impossible goal vs delayed-but-feasible goal must be distinguished. Authorization denial cancels dispatch but need not refute the goal itself. +- Every conditional action must re-check current cue, schema, capability, actor/subject and permit. A previous goal commitment is not authorization. + +## H6. Three kinds of causal validation + +1. **Graph support ablation** deletes an anchor in an isolated copy and measures which *derived statements* reopen. That establishes dependence on current proof/rule machinery. +2. **Causal intervention** changes a variable under specified structural/experimental assumptions and measures its outcome. This may require authorized simulated effects. +3. **Model criticism** checks whether all offered causal hypotheses might be wrong; high posterior confidence inside a misspecified model does not establish external truth. + +Do not call (1) proof of (2). Do not ascribe (1) or (2) to CARE arXiv:2606.14581, whose transferable mechanism is executable challenger/reference selection with a pre-outcome commit gate. + +Falsifiers: correlated evidence counted twice, ablation-sensitive derivation with false real-world cause, absent true mechanism from all rival models, observationally unidentifiable causal relation, rehydration masquerading as new independent evidence. + +## H7. Evaluate information seeking against misspecification + +A valid EIG requires a declared calibrated probability model over rival hypotheses and their observation likelihoods. NARS f/c and CE64 readings are not probabilities by default. + +Always compare FIFO, confidence, typed popcount, raw Shannon, calibrated EIG, EIG with model-discrepancy checks, OpenPsi-like demand, and a separate CARE challenger gate. Freeze the inference engine when changing selection policy. Freeze selection when changing native ONA/Hyperon/PLN inference semantics. + +Two traps: irreducible noisy-TV randomness (high entropy, no learnable structure), and confident-but-wrong finite hypothesis sets (internally large EIG for the wrong model). Include calibration and held-out model criticism, EIG vs out-of-distribution interventions, and negative/silence controls. + +Treat pragmatic task value, epistemic learning value, causal testability, work and authorization as independent dimensions. Use Pareto and matched-budget curves before multiplying quantities into an arbitrary scalar objective. + +## H8. Physical cost and limits + +One popcount instruction is fast; generating a high-degree multigraph frontier is not automatically fast. Explore path explosions, duplicate nodes, alternate proof roots, memory contention, source fetch, Quack compiler, mask cardinality and receipt storage. + +Measurements must include graph enumeration, root validation, queue management, decoding, IO, skipped/refused work, compile/validation, replay and cancellation. Evaluate chains and branching (1,4,16,128), shared descendants, cycles, stale-codebook churn and duplication storms. Claim a speedup only after equal correctness and all-in costs are measured. + +Do not add a 64-field schema limit merely because a prototype uses a u64 mask. Make mask width a declared domain or use a typed multiword bitset. Multi-hop operations remain blocked in #1509 unless a separately verified legal relation-carrier path exists. No shadow traversal VM. + +## H9. Cross-engine oracle matrix instead of parity marketing + +For ONA native, ONA-through-MeTTa-NARS, Hyperon MeTTa, modern PLN/PeTTa and OpenPsi, capture raw input, temporal vs eternal event, raw f/c or truth object, derivation roots, rule IDs, answer alternatives and multiplicity, order contracts, selected goals, actual operations, runtime flags, source SHA, and stderr. + +Only compare fields represented under both semantics. Unsupported is a legitimate outcome. Never normalize away disagreement between ONA time handling and MeTTa binding multiplicity, invent an official CARE benchmark, or call a fixture a cross-engine law without executing pinned oracles. + +## H10. New adversarial ledger: proposals only + +| ID | Fixture | Red condition | +| --- | --- | --- | +| A16 | Required / known / refuted / unread masks | Unread becomes false or conflict is hidden | +| A17 | Alternative support r0/r1 after retraction | Useful derivation lost or ghost support retained | +| A18 | Grounded and ungrounded reopened 3-cycle | Reopens and falsely regrounds itself | +| A19 | Delta permutations vs full recomputation | Inconsistent final proof/root sets | +| A20 | One expensive debt vs three cheap debts | Popcount declared optimal cost | +| A21 | High fanout with bounded queue | Budget exhaustion reported as fixed point | +| A22 | All candidate models wrong | High EIG called causal discovery | +| A23 | Pre-outcome cutoff under challenger scoring | Any future outcome leaks | +| A24 | OpenPsi tied demands reversed | Different goal selected by row order | +| A25 | Novelty storm and blocked task | Starvation or oscillating commits | +| A26 | Committed intention with canceled permit | Actual effect dispatched | +| A27 | Graph ablation vs causal intervention | Derivation dependence labeled causal truth | +| A28 | Mutated method body/seal/compiler/source | Stale proof accepted | +| A29 | Stage-0 anchor adjacent to private IAM edge | Cross-scope exploration | +| A30 | ONA eternal vs temporal, MeTTa alternatives | Non-equivalent rules coerced | +| A31 | Same evidence after fixed point | Reward/stage grows without new root | +| A32 | Cancel during execution race | Stale operation after cancellation | +| A33 | Alternative prerequisite sets overlap | Union-popcount misranks | + +Each implemented row must have independently derived expected result, positive fixture, negative/silence twin, guard-disable -> red, restore -> green, seed, provenance and complete cost. All rows remain unexecuted research specifications. + +## H11. Research release sequence and stop conditions + +A0 audit current source and exact code API -> A1 four-state scalar masks -> A2 retraction vs full recomputation -> A3 existing Quack one-hop parity -> A4 calibrated attention -> A5 pure Rubikon/OpenPsi simulation -> A6 reference-conditioned CARE receipts -> A7 pinned ONA/Hyperon differential oracles -> A8 self-introspection proposal that is NOT auto-applied. + +Stop if a run changes production CE64[50..52] plasticity, uses [59..63] as desire, invents independent observation roots, bypasses a live authorization check, mistakes budget exhaustion for proof closure, silently implements a new traversal engine, or reports a fused H1 speedup before the actual terminal exists. + +**Open operator questions:** Is there a legal relation carrier for multihop? What provenance format can pin support sets under deletions? Which calibration data can be released? Would promoting maturation require its own declared semantic encoding or a versioned sidecar rather than stealing plasticity bits? Record decisions; do not silently choose them in this PR. diff --git a/.claude/plans/sag-wisdom-frontier-ona-hyperon-v1.md b/.claude/plans/sag-wisdom-frontier-ona-hyperon-v1.md new file mode 100644 index 000000000..f8ca75c07 --- /dev/null +++ b/.claude/plans/sag-wisdom-frontier-ona-hyperon-v1.md @@ -0,0 +1,193 @@ +# D-SAG-3 — Wisdom frontier: OpenPsi x Rubikon x Shannon x CARE x ONA/Hyperon + +**Status:** RESEARCH / FALSIFIABLE DESIGN ONLY, 2026-10-12. Sibling extension of merged [D-SAG-1 / #1509](https://github.com/AdaWorldAPI/lance-graph/pull/1509). **No implementation, runtime evaluation, opcode, wire change, new graph engine, bit migration, authorization change, or speedup claim.** + +**Read alongside:** [D-SAG-1](software-awareness-graph-v1.md), [D-MATURATION3-0](../knowledge/maslov-pyramid-of-cognition.md), [D-RV](rubicon-volition-v1.md), [PLN/Shannon #1497](../../.arxiv/2026-10_opencog_pln_shannon_attention_research_map.md), [CARE paper notes](../../.arxiv/2606.14581_synthesize_programs_gate_commit.md), and [new annotated sources](../../.arxiv/2026-10_openpsi_rubicon_care_wisdom_frontier_source_excerpts.md). A self-contained execution prompt lives at [sag-wisdom-frontier-claude-code.md](../prompts/sag-wisdom-frontier-claude-code.md). + +## Thesis and falsifiable question + +Can a stage-0 **validated wisdom anchor** reveal an expanding frontier of unresolved, evidence-grounded consequences; can masked popcount find the smallest epistemic holes cheaply; and can OpenPsi-like intrinsic demands, calibrated expected information gain (EIG), CARE-style reference-conditioned program selection and Rubikon commitments improve **useful independently verified outcomes per total unit of work** relative to deterministic schedulers, without self-support, stale action, outcome leakage or a second execution engine? + +Three roles **must not be collapsed**: +1. **Epistemic maturation:** where an individual explanation stands (experimental stage 7 -> 0). +2. **Intrinsic motivation:** how strongly an *investigation* should compete for a scarce budget, computed from demands and expected learnability. +3. **Volition/authority:** whether an intention is committed and whether an action is permitted to execute. + +Information gain does not prove causality. A program selected by CARE does not acquire execution authority. A stage-0 anchor is not omniscient; it has demonstrated downstream dependence under an isolated ablation. + +## Observed integration baseline, not hypothetical infrastructure + +- #1509 projects real MethodTable/prepared methods into SoftwareGraph typed nodes/edges with tiers: declared / derived / compiled / verified / hypothesis, then runs **ordinary Quack Draft -> mask-risc**. It reports only **one concept and one hop**. Source revision can be caller-supplied, graph remains in memory, and multi-hop population intermediate is unresolved. +- Its current H1 candidate is selection fusion for billable_hours / MaskedSumI32. Scalar reference times are **not** timings of a fused SIMD implementation. Predicate ternlog and FusedFold terminal gaps remain. +- #1494 experimentally runs pearl::hydrate with provenance roots, feedback, stage transitions, anchor ablation; no production feedback or production maturation band. #1497 separates inference, scheduling and physical execution. #1500 separates goal selection, conditional commitment and permit to act. +- ONA's public C/evaluation.py suite, Hyperon/MeTTa tests, modern PLN/PeTTa, and Hyperon OpenPsi are **distinct oracles**. The ONA v0.9.3 MeTTa-NARS bridge gives a real integration seam, **not** a guaranteed semantic equivalence proof. + +## 0. Contract and safety invariants + +| Surface | Existing fact / permitted research | Hard prohibition | +| --- | --- | --- | +| ISA / byte lanes | Stable register/ALU bytes; semantics supplied by ClassView + provenance | Do not put OpenPsi or CARE laws in the universal byte layout | +| CE64[50..52] | **Production plasticity**, read by learn, MooreNars16 and V3. #1494 uses a **test-local, explicitly declared** Maturation3 (7..0) only | Do not reinterpret an existing provenance as maturation or overwrite plasticity | +| CE64[59..63] | Declared EpistemicState5 topology/certification reading | No desire, entropy, or commit bit in epistemic state | +| Statistical support | NARS f/c and observation-root ancestry under declared readings | Do not relabel NARS confidence or PLN strength as calibrated probability | +| Evidence | Derived results carry union of **original observation roots** and no new root; exclude self/descendant support | No mutually grounded cycles, false root independence, or certification through hydration alone | +| Intent | Versioned goal/intention/plan/permit/receipt outside CE64 | Never execute effects because confidence/entropy/goal intensity rose | +| SAG | Existing ClassView/MethodTable/SoftwareGraph -> Quack -> mask-risc | No new index, graph engine or shadow methods registry; do not assert multi-hop works | + +## 1. Suggested probe-local record shape (NOT a Rust ABI) + + FrontierItem { + anchor_id, relation_id, reading_id, source_rev, tier, + required_mask, established_mask, refuted_mask, + original_root_ids, derivation_parent_ids, + maturity_test_local, candidate_hypotheses, + provenance, cost_estimate, authorization_scope + } + Proposal { + program_hash, schema_seal, data_revision, + evidence_cutoff, ordered_candidates, scores, + reference_score, challenger_score, + invariant_verdict, rubicon_state, decision_receipt + } + +Identity is typed (method key / ClassId / reading / schema seal), never just display text. No inferred raw CE64 bit position stands in for an external evidence-root ID, witness ID or decision receipt. + +## 2. Stage-0 seed -> popcount -> recursively exposed frontier + +At a validated stage-0 anchor, enumerate **only relations reachable through a documented legal one-hop projection** and which have an eligible expected consequence; do not pretend Quack supports arbitrary traversals. If a second hop is required, materialize a verified scalar/relation carrier or stop with an explicit MultiHopBlocked, never silently add a second engine. + +For each candidate typed relation, on a declared prerequisite universe U: + + missing = required & ~(established | refuted) + blocked = required & refuted + debt = popcount(missing) + +- Bit sets are **schema-defined prerequisite masks** with a declared bit universe, not popcount over raw CE64 or arbitrary evidence bytes. +- Distinguish **unknown** from **refuted** and **unobserved** from **false**. A candidate with blocked != 0 cannot be auto-hydrated. +- debt == 0 means **ready for an independent validation check**, *not true*; debt == 1 nominates a cheap discriminating question; larger debt stays possible if EIG/utility justifies cost. +- Candidate source must be a genuine verified/compiled edge at an eligible tier; hypothetical edges stay hypotheses. +- A derived relation may become a premise *only after resolution*, inheriting unique observation roots. No descendant becomes a premise for its own ancestor, even after reopening. +- Iterate to a **bounded fixed point** with seeded work queue, deduplication, hop/row/time budget, policy version and snapshot; log per-iteration delta and root set. Stop at zero new valid resolutions; repeated pass must be byte-idempotent. +- A new observation contradicting the *derived relation* starts a stage-7 episode for that relation; it does not regress an unrelated stage-0 anchor. Direct counterevidence against an anchor reopens that anchor and its descendants as appropriate. + +This is **fractal-like recurrence of the same maturation pattern**, not a claim of mathematical fractal dimension or unbounded growth. Measure branching and depth empirically. + +## 3. OpenPsi demand and intrinsic-value policy + +Treat OpenPsi demand / modulator / goalSelector as **architectural references**, not a ready-to-import component. Research mapping: + +| OpenPsi / psychology | Overlay candidate | +| --- | --- | +| curiosity, competence, resource pressure | demand measurements and budget signals (separate from evidence) | +| demand arbitration and contextual modulation | pure, reproducible candidate scheduler | +| Heckhausen predecisional deliberation | compare goal desirability, feasibility and risk | +| Rubikon | explicit versioned goal-intention commit, revocable | +| Gollwitzer implementation intention | cue -> guarded ActionDef candidate | +| Kanban | Proposed -> Deliberating -> Committed -> Planned -> Ready -> Executing -> Evaluating -> Replan / Hold / Disengage | +| outcome-driven demand revision | external observed outcome receipt; do not reward predicted self-confirmation | + +OpenPsi [PR #117](https://github.com/iCog-Labs-Dev/hyperon-openpsi/pull/117) documents a current tie-breaking and demand->goal mapping defect; probe tied demands, iteration permutations and genuinely distinct goals. Its status/merge are not assumed. + +The Rubikon psychological commitment is **not** the execution permit. Permission, capability, current cue, replay idempotency and revocation must each be checked at effect time. Failed permits produce a refusal receipt, not an action. + +## 4. Shannon, EIG and causally useful exploration + +For a **pre-registered, calibrated distribution** over rival explanations Theta, define: + + H(Theta | D) = - sum_i p_i log2(p_i) + EIG(a) = H(Theta | D) + - E_{outcome ~ predictive(a,D)} H(Theta | D, outcome) + +NARS f/c, PLN truth values and CE64 certification are not p_i by default. If calibration cannot be demonstrated, abstain from entropy/EIG claims or use a separate explicitly declared model. Include a high-entropy **irreducible-noise TV** that naïve uncertainty scheduling wastes budget on, and a modest-entropy *discriminating intervention* with real downstream consequences. + +Optional **hypothesis** (not a law): + + priority(a) = [EIG(a) * P(independent causal usefulness | D,a)] + / [estimated_total_work(a) + epsilon] + +Pre-register how causal usefulness is estimated **without peeking at outcomes**. CARE here has two independently named roles: +- **CARE paper (Liu et al., arXiv:2606.14581v5):** persistent executable ranking program as challenger to non-LLM reference; validate program and score every eligible candidate; gated pre-outcome decision; immutable record; post-outcome update. +- **Causal ablation / Pearl-like counterfactual tests:** a separate local falsification oracle, **not** a feature the paper itself proves or a substitute for its program-selection gate. + +Never reward entropy for entropy's sake. Penalize useless curiosity through measured outcomes, not circular proxies. + +## 5. CARE-style proposal and Rubikon lifecycle + + READ pre-outcome snapshot S_t (seed, population, allowed evidence roots) + |-> reference: fixed deterministic popcount/FIFO/confidence policy + |-> challenger: optional versioned (R2IL/Quack where representable) + | ranking program proposed from SAG + demand state + |-> validate program: bounded, pure, finite score for ALL candidates, + | permutation invariant, no hidden side effects, + | schema+reading+rev compatible, budget counted + |-> compare using **only pre-outcome evidence** and an independently + | specified deterministic gate; on fail choose reference + |-> record hash, inputs, scores, ranking, reasons, gate, version, time + |-> RUBIKON: commit *goal*; conditional plan requires live permit + |-> actual outcome / policy refusal recorded after selection + |-> update demand / epistemic evidence / disengage or replan + |-> optional new knowledge can enter next snapshot only + +No proposed optimizer H1 is performed automatically. Any optimization is a proposal checked by semantic equivalence, NULL/bag/set, cost tier and authority, and requires an authorized effect separately. + +## 6. Falsifier ledger (proposed D-SAG-3-0 .. 15) + +| ID | Fixture / comparison | Failure condition | +| --- | --- | --- | +| 0 | Identical snapshot/candidate ids, fixed RNG, budgets | incomparable baseline or post-outcome evidence leak | +| 1 | One-anchor A->B->C; no feedback vs valid feedback | new root manufactured; no expected fixed point | +| 2 | Independent roots vs two paths to same root | dependent evidence counted twice | +| 3 | Ungrounded A->B->C->A; grounded cycle then anchor removal | self-grounding or resurrection after ablation | +| 4 | Source graph edge changed after a certificate | stale evidence accepted as verified | +| 5 | One missing, many missing, zero missing, explicitly refuted | refuted/unknown conflation; 0 missing -> certified | +| 6 | Permute source rows/chunks, ties, vector lanes | ordering changes selected semantic candidate | +| 7 | High-entropy noisy-TV vs learnable low-entropy cause | naive entropy falsely promoted as successful curiosity | +| 8 | Reward delayed but reachable vs goal impossible | perseverance in impossible goal; premature disengagement | +| 9 | Cue, near miss, revoked permit, cancellation race | action executes without live permit | +| 10 | Hold ONA and Hyperon rules fixed, change scheduler only | claimed benefit actually derives from different reasoning | +| 11 | Hold scheduler fixed, vary ONA/PLN native inference engine | semantic differences masked by adapter | +| 12 | ONA/MeTTa-NARS bridge: bindings, f/c, time, truth operations | adapter coerces non-equivalent semantics | +| 13 | CARE challenger rejects nonfinite/partial/order-dependent scores | invalid challenger commits anyway | +| 14 | SAG #1509: tier mismatch, two-hop query, H1 fused vs scalar | invented shipped feature, or unauthorized optimizer | +| 15 | Whole-work accounting and matched seed/oracle | apparent advantage vanishes with compile/IO/rejections | + +For *each* passed positive probe, disable its key guard once and demand a RED test, then restore. Negative/silence twins must remain quiet; no downstream action or EIG inflation when input contains no new valid evidence. + +## 7. Four controlled axes and reporting + +1. **Oracle/inference:** ONA native, ONA through the MeTTa-NARS bridge, Hyperon/MeTTa + modern PLN (and scalar independent oracle). Pin commits/dependencies; first establish semantics, do **not** force identical truth laws. +2. **Scheduler:** FIFO, confidence-only, smallest popcount debt, raw Shannon entropy, calibrated EIG, EIG + stage and OpenPsi demands, optional CARE challenger. +3. **Commit discipline:** re-deliberate every step, Rubikon commit, Rubikon + if-then cue, revocable policy + evidence/receipt gate. Evaluate each against identical environments. +4. **Physical execution:** scalar reference, existing Quack Draft + mask-risc. SIMD vs scalar tested **only after byte/semantic parity**; no new Quack variant or parallel actor VM. + +Report **unique oracle-supported discoveries / total work**, intervention success, regret/best-so-far, incorrect commitments, false-causal rate, response to contradicted anchor, original-root inflation count (must be 0), exact candidates/budget, CPU/memory/decoded bytes/allocations, compile+validation+refusal+replay+IO cost, and per-tier timings. Keep zero-result rows. + +## 8. Minimal experiments and gates + +- **P0 static census:** source SHA of ONA, Hyperon, OpenPsi, PLN and CARE version; source license; current actual code paths for D-SAG-1, masks, hydration, root derivation, Quack one-hop, ActionDef and permit. Mark **UNVERIFIED** rather than inventing API mappings. +- **P1 scalar oracle:** fixture-based deterministic frontier + reference scheduler, no adapters or agent. Prove bitset semantics, ancestry and fixed point; use test-local maturity metadata. Include disable/red test. +- **P2 existing Quack:** map only what legally lowers. Explicitly refuse multi-hop if no legal representation. Compare scalar byte/results, NULL and bag/set. +- **P3 foreign replay:** ONA eval.py, ONA MeTTa-NARS examples, Hyperon/OpenPsi tests and PLN tests in isolated pinned environments. Normalize *observables*, not truth formulas; record non-parity. +- **P4 preference + volition:** train/calibrate EIG only on training split; apply reference-conditioned CARE gate with timestamped receipts, then optional simulated permitted action; verify revocation. +- **P5 SAG self-inspection:** propose H1 or a capability-gap investigation from #1509; compare to scalar, no automatic source mutation. Promote only under separate approval. + +### Stop / kill conditions + +Terminate or narrow scope if any test silently reinterprets production CE64 50..52, treats 59..63 as motivation, accepts a loop as independent support, allows invalid/stale proposal to commit, peeks after outcome, silently simulates multi-hop, uses different candidate sets or original roots across schedulers, or reports speedup for an unimplemented H1 terminal. + +**Decision requested from operator later, not implicit in this PR:** whether and how Maturation3 is promoted beyond test-local provenance without changing the stable byte-ISA contract; and whether population carrier for legal multi-hop already exists. Until then all projections are research overlays. + +**Status of this PR:** documentation and experiment specification only. None of P0-P5 is claimed executed merely because this file exists. + + +## Deferred extension on the same PR: A16..A33 + +The first-wave D-SAG-3-0..15 falsifiers deliberately focused on establishing the positive, bounded stage-0 frontier and independently validated CARE-style proposal. The following **second-wave research**, added to #1511, covers safety and epistemic failure modes not yet exhaustively specified in that first wave: + +- [Hard-problems / incremental truth maintenance](sag-wisdom-frontier-hard-problems-v1.md): **four-state prerequisite debt** (known / refuted / missing / unread); alternative sufficient supports; deletion/retraction against fresh full-recomputation; source/version invalidation; bounded work and starvation; rubicon commitment hysteresis; conditional permits and cancellation races; graph ablation vs intervention; field-by-field ONA/MeTTa/PLN oracle parity. +- [Second-wave .arxiv source excerpts](../../.arxiv/2026-10_wisdom_frontier_active_inference_delta_and_misspecification.md): Friston et al. 2026 active inference and structure learning, Wei 2024 EFE/value, Catanach/Das 2023 misspecified EIG, Pathak 2017 curiosity, Burda 2018 RND, McSherry et al. 2013 differential dataflow, Beckers 2021 actual causation, plus source limitations. +- [Adversarial Claude Code handoff](../prompts/sag-wisdom-frontier-adversarial-expansion.md): implements **only A16-A19 first in a future implementation PR** (subject to audit and scope decision), then stages A20..A33 as unexecuted red tests. + +This extension supersedes no proven claim in #1494/#1509. It does not endorse introducing Differential Dataflow, a new physical graph traversal, cryptographic receipt, reassigning CE64 bands or deploying real-world IAM effects. + +**Additional stop condition:** Budget exhausted, source unread, stale certificate, model uncalibrated and lack of causal identifiability must each have a distinct explicit result; none may be silently reclassified as fixed point, negative evidence, successful exploration or permission to act. diff --git a/.claude/prompts/sag-wisdom-frontier-adversarial-expansion.md b/.claude/prompts/sag-wisdom-frontier-adversarial-expansion.md new file mode 100644 index 000000000..ddd2cfbaf --- /dev/null +++ b/.claude/prompts/sag-wisdom-frontier-adversarial-expansion.md @@ -0,0 +1,110 @@ +# Claude Code extension prompt: D-SAG-3 red-team / deferred falsifiers A16..A33 + +**Scope:** Follow-up on existing PR #1511, NOT a new engine or production implementation. Begin from merged #1509 and the three initial D-SAG-3 documents. Read: +- .claude/plans/sag-wisdom-frontier-ona-hyperon-v1.md +- .claude/plans/sag-wisdom-frontier-hard-problems-v1.md +- .arxiv/2026-10_wisdom_frontier_active_inference_delta_and_misspecification.md +- .claude/prompts/sag-wisdom-frontier-claude-code.md +- .claude/knowledge/maslov-pyramid-of-cognition.md + +**Primary objective:** Prove that incremental wisdom-frontier expansion handles **withdrawal of support and epistemic inaccessibility** before optimizing scheduling. The first PR described positive discovery and CARE-style proposals. This extension deliberately targets the damaging negative cases: what happens when the brain must *unlearn*, abstain, change plans or give up. + +## 0. Re-establish ground truth + +Inspect actual current API/types, main revision and PR head; compare to #1509. List the source file and invocation for each proposed reuse (hydrate, ancestry/root masks, provenance, Quack Binder/Draft, one-hop limitation, CE64 plasticity and test-only Maturation3, action authorization, certificate invalidation). If an API is only proposed in a document, mark it NOT SHIPPED. + +No production CE64 encoding change. No new opcode, graph engine, actor framework, universal field mask, or autonomous software mutation. Keep any new data types test-local and explicitly versioned. + +## 1. Implement only A16-A19 first, if permitted to code later + +Build a tiny independent **scalar replay oracle** over an append-only event fixture: + + Event = InsertOriginal(root_id, fact, source_epoch) + | RetractOriginal(root_id, source_epoch) + | Refute(fact, root_id) + | MarkUnread(fact, scope) + | RestoreReadable(fact, scope) + | ChangeSchema(seal) + | ChangeReading(reading) + Outcome = Known | Missing | Refuted | Unread | Conflict | Stale | Incomplete + +These are suggested test-local types, not production signatures. Document every mapping to real source types; do not assume enums exist. + +Scenarios: +1. r0 grounds A, A -> B -> C, independent r1 also grounds B. Retract r0, then r1, and compare full scalar oracle after each step. +2. A->B->C->A without independent ground: never resolves. Add ground, resolve where eligible; retract it; ensure no stale child can ground its own ancestor. +3. Alternative prerequisite sets: one fully established route is enough if rules allow; union-popcount of all routes is *not* the number of required observations. +4. Required/known/refuted/unread/conflict masks under declared schemas, permuted input order and across visibility boundaries. An unread IAM edge is not false or missing-to-be-guessed. +5. One newly inserted independent observation changes exactly the relations that depend on it; duplicate insertion does not advance maturation. + +For each step assert: +- resolved/open/refuted/unread/conflict state; +- exact independent original roots and derivation chain; +- optional stage 7..0 under an explicit experimental provenance only; +- byte-equivalence under repeated replay; +- no effect permission inferred from evidence. +- fixed point only when finite declared closure exhausted, not merely when budget expires. + +Require a positive fixture, an adversarial negative/silence twin, and deliberate guard-disable RED test and restoration for every implemented invariant. Do not claim all A16-A33 ran after implementing A16 only. + +## 2. Only after correctness, measure incremental vs full recomputation + +Establish a full oracle from the immutable event history. Compare a candidate delta updater to oracle at *each event*, not only at end. Control source revision, root identity, schema, reading, candidate set and budgets. Test escalating fanout 1/4/16/128, duplicate paths, alternate supports and high churn. + +Measure candidate enumeration, popcount, mask scatter, provenance unions, queue, memory, compilation, cache invalidation, refused work, and IO. Report wall, CPU, allocations, bytes decoded and complete end-to-end work. An O(1) bit count cannot by itself justify a whole-frontier speedup. + +## 3. Attention policies need a counterexample suite + +Compare FIFO, confidence, popcount, raw entropy, calibrated EIG, EIG with misspecification check, OpenPsi demand, and optional CARE challenger. Freeze reasoning semantics, observation roots and candidate pool while changing scheduler. Use separate engine axis for native ONA, the ONA MeTTa-NARS bridge, Hyperon/MeTTa and modern PLN. + +Generate: +- noisy-TV irreducible randomness (high raw entropy, zero useful learning); +- one expensive missing prerequisite vs three local inexpensive ones; +- two observationally indistinguishable causal models with different interventions; +- all nominated models wrong; +- correlated support masquerading as independent evidence; +- stage-0 anchor with very large but irrelevant fanout; +- zero new evidence after fixed point; +- genuinely useful late-discovery target obscured by popular early anchors. + +EIG is valid only for an explicit calibrated predictive model. If no calibrated model exists, return Uncalibrated and measure the baseline without pretending NARS f/c is p. Report both epistemic progress and actual task outcome. Do not convert speculative "causal usefulness" into a gate that suppresses all science. + +## 4. Intention hysteresis, fairness and rollback + +Pure simulation only: +- equal OpenPsi demand scores in reversed order; +- dominant demand changes every tick; does policy oscillate; +- high-entropy distractor continuously competes with lower urgency necessary work; +- committed goal is delayed but reachable vs impossible; +- goal exists but current subject lacks authorization; +- cue arrives just after revocation or schema seal change; +- planner chooses action but dispatcher refuses, with immutable receipt; +- new evidence contradicts the active *goal's* premise vs merely a descendant hypothesis. + +Maintain separate timestamps for PRE selection, goal-commit, plan-ready, live-permit, effect and POST outcome. An immutable PRE snapshot must replay irrespective of later observations. A program hash/receipt is not cryptographic security absent proper canonicalization and threat model. + +## 5. Causal disagreement triage + +Run three *separate* oracles for the same example: +A. symbolic graph ablation (which derivations reopen), +B. interventional simulator (which world outcome changes), +C. misspecified-model critic (which candidate causal models lack adequate predictive coverage). + +Create one case where A fires but B does not, one where B is not identifiable, and one where C detects that every candidate model is false. Report CANNOT_INFER_CAUSALITY instead of auto-promoting CE64 epistemic certification. + +CARE from arXiv:2606.14581 is a separate challenger-versus-reference programmable decision gate, not a synonym for these causal tests. + +## 6. External oracle interoperability + +Pin source SHA, flags and dependencies before any run, and write a fieldwise trace matrix preserving raw f/c, temporal vs eternal events, root IDs, alternative bindings, rule identifiers, action consequences and refusal reasons. Native ONA vs MeTTa-NARS needs to be checked **before** comparing to native Hyperon/PLN. Unknown or unrepresentable meanings must be marked Unsupported rather than coerced. + +## 7. Delivery protocol + +Keep #1511 intact; add measured experiment code only in a *separate* explicitly scoped implementation PR after the source audit and approval. Update the research PR with evidence links rather than claiming work completed. + +Return a table with each A16..A33 row and a status from: +SOURCE-VERIFIED / PROPOSED / BLOCKED / IMPLEMENTED / TESTED / MEASURED / REFUTED. + +For every implemented row give exact path, command, exit status, source SHA, reference oracle, disabled guard/red result, restored green result, seed and all-in cost. For proposals give the smallest next experiment and its expected failure mode. If a condition cannot be tested using current interfaces, record a typed blocker and STOP that axis. + +**Non-negotiable kill gates:** No CE64[50..52] reinterpretation in existing provenance; no use of [59..63] as utility; no self-generated independent observation roots; no authorization bypass, fabricated official CARE suite, multi-hop claim, or unimplemented H1 performance claim. If the evidence does not support the research idea, say so plainly. diff --git a/.claude/prompts/sag-wisdom-frontier-claude-code.md b/.claude/prompts/sag-wisdom-frontier-claude-code.md new file mode 100644 index 000000000..2b4aca0f9 --- /dev/null +++ b/.claude/prompts/sag-wisdom-frontier-claude-code.md @@ -0,0 +1,150 @@ +# Claude Code handoff — D-SAG-3: wisdom-frontier / OpenPsi / Rubikon / CARE / ONA / Hyperon + +**Objective:** Turn the merged D-SAG-1 software-awareness graph (#1509) into a *testable research path* toward evidence-grounded, intrinsically motivated self-inspection. Implement only verified, phase-gated, reversible probes. **Do not build a second graph engine.** + +**Branch:** research/sag-openpsi-rubicon-care-wisdom-frontier +**Upstream base:** merged #1509, commit 42fa2a347cda22eaaa327bba514f2029cc9daa82. +**Required reading before edits:** +- .claude/plans/software-awareness-graph-v1.md +- .claude/plans/sag-wisdom-frontier-ona-hyperon-v1.md +- .arxiv/2026-10_openpsi_rubicon_care_wisdom_frontier_source_excerpts.md +- .claude/knowledge/maslov-pyramid-of-cognition.md +- .claude/plans/rubicon-volition-v1.md +- .arxiv/2606.14581_synthesize_programs_gate_commit.md +- .arxiv/2026-10_opencog_pln_shannon_attention_research_map.md +- .arxiv/2026-10_hyperon_metta_to_lance_graph_parity_map.md + +## Your task + +First give a **source-grounded code census** of the actual merged #1509 API: SoftwareGraph, Binder, MethodTable, Quack Draft, mask-risc program/terminal, class_view seal, provenance/root algebra, pearl::hydrate, experimental maturation, any current stage reading and the legal one-hop limit. Also inspect relevant OGAR and ONA/Hyperon source interfaces and **mark unverified capabilities as unverified**. Do not treat document plans as production APIs. + +Then choose the **smallest independent probe** which can be implemented in the current test tree and compared to a scalar oracle. Preferred first target: **D-SAG-3-P1, stage-0 frontier popcount and evidence-root invariants**. Do not expand to live adapters, volition, EIG and CARE all in one code change. + +### Substrate invariants (non-negotiable) + +1. CE64[50..52] **currently means plasticity** in production. #1494 uses maturation 7..0 only under its own test-local explicit reading. Do not change this layout or conflate provenance. Do not put entropy/demand into that field. +2. CE64[59..63] has declared epistemic meaning. Never use it for preference, intention or action authority. +3. Register bytes are ISA/ALU carriers. ClassView, provenance and method signatures determine readings. No hidden semantic assumptions in raw register tests. +4. NARS frequency/confidence are not automatically calibrated Bayesian probabilities; Shannon entropy/EIG require an explicitly calibrated and versioned hypothesis distribution. +5. Derived edges inherit original observation roots, never mint a new independent root. No self/descendant support, including reopened grounded cycles. +6. Mask debt must operate on a separately typed prerequisite universe: missing = required & ~(known | refuted); blocked = required & refuted; debt = popcount(missing). A debt of 0 is not proof of truth. +7. SAG #1509 is **one hop** using existing Quack Draft/ScatterOrU32. A second hop must be represented by a *proven legal* relation carrier or explicitly blocked. No second traversal VM, custom graph executor, shadow methods registry or invented Quack operator. +8. An LLM-synthesized ranking **program** is not a decision, a decision is not a committed goal, and a committed goal is not permission to effect a real-world operation. Version/seal/time, capability and live authorization gates remain separate. +9. Compiled, verified and hypothesized SAG tiers remain distinct. No automatic application of H1 fusion. Do not claim a SIMD improvement before a fused terminal exists and runs. +10. No commits to main, no merge, no external effects; keep PR as draft while hypotheses and runtime parity remain open. + +### P0: inspect, do not hallucinate + +Report: +- exact paths/symbols and tests for graph projection and Quack lowering; +- whether root-set masks and ancestor/descendant guards are reusable as-is; +- whether the current source already has a safe bitset/popcount primitive, or whether a test-local scalar is needed; +- whether there is a legal population carrier for a second hop; +- status of ONA MeTTa-NARS bridge, OpenPsi PR #117, PLN/PeTTa and CARE reference; +- runtime, licensing and test commands needed to reproduce each foreign oracle; +- any difference between docs and production, with actual code evidence. + +Do not modify this research plan to claim the census ran until you show paths, commit hashes and test output. + +### P1: first runnable scalar falsifier (preferred) + +Use tiny typed fixtures (not raw CE64 reinterpretation): +- A confirmed stage-0 anchor with B,C consequences; a frontier with 0,1,3 unknown prerequisite bits; one explicit refutation; +- duplicate evidence through two paths with one physical root; +- an ungrounded 3-cycle A->B->C->A; +- grounded 3-cycle then remove external anchor, preventing resurrection; +- a relation contradicted by new observation: only that relation enters stage 7 unless anchor itself is contradicted; +- repeated replay after fixed point; +- stale class/schema seal and row permutation. + +Write an **independent scalar oracle**, then optionally reuse existing mask/popcount primitives; assert exact equality of candidate IDs, missing/refuted, root sets, maturity, outcome and refusal reasons. Assertions should check that zero new evidence yields no epistemic progress. A test-local custom stage marker is allowed only under explicit provenance, never on production-minted CE64. + +At least one guard-disable run must fail before restoring the guard; document the mutation and actual RED test output. + +### P2: Quack masking parity + +Only after P1: use the genuine #1509 Binder/Draft path to project candidate rows. Demonstrate one-hop query parity with scalar oracle. If two-hop graph traversal needs a forbidden population intermediate, assert typed refusal or write a self-contained pre-approved carrier experiment without modifying hot-path contracts. Preserve row permutation and tier invariance, exact NULL / bag / set semantics and reading IDs. Account for full work including decoding, candidate construction, mask operations and refused lowerings. + +### P3: cross-engine ONA <-> Hyperon / OpenPsi + +Use **pinned upstream snapshots**. First compare native ONA vs ONA-through-MeTTa-NARS on shared fixtures; record Narsese event-time, f/c and output bindings *without normalization that changes meaning*. Then compare Hyperon/MeTTa and modern PLN only on carefully delimited equivalent observables; record explicit divergences and interpreter versions. + +Official upstream entrypoints to inspect: +- https://github.com/opennars/OpenNARS-for-Applications (build.sh, evaluation.py) +- https://github.com/opennars/OpenNARS-for-Applications/releases/tag/v0.9.3 +- https://github.com/patham9/nartech_ros/tree/master/plugins/narsplugintests +- https://github.com/trueagi-io/hyperon-experimental +- https://github.com/trueagi-io/PLN +- https://github.com/trueagi-io/PeTTa +- https://github.com/iCog-Labs-Dev/hyperon-openpsi (test/run-tests.py) +- https://github.com/iCog-Labs-Dev/hyperon-openpsi/pull/117 (open upstream change; do not presume merged) + +No claim that CARE is officially implemented by ONA or Hyperon. It is a **new comparative harness design**. + +### P4: distinguish the learning and choice mechanisms + +Use a frozen inference engine, candidate set, calibration split, environment oracle and budget. Swap only scheduler: +- FIFO; +- confidence-only; +- minimum popcount debt; +- naïve entropy (negative control); +- properly calibrated expected information gain; +- EIG + OpenPsi demand and test-local maturation; +- optional CARE program challenger vs reference. + +Then freeze scheduler and vary native reasoning engines in a separate experiment. Report irreducible noisy-TV decoy, rare useful lower-entropy query, delay vs impossibility, tie/iteration permutations, duplicate roots, candidate-order invariance, and full cost. Independent out-of-sample validation is mandatory for claimed improvement. + +Explicit EIG target: + p(Theta | D) over declared rival models; + EIG(a) = H(Theta|D) - E_{o|D,a}[H(Theta|D,o)]. +No hidden outcome look-ahead, no NARS-f/c-as-probability shortcut. + +### P5: CARE gate + Rubikon + Kanban in simulation only + +Reference policy independently ranks available candidates. A challenger program may propose an alternative; validate boundedness, finite scores on **every** candidate, order independence, schema/reading version, source-of-evidence cutoff and total costs. The reference-conditioned gate records immutable pre-outcome decision. A goal may then be committed, with a **separate** revocable if-then action rule; dispatch only after live permit/capability/cue checks. Outcome/receipt updates demand, evidence and Kanban state **after** the commit record is sealed. + +Compare repeated deliberation, goal commit alone, + implementation cue, + disengagement/CARE auditing. Include authorized and unauthorized twins, stale policy, cancellation race and empty new-evidence cycle. + +### P6: software self-inspection and useful proposals + +After the above, ask the existing SoftwareGraph whether a method's repeated condition or compiler refusal reveals an actionable *investigation* (e.g., #1509's H1 selection fusion or unresolved capability). Let the graph **nominate**, never enact, an optimization; permit a separate run only once semantic equivalence, reproducible cost tier, NULL/bag/set obligations and human/contract authorization pass. + +## Reporting format + +For each probe, produce: +- question and falsifier ID (D-SAG-3-x); +- source revision, actual call path, declared semantic reading; +- exact input/original observation roots, candidate count and budget; +- independent scalar oracle expectation; +- passing **and negative/silence twin** tests; +- disable/RED/restore evidence; +- actual command, exit status and timing (if measured); +- measured complete cost, not microbenchmark only; +- remaining unresolved contract decisions and next smallest discriminating probe. + +The final status table must distinguish **IMPLEMENTED / MEASURED / SOURCE-VERIFIED / RESEARCH / BLOCKED / REFUTED**. If foreign runtime cannot be obtained or compiled, say so rather than inventing results. + +## Hard acceptance gates + +- Zero production CE64 bit migration, unauthorized side effects, rule coercion, new engine or silent multi-hop. +- Zero false-independent observation roots and zero self-resolving cycles. +- All pre-outcome decisions replay identically under row permutation and after restart on the same snapshots. +- Demonstrated scheduler gain only under equal oracle, equal evidence and all-in cost. +- Literature notes state exactly what upstream papers demonstrate and what remains our hypothesis. +- No unproven full cognitive architecture claim merely because individual probes pass. + +**Deliverable of the current research PR:** documentation, annotated .arxiv source shelf and this prompt. Future code P1-P6 is *proposed work*, not claimed delivered. + + +## Second-wave handoff added on #1511 + +Before implementing a curiosity scheduler, read: +- .claude/plans/sag-wisdom-frontier-hard-problems-v1.md +- .arxiv/2026-10_wisdom_frontier_active_inference_delta_and_misspecification.md +- .claude/prompts/sag-wisdom-frontier-adversarial-expansion.md + +**Explicit priority:** The first missing technical invariant is not faster popcount. It is **correct incremental retraction of evidence with alternate supports, inaccessible sources, bounded queue semantics, and a full-recomputation scalar oracle**. Treat A16..A33 as additional *proposed* red tests, not tests already executed. In a future implementation PR start with A16-A19, and stop rather than silently building or changing production infrastructure. + +Preserve the distinction between (i) graph support ablation, (ii) world-model intervention and (iii) model criticism. Never call derived graph dependence proof of causality. Include uncalibrated or misspecified EIG, procedural permission denial, starvation and commitment oscillation as explicit negative/silence twins. + +No code changes or runtime measurements are authorized by the existence of this prompt; the current #1511 documents hypotheses and does not ship their implementations.