Dune Tech · Mentat Cognitive Framework

Project Kwisatz Haderach proposes pretrained multimodal perception, a logic/probabilistic causal reasoner and layered graph memory. Rule-guided validation, analogy, counterfactual models and active re-observation are intended to support a universal traceability imperative. Dune supplies design provenance. The integrated capabilities remain an unvalidated research proposal.

Read the material

Material	Role
Complete original	Exact 58,261 bytes, 318 lines, three tables and 50 bibliography entries.
Full assessment	All 32 claims and 51 exact full-slice source passages.
Current catalog	Accepted historical source/review/lead identities.
Historical review guide	Three stages, author readbacks and independent closure.
Selected evidence	Exact 23+2 deltas, primary scopes and historical analyses.
Machine record	Separate statuses, six comparisons and one derivative question.
Provenance	Three full-copy origins, old bases and version semantics.
Rights	Unknown rights and separate material roles.
Manifest	Ten member hashes; handoff externally binds manifest.
Placement	Eleven unique prospective targets in a planned private repository.

Read each claim with its kind, status, rationale and actual primary scope.

The proposed data path

C01–C08, C12–C15. Pretrained perception or prompted/trained modality modules emit provisional triples. Each needs entity identity, relation meaning, time, provenance and calibrated uncertainty. Well-formed output is an inspectable interface rather than exposure of every neural computation.

The reasoner checks supplied types/rules and integrates uncertain facts. A plausibility check does not independently observe whether this particular bird sang. Separate unknown, negative, rejected, accepted and superseded facts. A rejection can suppress a novel true event; a false but consistent candidate can pass.

Formal deductions require a language, premises and proof method. A Bayesian posterior, a KG edge, a DAG and a causal estimand are different objects. Trace logs can establish what the program did under recorded inputs, without certifying perception, causal truth, goal legitimacy or external permission.

Memory buffers, accepted episodic subgraphs, semantic abstractions and compiled procedures have distinct update rules. The source's validated-memory descriptions at82/157 coexist with the unfiltered-history description at248; preserve both and require an admission/version/retraction contract. Storing episodes does not supply abstraction or skill-induction algorithms.

Re-observation can request an image crop, higher-resolution analysis or speaker focus. Define query objectives, available observations, cost, stopping, selection bias and uncertainty feedback. Reprocessing the same ambiguous evidence need not create independent evidence. Model requests and any eventual external effects also need separate feasibility and authority checks.

Components and the composition gap

C09–C11, C16–C18, C22–C23, C27–C28. NumPyro supplies model primitives and JAX-based inference; HMC/NUTS does not directly sample discrete latents, so enumeration must be tractable or another method chosen. This does not connect Prolog entailment to probabilistic events or give causal meaning to variables.

NOTEARS uses the matrix exponential in a smooth DAG constraint under a linear least-squares setup. Acyclicity and a stationary optimizer point do not identify world causes. FCI can leave equivalence-class directions unresolved; PC and a specific PyWhy/DoWhy integration were not verified by that historical reading.

Observation, intervention and individual counterfactuals need explicit structural functions, exogenous noise and identification assumptions. A retrieved analogous episode can supply context; it does not replace a justified causal model or the case-specific hidden conditions.

The one-shot relational learner ranks candidate tails using a trained metric, background neighbors and embeddings. New test relations are evaluated over an existing entity universe. Count meta-training, support pairs, candidate types, prior graph and held-out relation/entity information. This does not establish arbitrary unknown-relation learning or zero total-data semantics.

Lobster is a scoped GPU Datalog precedent with fixed-point evaluation and top-1 proof provenance under a size limit. Its benchmark host/device resources, including an 80 GB A100 regime, do not demonstrate a complete consumer-GPU workload. SmolVLM's compact pretrained variants do not establish MCR triple accuracy. Bibliography46 and48 identify the same KG paper.

TMS records reasons and supports revision; the accessed support is abstract-only. Comparable confidence, conflict selection, multiple supports, cycles, history and rollback need definitions. Neither consistency nor the least-certain-premise phrase guarantees world truth or stable external meaning.

Three historical conditional constructions

C30. Both worlds give a rule Sings(x)→Bird(x) and Bird(b). Sings(b) is true in one and false in the other. The same proposed Sings(b) is type-consistent in both. A filter consulting only the supplied rule and Bird(b) cannot distinguish them. Additional observations can resolve this example; no actual perceptual error rate is inferred.

C31. With fair binary U, model A uses X=U,Y=X while model B uses X=U,Y=U. Both assign half the observational mass to(0,0) and(1,1). Under do(X=1), model A sets Y=1 while model B retains P(Y=1)=1/2. Thus those episode distributions alone do not select the intervention prediction. This does not deny causal inference with valid extra assumptions or test an implemented MCR.

C32. The sole positive R(a,b) fits H1 containing only(a,b) and H2 containing (a,b) plus(c,d). They disagree on R(c,d). Background structure, meta-training or expert priors can select between them, as additional information. The example does not rule out task-specific one-shot learning.

These are preserved historical author arguments, not new simulations, recovered source formulas or blanket impossibility results.

Targets, resources and comparison routes

C04, C19–C26, C29. The four-stage 36-month roadmap and MCAS dimensions are prospective. Phase2 targets strictly >95% correct/complete triples on 1000 images with 1–3 objects. Phase3 targets strictly >80% OSGS on new relations. The later dynamic mystery and CEQ/CRA/OSGS/TAI need frozen tasks, denominators, independent truth, calibrated human protocols and held-out exposure. Clear or plausible explanations need not be causally correct.

The source explicitly uses pretrained perception and a rich earlier memory. Charge all rules/ontology/annotation, model data/compute, graph construction, matching/meta-training, causal discovery, searches, pruning, re-observation, validation, explanations and transfers. Incremental adaptation and total costs are separate. Offloading shifts memory and latency; it does not remove them. Fix graph/model sizes, weights, activations, optimizer/workspace lifetimes, joins/proofs, posterior samples and CPU/RAM/NVMe costs before a hardware claim.

All six actual simpler_comparisons remain addressable: the same perception with ordinary schema/provenance/rules; exact supplied-model planning; direct fixed-SCM queries and matched discovery; information-matched relation ranking; budget-matched active observation; and competent indexed CPU logic versus GPU acceleration including compilation and transfers.

One derivative reading question, unexecuted: “Do the causal and memory layers improve independently labeled triple accuracy and query success over the same pretrained perception model with schema validation, ordinary provenance storage and rule checks at equal total cost?” This new wording derives from simpler_comparisons[0]. It is not an original research_questions entry; the assessment has no such field.

Historical stages and local delivery

The earliest author snapshot, polished review input and final accepted assessment are separately bound. Exactly 23 author scalar changes cover spacing, neutral reference wording and a narrowed Lobster locator, including explicit C19–C21 statement/coverage spacing. Reversing them restores 127,556 earliest bytes. Then only review_status/review_records change; their inverse restores127,651 polished bytes. The historical independent review requested no correction. The final full assessment and all 51 source passages remain unchanged here.

The joint Dune/MENT-A review has 71 claims/117 passages/15 primary records, rather than Dune's 32/51/eight. Other-source 31 readbacks and four analyses retain attribution. Five declared register searches, with whole-word Dune/Mentat aliases, yield no matching values within scope; raw mentat inside implementation is not a link. No PDF association or global absence is inferred.

Flynn/bp-0048 is WR-owned and planned as a private repository. Actual repository, commit and platform IDs remain null. Rights unknown, known answers exposed, publication unperformed; current review/root adoption remain separate. Owner handoff stays local under the WR communication hold. No new scientific, primary-source, PDF, model/solver, architecture or repository action.