Divina Architectura

Divina proposes an HCKG and human-authored rule scaffold supporting concept bottleneck perception, graph/text representations, MAML and distillation, causal reasoning, error-driven learning, modular hardware use, Graph-RAG, planning and human-directed MEND edits. Its allegory organizes intended capabilities and a staged evaluation roadmap. Component evidence and documentary correspondence leave the integrated learning, grounding, causal and alignment claims open.

This first standalone documentary dossier preserves the complete original, accepted full assessment and current catalog. It adds reading guidance and exact historical selections. Current dossier review and lead adoption follow this frozen submission separately.

Read the material

Material	Role
Complete original	Exact 56,742-byte source, 391 lines, four tables and 77 bibliography entries.
Full assessment	All 44 claim objects, scopes, evidence references and 70 literal quotation passages.
Current catalog	Accepted scientific assessment identity and original review/lead links.
Historical review guide	Four assessment stages, joint review scope and separate documentary acceptance.
Selected evidence	Exact historical values/ranges, current PDF contract and three explicit links.
Machine record	Separate statuses and five existing comparison routes.
Provenance	Full-copy origins, original path bases and versioned external references.
Rights	Unknown rights, including four distinct historical and derivative roles.
Manifest	Ten member hashes; external handoff binds the manifest itself.
Placement	Eleven unique prospective repository targets and roles.

Read every claim with its kind, status, rationale and actual evidence limits. C31 is a documentary identity derivation. C40–C44 are five conditional assessor arguments. The original shared review has 79 claims across Divina and BP-0021; Divina has 44. Historical scientific reading, earlier PDF page reading and the current metadata checks retain separate attribution.

Knowledge, learning and control

C01–C16, C24. The HCKG ontology, human inviolable rules, causal assumptions, concept/target labels, worlds, goals and evaluation labels are given information. They are separate from learned representations, specialists, causal structure and revisions. Hierarchy/default logic, causal models and similarity graphs have different semantics; names and edges alone do not establish external truth. The Formele Definitie at lines128–129 describes constrained local interaction in prose; it supplies no mathematical dynamics, objective or guarantee.

Primary concept-bottleneck evidence uses concept and target supervision. Its model-level concept intervention differs from a world intervention. The Divina proposal itself does not say every concept activation is supervised. The cited GraphFormers work resolved to textual graph nodes, not a validated pixel/audio frontend. MAML relies on a meta-training task distribution, support/query data and inner/outer optimization. A larger external pretrained teacher is explicitly allowed at lines144–147. Charge those priors and teacher/meta preparation costs alongside the incremental small-learner adaptation.

C17–C23, C40–C42. Error-driven updates need acquisition, uncertainty, validation and retained-task criteria. Distillation/pruning and Graph-RAG are components. The inspected compression example uses an unpruned 8B teacher fine-tuned on 94 billion tokens; it is not a training-free system budget. The Graph-RAG tutorial was partly read and its code was not executed. Retrieved context alone does not logically constrain an unconstrained generator to entailed outputs. Deterministic templates or sound output validation are constructive alternatives under their own contracts.

The source explicitly supplies human-specified inviolable rules and immediate rejection of every violating plan at lines175–177. C41 preserves a positive conditional guarantee: all effect-relevant state is modeled, deterministic total transitions match reality, actions are enabled, initial state satisfies I, every serial effect passes a sound gate using fresh state, and unknown/timeout blocks execution. Induction preserves I at each finite step; universal checker completeness is unnecessary. Preconditions would require applicability checks. Omitted effects, concurrency, stale state, incorrect models or unauthorized rule changes fall outside those premises. Rule compliance is not all values or authority.

C40's explicit two fair-bit causal models have the same passive distribution and different do(X=0) answers. A supplied valid causal model can answer its own query; the witness concerns identification without directional assumptions/interventions. No trained Divina model or real causal-discovery campaign is evaluated here.

Causality, explanation and evaluation

C25–C36. Distinguish concept-label accuracy, plan validity, actual execution, social benchmark accuracy and human expert explanation judgments. A graph path, local sensitivity or surrogate rule does not by itself establish external causal fidelity. HUMAINE comparative human-experience judgments are separate from causal-graph truth. The source retains human experts for social explanations.

The CEF paragraph and reference67 name Causal3D; the overview table names CausalVQA. Both supplied formats preserve that discrepancy. The exact Causal3D reference supports scoped synthetic discovery/intervention regimes; it does not define Divina CEF. The distinct candidate CausalVQA video multiple-choice work has a short inspected HTML and unresolved header/internal date conflict. Repository search text is not a checkout or code audit, and this candidate is not asserted to be the intended replacement. CGS/CEF/ESR definitions, thresholds and held-out protocols still need specification.

The cited Rule Engine label4.5.3 differs from inspected current5.2.1 documentation; Experta was not audited. Primary MAML/CBM PDF text was partial; screenshot references did not produce actual inspected pixels. GraphFormers resolved v3. The blueprint's MEND expansion says using Gradient Decomposition; the inspected primary section 3 heading says Model Editor Networks with Gradient Decomposition. The paper title is Fast Model Editing at Scale. Preserve those roles literally.

Adversarial/OOD tests, fairness statistics, self-diagnosis and retention measures are proposed evaluation conditions, not achieved robustness, all ethical judgement, awareness or continual learning. Specify threats, shifts, groups/objectives, ground-truth error labels, old-task support, calibration and independent scoring. Classifying one's own error category is distinct from reliable self-repair.

Editing, hardware and comparisons

C37–C39, C43–C44. A human provides the atomic X/Y/Z correction. Symbol-to-pair translation, pretrained base, chosen weights, editor training, locality and equivalent-example data are additional inputs. MEND deployment on one pair differs from editor preparation. C43 considers two question strings about the same entity: an edit can fix one while leaving the equivalent one wrong. That is a possible function-level counterexample, not an observed MEND failure or an invented universally quantified source rule. Reliability, locality and generalization need separate tests.

Symbolic storage is mainly CPU/RAM; specialists load into GPU memory on demand. Raw packed parameter bytes Pi*bi/8 do not include activations, caches, scales, workspaces, training states or transfer buffers. FP8 and experimental FP4 do not guarantee accuracy, utilization or energy. Manufacturer AMP/HAGS GPU scheduling context does not demonstrate a general CPU-symbolic/GPU orchestrator. Nominal 32GB capacity and575W TGP are not application measurements. The four-phase year1/year2–3/year4/year5+ roadmap and all competence targets remain proposals.

The five actual simpler_comparisons routes use: a conventional graph/rule checker and planner with the same HCKG/authority; the same encoder with direct prediction, concept auxiliaries or CBM without GNN; adaptation and small-specialist baselines with equal teacher/meta data and total compute; ordinary causal methods with equal variables/assumptions/intervention access; and graph templates, verified retrieval, direct correction/local finetuning with separate fidelity, retention, edit, runtime and human-work scores. Charge all labels, rules, priors, acquisition, training/tuning/failed runs, CPU/RAM/GPU/storage, energy and expert work. There is no research_questions array and no additional question or experiment here.

Documentary chronology and delivery

The 23-page alternate-format PDF is accepted under reconciliation004_documentary_v1. Current supplemental /sources/8 and sidecar /entries/7 link D01→C10, D04→C39 and D05→C31. D01's title page is separate from the mapped prose on page8. ESR continues metric column3, locality/generalization column4; Implementatie continues phase-three activities, column4, before the deliverable in column5. The PDF and whole shared ledgers stay external.

The copied historical assessment predates full registration and keeps its old unregistered wording verbatim. Later live contract activation is recorded separately. Author/reviewer each originally viewed43shared pages, including Divina's23; lead twelve shared spot checks were bounded. No new PDF pixels, extraction, rendering, primary research, causal/editing/simulation or GPU run is performed for this dossier. Documentary agreement is not independent science.

Four assessment stages are pinned: earliest to intermediate73 scalar changes, intermediate to independent intake5, and intake to final exactly2 review fields. Earlier stages include substantive author refinements; their75 distinct changes overlap and must not be added. Only the last stage leaves scientific claim objects unchanged. Exact reverse operations reconstruct every earlier byte version.

Planned private destination Flynn/bp-0037 is WR-owned. Actual repository, commit and platform IDs are null. Rights are unknown, public release is unauthorized, and included answers are exposed study material. The lead decides adoption and sealing, retaining the exact owner handoff locally under the separate WR hold.