Project A'Tuin

Discworld's Project A'Tuin proposes a contextual knowledge graph, Soar-inspired memories and production compilation, specialized guild agents, a utility-based orchestrator, neural interfaces, trained graph propagation and narrative/causal plan ranking. The intended integration, low-data competence, causal/narrative equivalence, consumer-GPU feasibility and alignment remain unestablished.

This first standalone documentary dossier preserves the complete original, accepted full assessment and current catalog. The exact original filename is Discworld _Blueprint.md, including the space before the underscore.

Read the material

Material	Role
Complete original	Exact 79,162 bytes, 416 lines, four tables and 58 bibliography entries, no final newline.
Full assessment	All 22 claims and 26 main/supplementary literal anchors with evidence limits.
Current catalog	Accepted corpus assessment, review and lead identities.
Historical review guide	Two versions, other-source CITIES-R1 and bounded PDF scope.
Selected evidence	Exact scoped historical values and original table/citation ranges.
Machine record	Separate statuses, four matched comparisons and one existing question.
Provenance	Exact filename, copy origins, historical bases and read limits.
Rights	Unknown rights and distinct copied/derivative roles.
Manifest	Ten member hashes; external handoff binds the manifest.
Placement	Eleven unique prospective repository targets and roles.

Read each claim with its kind, status, rationale and actual primary scope. The historical cities review covers 48 claims across Discworld and Duckstad; Discworld has 22. Current dossier review and lead adoption remain later, separate records.

Context and knowledge

C05, C07–C09. Four architecture layers are described. The four advertised cognitive pillars are not enumerated as four concrete algorithms or interfaces. Component names leave ownership, state transitions and dataflow to specify.

Cyc-inspired microtheories organize context-separated statements. Consistency within each context and across imports still needs a logic, bridge rules, provenance and revision. Observed, believed, inferred, hypothetical and counterfactual statements have distinct roles. A readable edge is not by itself a grounded fact. Failed primary Cyc access remains disclosed.

The representative GCN uses graph features, trainable matrices and local neighborhood updates. Fixed depth does not automatically affect every node. Embedding propagation does not justify asserting or retracting symbolic facts. Depth/connectivity, objective, update order, conflict handling and memory cost remain obligations; the component check is not a claim about every GNN variant.

Semantic facts, episodic snapshots and guild-owned procedural rules are distinct stores. Recall needs identity, temporal indexing and retrieval criteria. Explanation needs actual decision dependencies, beyond a similar prior episode. Storing information and generating vectors are different from learning a rule.

Guilds and conditional learning

C10–C11, C13–C16. Decomposing tasks into guild contracts can be useful with typed inputs/results, dependencies, allocation, synchronization, budgets, failure/retry and termination. Coupled subproblems can reintroduce search. The global utility and protected constraints remain unspecified; a utility score does not establish collective good or action authority.

Soar supports a genuine conditional learning mechanism. Supplied procedural knowledge and a solved substate can yield a production by tracing dependencies and variablizing conditions/results. Any successful action sequence is not automatically a universally applicable rule. Count exceptions, rule interference, matching and generation costs before claiming net held-out benefit.

Phase I explicitly teaches structured domain knowledge. Socrates follows from given premises. Phase III preserves the second isomorphic-problem qualification: a new rule can avoid repeated search when its conditions match. New board-game rules are supplied prior information. Rule-count growth and five coherent proposals need independent utility/correctness criteria. Charge domain models, labels, perception/GNN/evaluator training, solver skills and communication.

Narrative and causal inference

C02, C04, C12, C18, C21. A coherent story can be false about an external process. Narrativium ranking has no operational scoring function or complete causal model. Logical validity, causal identification, narrative preference, perceptual truth and permitted effects remain separate.

DoWhy separates model, identification, estimation and attempted refutation. PC/FCI require data, tests, search settings and assumptions; FCI can retain latent or unresolved possibilities in a PAG. Narrative plausibility does not remove nonidentification. A benchmark key revealing a simulator cause cannot prove that it was identifiable from the agent's observations.

Neural feature translation needs labels, calibration, typed observations, uncertainty and retraction. Generated language must preserve proof premises. The source's Narrativium/Patrician plans remain proposals; this dossier runs no generator, planner or causal experiment.

Criteria, evaluation and resources

C03, C06, C14–C20, C22. Preserve the phase table's 10M-node wording and later nodes-and-edges target with queries below 100 ms. A fixed workload, encoding, concurrent load and complete graph/model memory budget are missing. The consumer hardware constraint does not establish feasibility.

The three phase plans, 1,000 proofs, 100 multi-guild tasks, phase-II 90% causal accuracy, five solutions and 24-hour rule growth are targets. Later evaluation targets include NCI>4/5, <1% detected inconsistency, >95% causal accuracy, >90% paradox detection and WES>4.5/5. Their differing stages are not collapsed into a measurement. A finite proof test does not establish universal consistency.

Citation 56 on the proposed fine-tuned judge names an engineering article, not an identified validated NCI judge protocol. Judge training, held-out labels, calibration and leakage control are additional inputs. State tracking needs a specified transition schema. The robustness test legitimately permits recognizing a paradox, asking for clarification or reporting limits.

Human-rated understanding, sufficiency, actionability and trust can measure intended experience under a defined protocol. They require independent checks of correctness and fidelity; persuasive error can also increase trust. No new psychometric campaign or current software/leaderboard update is performed.

The four actual competent_simpler_baselines use ordinary indexed rules or Soar with/without chunking; supplied causal and predictive models at matched information; one controller calling the same solvers; and indexed graph queries with template/provenance explanations. Count compilation, matching, coordination, training, inference and memory costs.

One existing optional question, unexecuted: “Does narrative ranking improve independently scored plans beyond a causal model and ordinary heuristic under equal evidence?” This is a reading route, not a new experiment.

Historical versions and PDF context

The two available full Discworld assessments differ only in review_status and disposition. Exact reversal recovers the initial bytes; all claims and primary reading scopes stay unchanged. The joint CITIES-R1 correction belongs solely to BP-0047, not Discworld.

Candidate /records/6 and accepted /accepted_scan_results/6 retain separate stages. PDFT-002-03 selects BP-0042: 29 pages, no_indication_within_scan, no follow-ups. Historical author overviews cover all 29 pages with enlarged 16/25. The reviewer sample is overviews 17–18, 21–22, 23–24 and the same enlarged pages, with exact Markdown/table locators. The sibling Markdown equals this original.

No omitted formal Narrativium function, new diagram or extra technical content was indicated within that scan. It establishes no full character/cell/reference identity, intended definitions, precedence or scientific validity. PDF, sibling, previews, enlargements and extracts remain external. No new PDF viewing, rendering, extraction or reconciliation occurs.

Five bounded source/prefix/title/PDF queries include the actual bp0042-c prefix and straight/curly apostrophe forms. Only light-triage fields match, identifying candidate and accepted stages rather than independent scientific evidence.

Rights unknown, answers exposed, planned WR-owned Flynn/bp-0042 and null actual repository/commit/platform IDs remain explicit. Lead adoption/sealing and exact local owner handoff follow separately under the WR communication hold.