Democratia

Democratia proposes groups of small neural perceptors with local symbolic knowledge and defeasible rules. For a task, agents would vote or delegate through a directed graph, aggregate weights at sinks, ratify knowledge and revise beliefs. The source selects weak computational emergence and excludes strong emergence. The integration's grounding, one-shot generalization, complete transparency, alignment and hardware targets remain unvalidated.

This first standalone documentary dossier preserves the complete original, accepted full assessment and current catalog, with scoped historical guidance.

Read the material

Material	Role
Complete original	Exact 78,083 bytes, 343 lines, intact displayed mathematics and 63 references.
Full assessment	All 15 claim objects, literal anchors, epistemic roles and evidence limits.
Current catalog	Accepted final assessment and historical review/lead identities.
Historical review guide	R1/R2 corrections, attested stages and later metadata closure.
Selected evidence	Exact historical selections, original sink-weight line and component index.
Machine record	Separate statuses, three comparisons and one existing unexecuted question.
Provenance	Copy origins, actual reference bases and unavailable-version limits.
Rights	Unknown rights and distinct copied/derivative roles.
Manifest	Ten member hashes; handoff binds the manifest externally.
Placement	Eleven unique prospective repository targets and roles.

Read each claim with its kind, status, rationale and actual primary scope. Historical content acceptance, the current independent dossier review and lead adoption remain separate. This documents the source's neural-symbolic engineering proposal.

Delegation, weight and evidence

C02–C03, C06. The source permits at most one outgoing delegation edge per decision. A sink's intact printed expression counts its own unit ballot plus one for each direct or indirect delegating ancestor. The exact original line 90 is selected unchanged. Cyclic members lose voting weight in the source; upstream voters feeding cycles, abstention, ties and changing snapshots still need precise behavior.

The checked uphill-delegation model is acyclic under its competence conditions. It assumes independent original voter ballots; effective votes become correlated after delegation. Neural agents sharing observations, models or layers cannot inherit independence from a count of delegates. The registered M-TRANSITIVE-DELEGATION record at /28 links only C02 and the same evidence. It is a component index, not another independent test or a truth certificate.

The source genuinely allows topic-dependent trust and revocation based on observed performance. Correctness feedback, calibration, exploration and delayed credit updates are missing. Using majority agreement itself as correctness can reinforce correlated error; this is a conditional risk, not a claim that every adaptive router must fail.

C06 is a historical assessor derivation. Unit-ballot ownership, neural certainty, reputation and multiple feature experts are different quantities. Separate proposition/feature graphs and an aggregation rule could reconcile them, but the blueprint does not define that arithmetic. Copying one perceptual observation along several paths does not create new independent support. No missing weighting formula is invented here.

Grounding, knowledge and revision

C04–C05, C08. Agents have trained neural components, local knowledge and suggested nonmonotonic rules. Default Logic is an example, not a complete semantics. Conflict priorities, extension selection, uncertain facts and the relationship among soft, probabilistic and defeasible logic remain open. A printed confidence of 0.85 is not automatically calibrated.

The H example supplies feature labels and counts while leaving relative geometry unspecified. Weighted agreement can admit a symbolic fact under a governance rule without independently establishing its referent or task correctness. The inspected softened-grounding work uses observed task outputs, learned mappings and predefined symbolic constraints. It does not validate consensus as a self-sufficient definition of meaning.

The spork example includes a user description and existing spoon/fork/texture detectors. Ratification can store that supplied example. Recognition of new category members, boundary discrimination and learning the prior detectors require separate held-out evaluation. Count pretraining, labels, knowledge engineering, feedback and communication instead of treating graph insertion as demonstrated one-example generalization.

C09–C10, C15. Preserve the real safeguards: weight decay, soft/hard delegation caps, minority-triggered second opinions, local committees, pre-vote consistency checking and conflict debate. The assessment evaluates their definitions and limits; it does not reduce the design to a naive fixed majority vote.

A direct incoming-edge cap need not cap transitive weight. Decay needs conservation or renormalization rules; second opinions and debate need termination budgets. Sparse committee communication may reduce traffic while losing evidence or creating bottlenecks. Grouping, migration, representative selection and stale state need protocols.

Checking direct conflicts with high-weight facts is constructive but does not check every consequence. Two concurrent acceptable insertions may conflict jointly; deletions affect dependent conclusions and separately voted propositions can be inconsistent together. The color example also needs single-color, time and context assumptions. Logical consistency, factual truth and effect authority remain distinct.

Explanation, metrics and resources

C01, C07, C11–C14. Auditable delegation paths and rule firings expose selected operational reasons. Completeness requires defined instrumentation, state, versions and replay; neural activation visualizations do not establish global causal or semantic explanation. Local rules and governance alone do not enforce all global invariants or authorize external effects.

The source's weak-emergence scope is preserved. A guided-self-organization editorial motivates constraints and research questions without proving unique optimality of this control scheme. Modern hardware can accelerate specified workloads; it does not remove unspecified search and communication costs.

CUDA streams allow asynchronous dispatch; actual overlap depends on resources and synchronization. Thousands of queued kernels are not thousands of resident models. Include active-model distributions, shared layers, training, graph states, CPU symbolic work, transfers, batching and end-to-end scheduling. GPU feasibility and the full R&D/training/inference budget are targets.

TS requires an independent definition of which causal steps should be recorded. RRI needs a task oracle beyond internal consistency. DER needs independent correctness and held-out recognition rather than successful insertion alone. PCM is only inspired by Gini: ordinary finite-population Gini has maximum (n−1)/n, so a target value of 1 would require a declared normalization. No metric or concentration value was measured.

The four curriculum phases and proposed accuracy, robustness, transparency, acquisition and concentration outcomes are prospective. Three actual comparisons use the same perceptors with calibrated ensembles/direct voting; a fixed expertise router with bounded communication; and a central typed knowledge graph with provenance, consistency and explicit revision. Match information, model capacity, communication and total compute.

One existing optional question, unexecuted: “Does dynamic delegation improve held-out accuracy and calibration over direct voting or calibrated ensembles using identical base models, especially with correlated errors?” No new voting experiment is commissioned.

Historical evidence and local delivery

R1 removed an inaccurate carryover description of damaged mathematical cells. The source formula is intact. R2 corrected the algorithm/theorem locator to printed 1238–1239 and distinguished original ballot independence from effective correlation. Original reviewed and corrected-content identities are attested by the historical review/ledger. A bounded lane search finds only the final full assessment. Metadata reversal matches the attested corrected-content digest, without claiming a direct comparison with an unavailable old full-file snapshot.

The joint review covers 30 claims across Democratia and Machiavelli–Nonary. Its game/equilibrium/coalition/Original999/counting material remains other-source context. The five current mechanism/PDF-register receipts use bp-0041-c and actual title terms. Two field hits identify one existing C02 mechanism; the four other destinations have no direct source relation within this bounded query.

Rights are unknown and all included answers exposed. Planned private destination Flynn/bp-0041 is WR-owned; actual repository, commit and platform IDs are null. No new scientific/political campaign, model, paper/PDF retrieval or hardware benchmark occurs. The lead decides adoption/sealing and retains the exact owner handoff locally under the separate WR communication hold.