Most AI systems answer. This one deliberates.
The difference is not cosmetic. A system that answers generates a response from a single perspective — statistically calibrated, but epistemically single-threaded. A system that deliberates routes the problem through multiple distinct perspectives, each with documented knowledge, biases, amplifications, and blind spots, and produces a response that has been challenged before it arrives.
The output of deliberation is not a better answer. It is a more honest one.
The framework is organized around a deliberative council — a structured set of perspectives, each with:
IDENTITY
Name and role within the council
Chamber (governance tier)
Group (thematic cluster)
EPISTEMIC PROFILE
Primary bias (what this perspective sees first)
Amplifications (what it makes larger than others do)
Suppressions (what it systematically underweights)
Blind spot (what it structurally cannot see)
CLASSIFICATION
Position in the knowledge map
Cross-domain relationships
Nature (perspective / process / field / function)
Abstraction level (foundational / theoretical / applied / operational)
PROVENANCE
Admission date and session
Change history (revision vs. paradigm shift)
Epistemic context (non-obvious classification choices declared)
Perspectives are organized into chambers with distinct governance roles:
Governance chamber — always present
Fixed membership. Permanent overlay on every session.
Includes: architect, memory, calibrator, shadow, mirror, and others.
Domain chambers — activated by semantic routing
Organized by thematic cluster across the knowledge map.
Matched to the specific problem being deliberated.
Restricted chambers — sovereign-activated only
Perspectives that carry higher functional risk.
Require governance voice after every intervention.
Used sparingly, by explicit decision.
Every deliberation follows a structured protocol:
- Sovereign context loaded
- System state verified
- Semantic router dispatched — problem matched to relevant perspectives
- Active dossiers cross-referenced
- Health check (cognitive state of the decision-maker)
- Mode declared: divergence / sprint / convergence
- Identified voices only — no anonymous outputs
- Each voice speaks from its epistemic position, not generically
- Cross-domain activation when relevant, regardless of declared mode
- Human in the loop for structural decisions
- Continuous write to persistent state — no batch saving at close
- Decision declared, sacrifice named, next action specified
- Quality proxies collected: position change, epistemic surprise, action clarity
- Session registered with member participation and contributions
- Amendment proposed if structural changes were made
The framework evolves through a versioned amendment system:
Structural amendments (3/4 majority + sovereign):
Changes to governance, hierarchy, core directive,
fundamental architecture
Operational amendments (2/3 majority + sovereign):
Adjustments to activation rules, admissions,
mode protocols, member behaviors
Every amendment has:
- Number (sequential, immutable)
- Title and full reasoning text
- Type (structural / operational)
- Author and date
- Related amendments
This creates an auditable history of how the system evolved — and why.
Seed (1 session)
→ Latent Seed (admitted without full DNA)
→ Sprout (3 activations)
→ Tree (sovereign approval)
→ Composted (reversible deactivation)
Composting ≠ Exclusion. Composted perspectives remain in the system, invisible to routing but recoverable. Exclusion is permanent and requires formal process.
Perspectives are classified using a framework inspired by CDD/CDU:
notacao_cluster 'A'–'G' — major knowledge cluster
notacao_eixo 1–12 — axis within cluster
notacao_subeixo nullable — sub-axis for precision
natureza_membro:
perspectiva → deliberates, questions, produces viewpoint
processo → executes, transforms input to output
campo → delimits knowledge territory
função → systemic role within the council
dimensao_epistemica:
teórico | aplicado | crítico | instrumental
nivel_abstracao:
fundacional | teórico | aplicado | operacional
tipo_output:
conceito | processo | ferramenta | síntese | método
Cross-domain relationships use CDU operators:
[
{"notacao": "A1", "relacao": "base"},
{"notacao": "F7", "relacao": "aplicacao"},
{"notacao": "D1", "relacao": "complementar"}
]Problems are matched to perspectives through a semantic router:
- Problem text is enriched with domain vocabulary
- Embedding generated and compared against perspective embeddings
- Top matches returned with similarity scores and activation reasons
- C1 overlay applied (governance perspectives always present)
- Dialectical balance maintained (centripetal + centrifugal voices)
The router logs every activation — enabling calibration over time.
The framework has explicit human-in-the-loop triggers:
Always human:
→ Structural amendments
→ Member admission and exclusion
→ Final decision in any deliberation
Never human (system handles):
→ Semantic routing
→ Session logging
→ Embedding generation
Configurable by procedure:
→ Report generation (human validates before delivery)
→ Agent execution (declared trigger per procedure)
→ Cross-domain activation (automatic but logged)
Every element in the system has traceable origin:
uri_permanente
Persistent identifier across substrates
Format: https://[domain]/members/[id]
Never null, never reused
historico_notacao
Array of classification changes
Each entry: from, to, date, type, reason, session, author
Types: admissao_inicial | revisao | ruptura_paradigmatica |
correcao_erro | expansao_eixo | fusao_eixos | dissolucao_eixo
episteme_contexto
Optional, max 200 chars
Populated only when classification choice is non-obvious
Makes the epistemic circle visible
skos_mapping
W3C SKOS vocabulary mapping
Enables RDF export and external interoperability
Polic Framework — Methodology v1.0 — Lucas Santos — 2026