Faster decisions. Onboarding in days, not weeks. AI that responds with evidence.
Documentation that ages creates decisions based on obsolete facts. AI tools without grounding hallucinate. O Knowledge OS is the infrastructure that transforms corporate knowledge into executable asset — with limits, traceability and costs under control.
- In a sentence
- Knowledge OS treats corporate knowledge as a governed cognitive infrastructure, not as a passive repository of documents: a platform that turns it into an executable asset, with limits, traceability and costs under control.
- Format
- Presentation
- Depth
- In-depth reading · ~3 min for central idea · ~20 min complete
You prefer the essentials first? Read synthesis (~1 min)
Why Corporate Knowledge Fails — And Why It Costs
Before any architecture, three real problems corrode the strategic value of knowledge. Each of them has operational, financial and risk costs.
Documentation ages. Not the decisions.
RFCs approved in 2022 still guide architectures in 2026. Seniors who know the story real are at 1:1s — not in the wiki. When they leave, knowledge disappears.
AI that hallucinates is worse than absence of AI.
Generic LLM tools respond with confidence about context they do not know. A product decision made upon an invented response has real cost.
Critical knowledge lives where it cannot be audited.
Slack DMs. Unversioned presentations. Whose head is gone. Unrecoverable knowledge is not equity — it is strategic risk.
To scale the impact
- • Estimated 40% of senior engineers' time be spent answering questions that have already been answered — informally, without auditable registration.
- • Average onboarding in companies without structured KM: 3 to 6 months to full productivity. Each additional month is direct payroll cost without return.
- • When a senior engineer leaves, the knowledge that goes away is not in Confluence — is in conversations that have never been coded.
Because it matters to the board
The original specification was dense in engineering, but silent in business outcomes. This review translates architecture into three measurable outcomes and three avoided risks.
Time-to-Decision −60%
From "who knows this?" to "what evidence?" Architectural and product decisions no longer depend on tribal knowledge.
Cognitive reuse +3x
Specs, decisions and patterns are consumed by agents — they don't die in PDFs. Each new project is part of a semantic ballast.
Compliance by Design
Lineage between requirement → code → test → deploy. Audit is no longer archeology.
Three risks this architecture avoids
- Risk 1 — Knowledge Decay: fictional misaligned documentation of the production code, generating decisions upon obsolete facts.
- Risk 2 — Hallucination at Scale: AI agents responding without grounding in auditable sources, contaminating product decisions.
- Risk 3 — Shadow Knowledge: critical knowledge in DMs, decks and senior heads — not recoverable, not auditable, not scalable.
"The organization's most underused asset is not the raw data. It's the contextual judgment that no one could encode."
10 dimensions of critical review
Each item lists the gap found in the original proposal and the concrete countermeasure applied in this version.
01 Executive Clarity
Original Gap
- 100% technical language; absence of financial and operational outcomes.
- No transformation narrative — only component catalogue.
- No mapping between architecture and business KPIs.
Correction applied
- Executive layer on top: 3 outcomes, 3 risks, 1 quote.
- Each architectural component justified by business capacity.
- Roadmap with phase output criteria in board language.
02 Operationalisation
Original Gap
- Conceptual — "we'll use GraphRAG" without real-use flow.
- No engineer journeys / PM / SRE / architect.
- Lack of integration with existing tools (Jira, GitHub, IDE).
Correction applied
- Four knowledge journeys documented (see §4).
- Integration points: PR check, plugin IDE, Slack agent, ChatOps.
- Operational workflow with explicit handoffs between human/agent.
03 Governance Enterprise
Original Gap
- "Governance" mentioned without ownership, stewardship or approval flow.
- No semantic versioning of specs or decisions.
- Lifecycle of implicit knowledge (creation, review, deprecation).
Correction applied
- RACI model by knowledge domain (§5).
- WithoutSee of specs + ADRs + ontology; deprecation with window and replacements.
- Knowledge Council with monthly cadence and defined mandate.
04 Organisational scalability
Original Gap
- Central model — not scale for multiple teams/BU.
- High risk of semantic duplicity between domains.
- No contribution policy, no federation.
Correction applied
- Federated model: core ontology + domain extensions.
- "Knowledge Mesh" pattern inspired by Data Mesh.
- Hire semantic testing between domains to avoid drift.
05 Cognitive Architecture
Original Gap
- Monolithic memory — episodic and semantics mixed.
- No retrieval policy (recency × authority × proximity).
- Semantic conflicts implicitly resolved by the LLM.
Correction applied
- Separation of memories: episodic · semantic · procedural · contextual.
- Retrieval policy declarative with precedent rules auditable.
- Conflict solve explicit + drift contextual detector.
06 IA-First Strategy
Original Gap
- Agents mentioned no feedback loop, no self-healing.
- No cognitive quality gate between retrieval and generation.
- No confidence score, no contextual tracing.
Correction applied
- Cognitive loop Sense → Retrieve → Reason → Validate → Act → Learn.
- Quality gates: groundedness, faithfulness, answer-relevance, quote cover.
- OpenTelemetry + LLM spans.
07 Knowledge Engineering
Original Gap
- Undefined documentary granularity (chapter? paragraph? block?).
- Frail metadata; no canonical entities.
- No capability map or domain catalog.
Correction applied
- Atomic units: Claim, Decision, Constraint, Pattern, Risk.
- Base ontology + semantic schemas versioned in JSON-LD.
- Capability map linking domain → entity → spec → agent.
08 Traceability
Original Gap
- Lineage shallow: spec ↔ code without transitional dependencies.
- No impact graph for semantic changes.
- Validation does not cover requirement.
Correction applied
- Traceability matrix: requirement → spec → ADR → code → test → deploy → telemetry.
- Impact graph navigable. Human-in-the-Loop gate: changes that orphan requirements trigger assisted review — not automatic blocking.
- Cognitive cover: % of clans with traceable evidence.
09 Cognitive observability
Original Gap
- No retrieval metrics, semantic quality or coverage.
- No cognitive ALS (latency, freshness, groundedness).
- No health health dashboard of knowledge.
Correction applied
- 18 cognitive KPIs divided into 4 layers (§7).
- SLOs per domain: P95 retrieval < 800ms; freshness < 24h.
- Dashboard "Knowledge Vital Signs" with alerting.
10 Roadmap & Adoption
Original Gap
- Implicit Big bang — adopt everything, wait miracle.
- No quick wins, no pilots, no exit criteria.
- No map of dependencies between components.
Correction applied
- 4 incremental phases: Foundation → Pilot → Federate → Autonomous.
- Quick wins 30/60/90 days with measurable KPI per phase.
- Criteria for promotion between auditable phases (§8).
Traditional documentation × Knowledge OS
Board view: 9 dimensions in 30 seconds.
| Size | Current Template | Knowledge OS | Impact |
|---|---|---|---|
| Organisation | By technological stack — silos React/Node/DB. | By bounded context. | Cohesion +80% |
| Recovery | Manual navigation + lexical search. | Hybrid retrieval: vector + BM25 + graph traversal. | Precision 3× |
| Engineering | AI generates with shallow and ambiguous context. | Agents consume pre-computed and citable context. | Velocity 3–5× |
| Governance | Implicit Ownership; head decisions. | Versioned ADRs + Federated RACI + audit trail. | Compliance ↑ |
| Scale | Slow onboarding; tribal knowledge. | Knowledge mesh with domain ownership. | Onboarding −70% |
| Validation | Tests as a later check. | Tests as executable spec (contract/golden/property). | Bugs −60% |
| Traceability | Zero — code does not connect the requirement. | Policy → Contract → Spec → Test → Code → Deploy. | Audit 100% |
| Cognition AI | Doc long pollutes the context window. | Atomic units typed; chunk-optimal. | Hallucination < 2% |
| Decision | Meetings to find out what's going on. | Question → Answer quoted in seconds. | TTD −60% |
Seven layers of Knowledge OS
Each layer has unique responsibility, contracts explicit and SLA itself. Replaceable without rewriting stack.
Standard Cognitive Loop
Every interaction goes through this pipeline. Each step is instrumented and auditable.
Self-healing: VALIDATE failures re-activate RETRIEVED with expanded context. Backpressure: Low confidence in REASON triggers scaling for human via approval gate.
Your AI has a limit. And you know exactly where he is.
Self-employed agents can save hundreds of hours — or burn budgets in infinite loops. Knowledge OS comes with Cognitive Circuit Breakers set by default: autonomy with containment.
Hard Escalation Flow
Every cognitive execution runs through the default loop. Failures in Validate return to Retrieve with expanded context — but only within a limited window of attempts. Once the limit is reached, the Circuit Breaker is triggered.
Limits per autonomy phase
Each phase of the roadmap operates with distinct limits and scaling mechanisms. Limits are configurable patterns via admin panel.
| Phase | Agent autonomy | Limit of attempts | Scale mechanism | FinOps Protection |
|---|---|---|---|---|
| 1 — Foundation | Read-only | — | Always human. | Predictable cost (without writing) |
| 2 — Pilot Cognitive | Consultation + suggestion (HITL) | 1 attempt | Human-in-the-Loop in each action | Budget alerted at 70% / 90% |
| 3 — Federate / Agentic Assist | Supervised execution | 3 attempts | Hard Escalation (human review) | Monthly Budget + Scheduled Alerts |
| 4 — Autonomous | Autonomous execution (write) | 3 attempts | Hard Escalation Mandatory | Required configured Budget |
FinOps for AI — operated metrics
Each agent execution is treated as a measurable resource — analogous to the FinOps of cloud, but for cognitive infrastructure.
Non-negotiable cost governance principles
- • No roof, no autonomy. Phase 4 (Autonomous Agents) cannot be activated without a monthly budget configured and approved.
- • Every lineup is auditable. Each Hard Escalation generates event with context, estimated cost in USD and consumed tokens.
- • Scheduled alerts. Notifications at 70%, 90% and 100% of the monthly budget — before the incident, not later.
- • Configurable limits, not hard-coded. The 3× attempts are standard; each org adjusts via admin panel as an appetite for risk.
Documentation that pays for those who document well.
The reason why documentation has always been outdated is not lack of discipline — It's a visible lack of incentive. Knowledge OS changes the equation: who documents well has their work multiplied by agents, And that comes out for the whole team.
Punishment model (which does not work)
- Motivation: prevent the deploy from being blocked.
- Quality sign: pass the quality gate.
- Behavior: minimum viable documentation to pass.
- Feedback: Pipeline error — no recognition for good work.
- Result: malicious compliance. Low quality. Frustration.
Reuse Index (which changes)
- Motivation: have the work recognized and multiplied by agents.
- Quality sign: Your content solves problems from other teams.
- Behavior: documentation as a visible professional investment.
- Feedback: personal impact dashboard cross-team in real time.
- Result: organic culture. Growing quality. Pride to contribute.
How the Reuse Index Works
From contribution to recognition, in 5 traceable steps. Each step generates an auditable event in Knowledge Graph.
You document a decision or pattern in Knowledge OS
An ADR, a spec, a runbook. Tipped, quoteable, with semantic frontmatter.
An agent on another team uses its content to solve a problem
Citable retrieval: each answer loads the source. Your work is no longer invisible.
The event is registered — you are credited as source
Quote appears in the Trace OpenTelemetry. Complete lineage: query → reply → your spec.
Your Reuse Index goes up. Managers and peers see the contribution
personal Dashboard and team. 1:1 conversation and review stop being about assumptions.
In the next sprint, you document more — because it's worth it
The incentive becomes structural. Culture emerges from the system, not from a poster on the wall.
Reuse Index Metrics
Individual and team indicators, displayed on personal dashboard. They measure professional impact rather than policing.
How often did your content help other teams
Weighted sum of decisions influenced
% of its contributions still valid
How many specs refer to your work
" I used to document it because I had to. Now I document because when an agent uses my spec to solve a bug on another team, my name appears on trace. It changed how I think about documenting."
Typed, granular and addressable knowledge
Without computable typology, agents receive noise. These are the 6 canonical types + the frontmatter that makes each artifact citable + the layout single-source-of-truth.
The 6 canonical types of atomic unity
Truths of the realm
Objective and stable invariants. Not an opinion.
type: fact id: FACT-001 # "Pedido tem estado"
Active rules
Business rule valid and enforceable.
type: policy id: POL-042 # "Cancelar requer auth"
Versioned ADRs
Architectural decision with context and consequences.
type: adr id: ADR-019 # "Adotar CQRS em Orders"
APIs and schemas
Formal interface between components/services.
type: contract id: CNT-007 # "POST /orders spec"
Non-functional restrictions
Performance, compliance, safety — hard limits.
type: constraint id: CON-003 # "p99 < 200ms"
Executable Runbooks
Step-by-step operational process, auditable.
type: procedure id: PRC-012 # "Deploy rollback"
Semantic Frontmatter — Full Example
Each document loads metadata that makes it citable, traceable and indexable by agents. No frontmatter, no computable semantics.
--- id: FEAT-042 type: feature-spec title: "Checkout Express" domain: orders status: approved version: 2.1.0 owner: "@squad-checkout" created: 2026-03-15 updated: 2026-04-20 depends_on: - FEAT-018 # auth-flow - CNT-007 # orders-api implements: - POL-042 # cancel-policy - ADR-019 # cqrs-pattern validated_by: - TEST-FEAT-042-001 - CONTRACT-FEAT-042 tags: [checkout, payments, critical-path] context_for_agents: - checkout-agent - payments-agent retrieval_weight: 0.92 freshness_sla: 7d ---
Canon directory structure
Layout single-source-of-truth. Each folder has owner, contract and CI gate own. Semantic indices generated in build.
docs/ ├── domains/ # Bounded Contexts DDD │ ├── orders/ │ │ ├── capabilities.md │ │ ├── ubiquitous-language.md │ │ └── specs/ │ ├── payments/ │ └── identity/ │ ├── specs/ # Spec Driven Dev │ ├── features/FEAT-001-checkout.md │ ├── apis/orders-api.openapi.yaml │ └── contracts/CNT-007.md │ ├── architecture/ # Decisões │ ├── adr/ADR-019-cqrs.md │ ├── diagrams/ │ └── capability-map.md │ ├── engineering/ # Padrões + DSL │ ├── patterns/ │ ├── dsl/ │ └── runbooks/ │ ├── validation/ # Testes como spec │ ├── golden-tests/ │ ├── contract-tests/ │ └── property-tests/ │ ├── ontology/ # Modelo formal │ ├── domain-model.ttl │ └── relationships.yaml │ ├── glossary/ # Linguagem ubíqua │ └── terms.yaml │ ├── agents/ # Context por agente │ ├── orders-agent.ctx.yaml │ └── payments-agent.ctx.yaml │ └── index/ # Gerado em CI ├── knowledge-graph.json └── embeddings.index
Four actual journeys of use
Out of concept: these are auditable flows with explicit human handoffs ↔ agent.
Journey 1 — Engineer writing PR
- IDE plugin recovers specs and ADRs related to file.
- Agent suggests patterns already established for the domain.
- PR check valid: cover spec, contract drift, mandatory citation.
- Self-trading ADR when architectural decision is detected.
- Merge generates event → KG indexes new evidence → updated metrics.
Day 2 — PM defining feature
- Agent recovers related capabilities in capability map.
- Suggests ADRs and historical constraints that limit scope.
- Generates spec-stub on DSL with verifiable requirements (Given/When/Then).
- Approval gate: architect + tech lead validate impact graph.
- Spec versioned becomes contract consumed by downstream agents.
Day 3 — SRE in incident
- Symptoms in telemetry → query in KG returns runbooks + ADRs of the component.
- Agent correlates with recent changes (reverse lineage).
- Suggests already validated mitigations; if new, it requires explicit application.
- Postmortem self-traced with linked evidence (logs, strokes, ADR).
- Lessons learned saw Pattern or Risk KG.
Day 4 — Auditor / Compliance
- Question: "evidence of which X control is implemented in Y".
- Lineage navigates: regulatory requirement → spec → ADR → code → test → deploy.
- Each node with timestamp, author, approval and version.
- Answer always cited; no citation, no statement.
- Export auditable (PDF + signed JSON-LD).
End-to-end traceability chain
Each chain artifact carries stable ID, version, author, timestamp and quote to the top source. Audit is no longer archeology.
Capabilities that the graph enables
Impact Analysis
"Dado ADR-019, which specs/contracts/tests will be affected?" — auto-response via traversal.
Conflict Detection
Contradictory policies detected by analysis of conflicting edges in the graph.
GraphRAG Retrieval
Agent navigates the graph to recover structured context, not Chunks isolated by similarity.
Full Auditability
Any line of code traceable to the business policy that originated it.
Contextual Onboarding
"What do I need to know to work in Orders?" — automatic structured response.
Tests as executable knowledge
A test that fails is a Spec violated, traceable to the policy that originated it. Three complementary modalities.
Contract Tests
They find that implementation honors the formal contract. Executed to each PR; they block in violation.
// refs: CNT-007, POL-042 describe('POST /orders/cancel', () => { it('requires auth (POL-042)', ...); it('returns 403 for anon', ...); });
Golden Tests
Behavior snapshots approved by the team. They detect semantic regressions, not just syntactic.
// golden: FEAT-042-checkout expect(result).toMatchGolden( 'checkout-express-v2.1' );
Property-Based
Random cases generated from formal constraints. It guarantees invariants in extreme cases.
// constraint: CON-003 property('p99 < 200ms', () => { fc.assert(latency < 200); });
Agent context file — pre-computered, versioned, governed
Each agent specialized in bounded context receives pre-computered context in build-time. Clean window context, predictable cost, deterministic generation.
# agents/orders-agent.ctx.yaml — v3.2.0 agent: orders-specialist version: 3.2.0 domain: orders owner: "@platform-cognition" context_sources: facts: - domains/orders/capabilities.md - glossary/orders-terms.yaml policies: - POL-042 # cancel-auth - POL-055 # refund-window - POL-071 # fraud-detection contracts: - CNT-007 # orders-api - CNT-012 # payments-api decisions: - ADR-019 # cqrs pattern - ADR-031 # event-sourcing constraints: - CON-003 # latency p99 - CON-008 # data-residency patterns: - engineering/patterns/saga.md - engineering/patterns/outbox.md memory: episodic: pr-context semantic: kg.orders.v3 procedural: dsl.orders retention: 90d retrieval: strategy: hybrid graph_depth: 2 semantic_threshold: 0.82 max_tokens: 8192 precedence: [authority, recency, proximity] min_groundedness: 0.85 capabilities: - spec-generation - code-review - contract-validation - impact-analysis quality_gates: - groundedness - faithfulness - citation_coverage - safety on_failure: escalate_to_human observability: trace: otel metrics: [latency_p95, groundedness_score, escalation_rate, cost_usd]
Why specialized agents?
Domain focus
Order Agent does not carry Billing — clean window context, lower cost, higher accuracy.
Pre-computer
Retrieval in build-time. Faster and more deterministic response.
Governance as API
Each agent has own ownership, SemVer version and audit trail.
Independent development
Context evolves with the domain — without coupling between agents.
Agent catalogue
Federated governance model
Inspired by Data Mesh: core central + domain ownership — without falling to the extreme of the monolithic date lake or the chaos of the pure bottom-up.
| Paper | Scope | Liability | Cadence | SLA |
|---|---|---|---|---|
| Knowledge Council | Org-wide | Approval of changes in core ontology, retrieval policy, DSL patterns. | Monthly + ad hoc | Decision in ≤ 10 working days |
| Domain Owner | Knowledge Domain | Curated specs, ADRs, domain ontology. Quality gates inside. | Weekly | Freshness < 72 h |
| Knowledge Steward | Subdomain / team | Triage of contributions, deprecation, evidence review. | Daily (asynchronous) | Triage < 24 h |
| Platform Team | L1–L5 (substrate → Reasoning) | Platform operation, observability, cognitive SLOs. | Continuous | P95 retrieval < 800ms |
| Contribution | Any | PR de spec, ADR, evidence. Under template + semantic lint. | Just-in-time | — |
| Auditor | Read-only cross-domain | Lineage check, evidence, regulatory compliance. | Quarterly | Export < 1h |
Knowledge Lifecycle
- • Not required (MAJOR.MINOR.PATCH).
- • Deprecation with minimum 90-day window and replacement pointer.
- • Specs are still navigable.
Federation policy
- • Core ontology Shared — Changes Require Knowledge Council.
- • Domain extensions — autonomous under the curatorship of the Domain Owner.
- • Semantic Testing Contract between domains in IC.
- • Controlled aliases avoid duplicity (e.g. "client"
- • Federated query with policy enforcement by domain.
Four memories, not one.
The original proposal treated memory as a unique bucket. Here we separate by cognitive function, with different retention and retrieval policies.
What happened?
Talk, agent runs, punctual decisions.
TTL: 30–180 days.
Usage: Short context, debugging.
What's true
Canon entities, ontology, versioned facts.
TTL: permanent (versioned).
Usage: grounding, main retrieval.
How to do
Workflows, runbooks, agent DSLs, patterns.
TTL: by SemVer.
Usage: execution of tasks.
Where are we?
Tenant, user, project, environment, permissions.
TTL: session.
Usage: scope + authorization.
Retrieval policy declarative
No heuristics hidden in the code. Every decision to "recover" is declarative, versioned and auditable.
precedence: - authority # ADR aprovado > spec draft - recency # menor decay rate por domínio - proximity # entidade no contexto atual - citation_count # quantas vezes referenciado conflict_resolution: strategy: explicit on_conflict: surface_both_with_provenance never: silent_majority_wins
Drift contextual detector
Detects when the recovered context begins to diverge from the actual state of the system. Prevents answers based on obsolete memory.
- • Schema drift: change in canonical entities triggers re-index.
- • Authority drift: ADR superseded → Invalidation of cached answers.
- • Behavioral drift: Drop of groundedness in successive runs.
- • Time drift: Clams with expired freshness are marked as "stale evidence".
22 5-layer cognitive KPIs
SLOs per domain + explicit FinOps layer. Dashboard "Knowledge Vital Signs" plugged into default warning (Datadog / Grafana).
| Layer | KPI | Definition | SLO suggested |
|---|---|---|---|
| Substrate | Knowledge Coverage | % capacity with at least 1 approved spec. | > 90% |
| Source Freshness | Average age of last update of ingested sources. | < 24 h | |
| Orphan Rate | % of unrequited specs or attached ADR. | < 5% | |
| Ingestion Lag | Latency between event and availability at KG. | P95 < 5min | |
| Retrieval | Recall@k | % of relevant docs in recovered top-k. | > 0.85 |
| Precision@k | % of top-k that are really relevant. | > 0.75 | |
| MRR | Mean Reciprocal Rank — where is the first relevant. | > 0.7 | |
| Retrieval Latency P95 | Retrieval layer response time. | < 800ms | |
| Citation Reachability | % of citations pointing to live source. | > 99% | |
| Reasoning | Groundedness | % of the response supported by cited evidence. | > 0.9 |
| Faithfulness | Answer does not introduce facts outside the recovered context. | > 0.95 | |
| Answer Relevance | Answer the question asked. | > 0.85 | |
| Hallucination Rate | % unsupported clan response. | < 2% | |
| Escalation Rate | % of queries requiring human intervention. | 10–25% (healthy) | |
| Outcome | Time-to-Decision | Average time of question the decision actionable. | −60% vs baseline |
| Reuse Index | How many times one spec is quoted by another. | > 3x quarter | |
| Lineage Depth | Average depth of the required graph → telemetry. | > 5 levels | |
| Audit Time | Time to generate evidence of compliance. | < 1h | |
| FinOps | Cost per Query (USD) | Average cost (LLM + retrieval) per answered question. | Monthly trend ↓ |
| Token Burn Rate | Tokens consumed by agent/day. | Within the monthly budget | |
| Hard Escalations/week | Circuit Breaker drives. | < 5% of executions | |
| Budget Usage | % of monthly budget consumed. | Alerts at 70% / 90% / 100% |
Semantic tip-to-point tracing
OpenTelemetry instrumenting agents, retrieval and LLMs. Each span loads:
- Trace. id — tip-to-point of the chery
- retrieval.policy — version of the policy applied
- Grounding. score — Reasoning confidentiality
- quotes[] — IDs and versions of sources
- model. id + prompt.hash
- cost.tokens + cost.usd
- decision. gate — what quality gates have passed
- Human. in.loop — scaling if applicable
Adoption in 4 incremental phases
Each phase has objective exit criteria. No milestone without measurable evidence.
Phase 1 — Foundation
0–90 days Quick winsEstablish substrate, ontology base and first read-only agents.
Delivery
- • Integrated Docusaurus + OpenKB
- • Core ontology v1 (10–15 canonical entities)
- • Ingestion of existing specs and ADRs
- • GraphRAG read-only for 1 pilot domain
- • 5 KPIs cognitive baseline
Output criteria
- ✓ Knowledge Coverage > 60% no domínio piloto
- ✓ Recall@10 > 0.75 em benchmark interno
- ✓ Knowledge Council operando
- NPS internal > 40 between engineers
Phase 2 — Pilot Cognitive
3–6 monthsFull cognitive loop, quality gates and first automated actions with human-in-the-loop.
Delivery
- • Agents with loop SENSE→... →LEARN complete
- • Quality gates (groundedness, faithfulness)
- • PR check with assisted contract drift review (HITL)
- • Tracing complete OTel
- • Dashboard "Knowledge Vital Signs"
Output criteria
- ✓ Groundedness > 0.85 em produção
- ♪ Hallucination rate < 5% ♪
- ✓ Time-to-Decision −30% vs baseline
- At least 1 incident solved with lineage
Phase 3 — Federate
6–12 monthsKnowledge Mesh: domain extensions, semantic testing contract, scaling to 3+ domains.
Delivery
- • Federated query with policy enforcement
- • Contract tests between IC domains
- • Domain Owners + Stewards established
- • Domain onboarding self-service
Output criteria
- 3+ areas in production without central coordination
- ✓ Reuse Index > 3x
- ✓ Audit Time < 4h
- ✓ Drift entre domínios < 1% por trimestre
Phase 4 — Autonomous
12–24 monthsAgents with delegated authority (write actions), self-healing, learning loops closed — operating within configured Circuit Breakers.
Delivery
- • Write agents (authorial PR, self-published ADR)
- • Self-healing in retrieval and Reasoning
- • Closed-loop learning with user feedback
- • Internal marketplace of agents/skills
Output criteria
- ✓ Time-to-Decision −60% vs baseline
- ✓ > 30% de PRs com contribuição cognitiva
- Past regulatory audit with lineage
- Knowledge operating as a product
Circuit Breaker active at this stage
Self-employed agents operate with 3 attempts. On Wednesday, the scale system for human review and records the FinOps audit event. Without budget configured, this phase is not enabled. Details in Security & FinOps.
Anti-standards to avoid
- Big bang ontological — design the perfect ontology before any delivery.
- Agent without human-in-the-loop in phase 1–2.
- RAG without mandatory citation — becomes a generator of plausibility.
- "Let's add governance later" — Ontological output is exponential.
- Late centralisation — when five areas have already diverged, mergers are expensive.
Non-negotiable principles
- No claim without quotation.
- Specs are versioned, caution is explicit.
- Every agent action is traceable end-to-end.
- Quality gates fail closed, not open.
- Domain owns its semantics, platform owns its substrate.
What changes when knowledge is a operating system
The original version was a stack. This version is a platform: has users, SLOs, contracts, auditable governance and roadmap. More Important — It Treats Knowledge how internal product, not as a by-product of delivery.
From documentation
for versable, citable, executable specs.
From search
for retrieval governed, with declarative policy and lineage.
From Chatbot
for agents with cognitive loop, quality gates and tracing.
From the homeless AI
for Cognitive Circuit Breakers with FinOps integrated.
Punitive governance
for Reuse Index that rewards document well.
Of the technical jargon
for Value anchors — every technical layer explained for the board.
"Knowledge that it is not executable turns folklore. Folklore governing decisions is strategic debt."
Next step
This specification is ready for a 90-minute workshop with architecture, platform and pilot domain. Output: Priority backlog for Phase 1 — Foundation.