82a358f18d
ontoref-derive: #[onto_mcp_tool] attribute macro registers MCP tool unit-structs in
the catalog at link time via inventory::submit!; annotated item is emitted unchanged,
ToolBase/AsyncTool impls stay on the struct. All 34 tools migrated from manual wiring
(net +5: ontoref_list_projects, ontoref_search, ontoref_describe,
ontoref_list_ontology_extensions, ontoref_get_ontology_extension).
validate modes (ADR-018): reads level_hierarchy from workflow.ncl and checks every
.ncl mode for level declared, strategy declared, delegate chain coherent, compose
extends valid. mode resolve <id> shows which hierarchy level handles a mode and why.
--self-test generates synthetic fixtures in a temp dir for CI smoke-testing.
validate run-cargo: two-step Cargo.toml resolution — workspace layout first
(crates/<check.crate>/Cargo.toml), single-crate fallback by package name or repo
basename. Lets the same ADR constraint shape apply to workspace and single-crate repos.
ontology/schemas/manifest.ncl: registry_topology_type contract — multi-registry
coordination, push targets, participant scopes, per-namespace capability.
reflection/requirements/base.ncl: oras ≥1.2.0, cosign ≥2.0.0, sops ≥3.9.0, age
≥1.1.0, restic declared as Hard/Soft requirements with version_min, check_cmd, and
install_hint (ADR-017 toolchain surface).
ADR-019: per-file recipient routing for tenant isolation without multi-vault. Schema
additions: sops.recipient_groups + sops.recipient_rules in ontoref-project.ncl.
secrets-bootstrap generates .sops.yaml from project.ncl in declarative mode. Three
new secrets-audit checks: recipient-routing-coherent, recipient-routing-coverage,
no-multi-vault. Adoption templates: single-team/, multi-tenant/, agent-first/.
Integration templates: domain-producer/, mode-producer/, mode-consumer/.
UI: project_picker surfaces registry badge (⟳ participant) and vault badge
(⛁ vault_id · N, green=declarative / amber=legacy) per project card. Expanded panel
adds collapsible Registry section with namespace, endpoint, and push/pull capability.
manage.html gains Runtime Services card — MCP and GraphQL toggleable without restart
via HTMX POST /ui/manage/services/{service}/toggle.
describe.nu: capabilities JSON includes registry_topology and vault_state per project.
sync.nu: drift check extended to detect //! absence on newly registered crates.
qa.ncl: six entries — credential-vault-best-practice (layered data-flow diagram),
credential-vault-templates (paths A/B/C), credential-vault-troubleshooting (15 named
errors), integration-what-and-why (ADR-042 OCI federation), integration-how-to-implement,
integration-troubleshooting.
on+re: core.ncl + manifest.ncl updated to reflect OCI, MCP, and mode-hierarchy nodes.
Deleted stale presentation assets (2026-02 slides + voice notes).
5.3 KiB
Extractions — Steal This Deck (Talisman, KGC 2026)
Source: Jessica Talisman, "Stop Betting, Start Building", Knowledge Graph Conference 2026.
Newsletter: Intentional Arrangement (Substack). Received: 9 May 2026.
Full deck: Steal_This_Deck.pdf
Framing / Intro hooks
Para abrir cualquier presentación de ontoref:
"AI is a knowledge tool. Not a data tool. That breaks every assumption underneath many AI strategies."
Adaptado a ontoref: los agentes AI que trabajan sobre tu repositorio operan sobre correlación estadística si no hay infraestructura de conocimiento. Ontoref es esa infraestructura, pero para proyectos de software.
"Agentic AI is not a model upgrade away. It is a shared language, from which to act."
Este cierre del deck es el mejor hook de intro para ontoref: un modelo más grande no sabe qué es tu proyecto, qué decisiones tomaste, ni en qué estado está. Eso requiere conocimiento estructurado, no un token window mayor.
Datos citable para el problema
Todos con fuente verificable — útiles en posts y slides sin necesitar justificación propia:
| Dato | Fuente |
|---|---|
| 89% de empresas reportan cero impacto de productividad de AI en 3 años | NBER, Feb 2026, n=5.937 execs |
| Developers experimentados fueron 19% más lentos con herramientas AI | METR RCT, 2025 |
| Ganancia neta semanal del trabajador AI-promedio: −14 minutos | Foxit/Sapio, March 2026 |
| 1.7× más issues en PRs escritos por AI | CodeRabbit, 2025 |
| 46% del código en GitHub escrito por Copilot | Octoverse 2025 |
El argumento que emerge: AI is not saving time. It is generating volume. Without a knowledge backbone, its only measurable output is noise.
El argumento central — para slides y posts
El stack de conocimiento (Talisman lo llama "knowledge infrastructure"):
1. Controlled vocabularies — términos con significado único y autoritativo
2. Taxonomies — organización jerárquica
3. Thesauri — equivalencias, relaciones cruzadas
4. Ontologies — compromisos formales: clases, propiedades, constraints
5. Knowledge graphs — live, queryable, governable
"You cannot skip layers. You cannot start at five and reverse-engineer to one."
Ontoref implementa este stack para proyectos de software:
- Schemas NCL → vocabularios controlados
.ontology/core.ncl→ nodos Practice/Concept con edges tipados- DAG-formalized knowledge → el knowledge graph
- ADRs + migrations +
state.ncl→ governance
El argumento de precisión — para posts técnicos
"The accuracy gap is not closed by a bigger model. It is closed by a defined schema, an ontology, and a validated query."
Con datos:
- 16% → 72% en question-answering sobre SQL enterprise con ontology checks (Allemang & Sequeda, data.world AI Lab, 2024)
- 3.4× GraphRAG vs vector RAG en 43 queries enterprise (Diffbot KG-LM Benchmark, 2023)
- Vector RAG colapsa a 0% past 5 entities per query. KG-grounded retrieval sostiene.
"Defining your terms is the cheapest accuracy and cost lever in the LLM stack."
Ontoref es exactamente esto: ontoref describe y el sistema Q&A (ADR-003) son el vocabulario controlado que reduce el espacio de alucinación cuando un agente trabaja sobre el proyecto.
El argumento de MCP — crítico para el posicionamiento técnico
"MCP and A2A are transport. Not semantics. They move bytes between endpoints. They do not establish shared meaning."
"Two agents connected by MCP exchange text bytes. They do not truly share context."
Ontoref cierra este gap: tiene superficie MCP (ontoref-daemon/src/mcp/) encima de una capa ontológica. El protocolo mueve bytes; ontoref provee el significado que hace que esos bytes sean conocimiento, no texto.
Esto es diferenciación directa respecto a "simplemente exponer tu repo por MCP".
El argumento cultural — para posts de opinión
"The industry is optimized to ship solutions, not to own problems."
| Celebrado | Huérfano |
|---|---|
| Launching new platforms | Maintaining existing platforms |
| AI-generated content | Taxonomy work |
| Ontologies built on the fly | Domain expertise |
"Knowledge work IS the maintenance. Yet it keeps getting cut."
Ontoref formaliza exactamente lo que siempre se corta: las decisiones arquitectónicas (ADRs), el estado del proyecto (state.ncl), la memoria operacional (.coder/). Lo hace queryable y machine-readable para que no dependa de que alguien lo recuerde.
El close — para cualquier formato
"The organizations that close the perception/reality gap first won't be the ones with the best models. They'll be the ones who finally did the work."
Adaptado: los proyectos donde los agentes AI trabajan mejor no son los que tienen el modelo más grande. Son los que tienen conocimiento estructurado sobre sí mismos.
Atribución
Jessica Talisman, MLS — Semantic Engineer, Information Architect, Knowledge Infrastructure Strategist. Newsletter: Intentional Arrangement (Substack). Talk: "Stop Betting, Start Building", Knowledge Graph Conference 2026, Technology Track, May 6, 2026.
Al usar cualquiera de estos puntos en público, citar la fuente — es un argumento de autoridad que refuerza, no debilita, la posición de ontoref.