cognify.

Applied R&D · AI · Data · Analytics

Logbook

Open record of experiments, notes, and exchanges with IA models.

  1. Log 008Oct 26, 2026
    • Tessera
    • inteligencia colectiva
    • multiagente
    • memoria colectiva
    • investigación
    Tessera: Artificial Collective IntelligenceRecord of the first multi-agent working prototype: what was tested, why roles are not enough, the six initial artifacts, and a hypothesis regarding tacit collective memory.
  2. Log 007Sep 29, 2026
    • Routing
    • Qwen
    • Ollama
    • Experimento
    Skill Availability and Routing in Qwen 3.5 0.8BWe evaluated Qwen 3.5 0.8B as a routing mechanism on the same benchmark as Router v0.1. With thinking=false it reached 46% operational accuracy versus the specialized router's 83%, but outperforms the baseline in CALCULAR. The result introduces the distinction between possessing a competence and recognizing when to use it.
  3. Log 006Sep 29, 2026
    • Ensayo
    • Routing
    • Skills
    • Agentes
    • Baseline
    Skill RoutingRouter v0.1 is the first experimental baseline for a line of work on skill routing in AI agents. A TF-IDF classifier with cosine similarity over centroids learns to route queries to five known capabilities. On 100 external queries, it achieves 83% accuracy, with heterogeneous performance across routes and a marked lexical limitation in CALCULATE.
  4. Log 005Sep 25, 2026
    • Exploración
    • Prompt injection
    • LLMasHCI
    • Qwen
    • Alineación
    LLM as HCI, execution and breakdown via prompt injectionThe use of a language model as a linguistic HCI interface instead of an agent is explored. A prompt injection attack breaks the role and reveals that capacity restriction cannot depend on model obedience.
  5. Log 004Sep 25, 2026
    • Exploration
    • Qwen
    • Wikipedia
    • Skills
    • Alineación
    Qwen 3.5 0.8B + Wikipedia: Incorporating External Knowledge AccessWe explore whether Qwen 3.5 0.8B can incorporate a Wikipedia access skill and autonomously combine it with other system capabilities. The composition works, but it is fragile: premature termination, incorrect tool selection, and a lack of operational boundaries.
  6. Log 003Sep 24, 2026
    • Exploration
    • Qwen 3.5 0.8B
    • Skills
    • Math
    Deterministic mathematical capabilities in Qwen 3.5 0.8BWe explore the incorporation of a deterministic mathematical skill including calculator, square root, power, and factorial. Tests showed that Qwen can select and combine tools sequentially, using the result of one operation as input for the next.
  7. Log 002Sep 24, 2026
    • Essay
    • LLMs
    The limitations of LLMs as an exploration spaceLLMs present structural limitations derived from their probabilistic nature that involve relational and dependency risks, reasoning failures and overconfidence, and inconsistencies in memory and context. These limitations open up an area of research for external architectures that improve their reliability, control, and alignment with human agency.
  8. Log 001Sep 24, 2026
    • Exploration
    • Qwen 3.5 0.8B
    • Skills
    Qwen 3.5 0.8B + Deterministic SkillsA local agent was built with Qwen 3.5 0.8B that uses external tools for verifiable tasks. Qwen interprets the intent and decides whether to respond on its own or delegate to deterministic counting functions and Hunspell for spell checking. Early tests show that the model can select the correct tool, pass arguments, and respect the result, even though its general knowledge remains limited.