"The specific frustrating moment when you realize your knowledge is
leaking away, or siloed, or dying with you, or being reinvented
by the next person who walks in the door."
That moment is what Cognitome is for.
02 / 07
What Cognitome does
The Problem
Knowledge workers accumulate deep domain understanding through conversation,
research, and experience. That knowledge lives in their heads, in scattered
chat histories, and in documents nobody can search effectively.
Keyword search finds words. It doesn't find meaning.
Similarity graphs find proximity. They don't find structure.
Neither captures what the knowledge actually is.
The Solution
Cognitome ingests a corpus — Claude conversation exports, PDFs, documents,
links — and constructs a formal OWL 2 ontology: a structured
knowledge base that captures not just concepts, but the
relationships between them.
The result is shareable, reusable domain knowledge that software agents
and humans can reason over — grounded in your own words, traceable to
its source.
03 / 07
The mechanism: Hypothesis → Antithesis → Synthesis
🧠
Hypothesis
claude-sonnet-4-6
Similarity graph computed from corpus. Proposer extracts concepts,
relationships, and hierarchies. Anchors each claim to a verbatim
source quote.
→
⚔️
Antithesis
kimi-k2.5 (free)
Deliberately chosen from a different training lineage than the Hypothesis model.
A model from the same training lineage critiquing its own proposal shares blind spots —
correlated errors that neither catches. Kimi K2.5 (Moonshot AI) was
independently trained on different data, different RLHF, different
architectural choices. Its failure modes don't overlap.
True antithesis requires genuine independence.
→
⚗️
Synthesis
claude-sonnet-4-6
Synthesizer integrates antithesis, enriches the ontology, resolves
conflicts. Human reviews and edits result in Protegé —
human edits within Protegé always win.
OWL 2 DL output, round-trip validated.
Key principle — The human remains the epistemic driver throughout. The pipeline surfaces structure; the human validates meaning. This is not autonomous ontology generation — it is augmented epistemology.
04 / 07
Sample output: hemodialysis device corpus
Extracted Concepts
UseErrorFailure in perception, cognition, or action during device operation
RiskControlDesign or procedural measure reducing identified hazardous situation
AlarmFatigueClinician desensitization from excessive alarm frequency
DialysateErrorSubclass of UseError — wrong concentrate, temp, or contamination
ResidualRiskRisk remaining after ALARP controls; accepted on therapeutic benefit
ThreatModelSTRIDE-based analysis of cybersecurity attack surfaces
TraceabilityMatrixFDA-required evidence chain: use error → control → validation
Extracted Relationships
UseErrortriggersHazardousSituation
HazardousSituationmitigated-byRiskControl
RiskControlvalidated-bySummativeEvaluation
SummativeEvaluationevidencesTraceabilityMatrix
AlarmFatigueis-subtype-ofUseError
AlarmFatigueaddressed-bySmartAlarmAlgorithm
ThreatModelrequiresSBOM
SBOMenablesVulnerabilityTracking
representative sample — generated from dialysis corpus
05 / 08
It's not hallucination. It's a wrong destination.
"An LLM doesn't lie. It navigates probability space and
lands somewhere that felt geometrically right
— but that coordinate happens to be factually empty in the real world.
It landed in the suburbs of Truth instead of the city centre."
The actual failure mode — not lying, not confusion. Wrong destination.
Without anchor quotes
The pipeline proposes UseError → causes → AlarmFatigue.
It's structurally plausible. The model is confident.
The relationship propagates through every downstream inference.
But the direction is reversed. AlarmFatigue is a subtype of UseError —
not caused by it. A wrong destination, silently compounding.
With the semantic checkpoint
Every relationship surfaces its anchor quote — the verbatim sentence
from the source corpus where the claim originated.
The human asks one question:
"Does this relationship actually follow from what I said here?"
That question is the only reliable catch for a plausible but wrong destination.
Not theater. A checkpoint.
06 / 08
Why ontology. Why not just RAG.
vs keyword search
🕸️
Structure, not proximity
Keyword matching finds co-occurrence. Similarity graphs find
semantic proximity. Neither captures that UseError triggers
HazardousSituation which is mitigated-by RiskControl.
That chain is causal knowledge. Only an ontology holds it.
vs RAG / vector DB
🔍
Reasoning, not retrieval
RAG retrieves relevant chunks. OWL 2 DL enables inference —
you can ask questions the corpus never explicitly answered,
because the ontology encodes the domain's logical structure.
Agents can traverse it. Humans can audit it.
vs LLM summarization
📎
Provenance, not hallucination
Every concept and relationship is anchored to a verbatim
source quote from the original corpus. Nothing is invented.
The human can trace any assertion back to the conversation
where it was established. Auditable by design.
07 / 08
Current state & roadmap
✓ Validated & Working
Three-model pipeline — Hypothesis/Antithesis/Synthesis end-to-end, validated on 400+ concepts
MCP server — Live SQLite backend, queryable by Claude and other agents
Multi-corpus extraction — Biology, legal, medical, social science corpora tested
Provenance anchoring — Every assertion traceable to source conversation
→ In Development
Document ingestion — PDF, DOCX, TXT as pseudo-conversation objects. Corpus-to-ontology engine, not just chat-to-ontology.
Web scraping — Ingest external links directly into the corpus pipeline
North star — Ontology-grounded Q&A. The knowledge graph becomes an agent that answers questions the corpus never explicitly addressed.
Built by
Solo founder. Former Gulf Oil (seismic Fortran), Cray Research (vectorization, XMP/YMP), FICO. 40+ years of domain knowledge leaking away — then building the tool to stop it. Hypothesis → Antithesis → Synthesis.
The Ask.
Cognitome belongs in the AI ecosystem. Not as a competitor
to anything being built — as a demonstration of what LLM
conversations can become when the knowledge in them is made
explicit, structured, and reusable.
01
A technical conversation about whether Cognitome's MCP integration and ontology pipeline belongs as a showcase in the Claude ecosystem
02
Feedback on the architecture from people who think about agentic knowledge management every day
03
A path to users — the latent demand for corpus-to-ontology tooling is real. The medical device usability manager. The solo researcher. The enterprise knowledge manager whose institutional memory walks out the door every time someone retires.
cognitome · built on claude-sonnet-4-6
The knowledge is already in the room.
Cognitome makes it visible.
If this belongs in your ecosystem —
point me toward the right conversation.
I'm not looking for a job. I'm looking for a collaborator.