Research annex series
Research Annex · Ontology Pipeline
Research diary from Hadto's ontology pipeline: what the discovery gates, benchmarks, and review loops are teaching us.
This series is source work, not the reading path. If you run a business, start here — or browse the rest of the research annex.
- What the ontology research pipeline is teaching us
Our ontology research loop is already showing a useful pattern: home services looks like the fastest packaging opportunity, professional services looks like the deeper differentiation play, and dental remains the best place to keep learning.
- Research velocity is rising, but the gates are doing what we expected
Discovery-backed competency questions are still climbing, but maintenance proposals are being rejected by answer-path quality gates—exactly the kind of friction that protects the platform from scaling with weakly grounded ontology edits.
- When the same escalation pressure appears in four businesses, it is a platform signal
The newest research cycle did more than add competency questions. It showed the same escalation pattern recurring across dental, home services, professional services, and franchise operations—a strong sign that Hadto should look for a reusable cross-venture operating primitive.
- 100% ontology coverage is not the finish line
A fully green ontology dashboard means we answered the questions already in scope. It does not mean we have finished learning what future owner-operators need from the platform.
- A research pipeline has to keep its evidence attached
A research loop becomes an asset only when every claimed discovery can be traced back to source material and review.
- A business has not learned until the playbook changed
Discovery is only intake. Real learning happens when evidence survives review and changes the shared method the next operator inherits.
- Research should hand sales a packet, not a pile of findings
A research loop starts to matter commercially when it can turn model pressure into a ranked offer, a buyer hypothesis, a discovery script, and a proof surface another operator can run.
- Differentiation only counts when the queue is visible
A premium service line stops depending on private judgment when the business names where work is blocked, who owns the review, and what requires manual follow-up.
- A repair only counts when it creates a reviewable decision
The useful proof after a repair is not that the machinery ran. It is that the next operator can see the decision, the remaining risk, and the path to judgment.
- A learning system has to let the score get worse
A green dashboard can protect pride. A teachable business needs something harder: a system that can admit a new gap, keep the evidence visible, and show the next operator where judgment belongs.
- Silence is not a clean signal
A quiet queue can mean safety, failure, rejection, or blindness. Future owner/operators need systems that preserve the reason for absence.
- Autoresearch for ontologies needs a field crew
Karpathy's autoresearch loop is useful because it keeps score. Ontology work needs its own score: less uncertainty without breaking the commitments the business already runs on.
- Benchmark the ontology against the business
The first OntoMoBench proof is a shipped benchmark contract. The next proof is a deterministic scorer that checks syntax, runtime behavior, transition conformance, and invariants against business traces.
- The ontology learned when the proof got better
OntoGPT did not teach us to let a model invent the ontology. It taught us to bind suggestions to schema, source text, competency questions, and scores that catch unsupported facts.
- An agent brain still needs an ontology overlay
Hadto's internal proof is a documented overlay contract, JSON schema, deterministic scorer, and fixture-backed tests for three operator surfaces.