Home/Gothink AIR/Business ontology
Gothink AIR / Business ontology SELF-EVOLVING

An ontology that updates itself.

It learns from your real documents and questions, changes only when a change scores better than the version in force, and never without your approval.

Learns from real work Must beat the version in force You approve every change One-click rollback
run_evolve·ontology evolution·insurance claims RUNNING
evidencepaired scoreapproval
6
Kinds of editEach with its own pass test.
0
Changes without approvalAutomatic only if you switch it on.
1
Click to roll backAny earlier version, as a new one.
0
LLM keys neededDeterministic, no network.
01 / the screen

Every proposed change shows its evidence and its score.

demo.gothink.ai/ontology/ · Self-evolving
Self-evolving tab: a proposed synonym with its evidence, a scorecard comparing the version in force with the candidate, three passed checks and Approve / Reject buttons1234
1 the proposed edit and why  ·  2 version in force vs candidate  ·  3 checks it must pass  ·  4 you decide.
02 / the loop

Observe. Propose. Score. Approve. Commit.

1 · observe

Real work

Fields in AI-ready documents, amendments, Playground questions.

2 · propose

One edit per signal

With the documents and pages behind it.

3 · score

Against the version in force

Same held-out work, same budget.

4 · approve

A person decides

Re-scored at approval.

5 · commit

New version

Snapshot, diff, Witness receipt.

Answers marked Correct in the AIR Playground become the questions every change is scored on. Rejected ideas are not proposed again until their evidence doubles.
03 / when a change wins

A change ships only if it beats what is in force.

Exhibit A· two edits from the demo above, scored on held-out workk = 5 · τ = 0.01
version in force (v2)with the edit
Synonym ‘aside’question retrieval
+0.5
Synonym ‘aside’overall score
+0.25
Map ‘adjuster_name’fields the ontology names
+14 pts
No question ranks worseContext budget keptBeats the parent by τ
Numbers from the demo screen in 01. People ask about money “set aside”; the claim files say “reserve”. Adding the synonym lets the question reach the right passage.
04 / six kinds of edit

Content changes only. Each with its own test.

low risk

Synonym

A word people use that the documents never do.

low risk

Map a field

A field that is another name for a term.

medium

New term

A field no term names. Definition from the estate pack.

medium

Rule

From the estate pack; must hold on 95% of live values.

medium

Relationship

Terms that always appear together.

high

Deprecate

Only if nothing uses it. Always a person.

05 / versions

Every version kept, compared, receipted and reversible.

demo.gothink.ai/ontology/ · Versions
Versions tab: a diff from v2 to v6 listing two new terms, a new synonym and a new relationship, and a version table with scores and Witness receipt ids1234
1 what changed between any two versions  ·  2 the score each change earned  ·  3 its Witness receipt  ·  4 restore, as a new version.
06 / for people and agents

It grows without bloating the prompt.

ontology_manifestNames only83tokens for the whole demo topic — what an agent gets up front.
ontology_browseFind the termsTop matches for the step at hand — ids and labels.
ontology_resolveFetch the sliceDefinition, synonyms, rules, relationships, evidence — just for those terms.
demo.gothink.ai/playground/ · answer
AIR Playground answer showing 'Business ontology used: aside → reserve amount (claims v6)', a Marked correct label, and the ontology tools listed under MCP / RAG access123
1 the Playground says which terms it used  ·  2 mark an answer correct — it becomes a scoring question  ·  3 the same ontology over MCP.
07 / what changed

From a document you maintain to a state that sharpens as it runs.

Before · build 17

Maintained by hand
  • Suggestions from fixed source lists
  • No test that a change helps
  • A version number, no way back
  • Agents inferred meaning each time

Now · build 18

Self-evolving, under your control
  • Learns from real documents and questions
  • Must beat the version in force
  • Compare and roll back any version
  • Agents fetch only the slice they need
Next step / see it on your own documents

Bring one business topic. Watch it sharpen.

Connect your documents, confirm a few answers, and review the first round of evidence-backed changes.