Add a document. Nothing else moves.
A document is read into TOMX's knowledge through a gate that decides, per fact, whether to write it, skip it as already known, or flag it as contradicting what is already held. No retraining: the knowledge lives on disk, so adding to it is a write.
Every entity this deployment serves already holds its document, so there is no blank entity to teach. What you can do is withhold one entity's document for your session, watch its answers degrade, then paste the document back and watch them return. That is a real write through the real gate, and it is how the research run itself was measured. It is not a document about something the corpus has never seen, and this panel does not claim that.
Grow once. Rerun the identical old world.
First hash and run the server-owned probe500 manifest. Then add one safe, previously absent fact, answer its new query, and rerun that exact manifest. The checkpoint identity is read before and after; no-retraining passes only when it stays identical while overlay rows grow.
Clear this session overlay; run and hash the fixed probe.
Write the fixed safe document and answer the new fact.
Rerun the same probe and show every old-world delta.
business
Q4830453| Relation | What the model answers now | Margin | Graph says |
|---|---|---|---|
| P31- | United States | 0.0858 | AB Volvo |
| P279- | Mean Streets | 0.0139 | dot-com company |
| P2578- | Říčany | 0.1241 | business administration |
| P361- | Raphaël Personnaz | 0.0092 | factory |
| P460- | United States | 0.0858 | company |
| P2360- | Mean Streets | 0.0139 | Architecture Firm Award |
| P921- | Mean Streets | 0.0139 | Dragons' Den |
| P460 | United States | 0.0858 | company |
Paste the document
Facts are recognized only between entities this corpus already holds, joined by a relation it already knows. This is not an open-domain extractor. When it finds nothing it reports that instead of inventing a fact.
Does adding break what was there?
Each point is a real bench run over the probe set under your session's state: relational-hop accuracy, gate off. If growth damaged what was already held, these bars would fall. Each number is measured fresh when you press the button.
no measurements yet; take a baseline before you grow, then again after