Tip: select any text on this page to bring up the Ask tooltip and question the Synedrix OS assistant about it.
Why we keep a playground around
We can't resist a promising new tool. The moment something looks like it might solve a problem well, we want to run it, not just read the README. Most of what we try doesn't end up in the stack, and that's fine; that's what trying things is for. What wasn't fine was having nowhere real to do the trying: it either turned into an accidental production dependency nobody signed off on, or a laptop experiment nobody else ever saw. Synedrix OS keeps a real, visible home for that curiosity instead, and this page is it.
Playing, without breaking anything
Everything here runs boxed off from the real thing: its own network, its own datastores, no path into production, no production data anywhere near it. Nothing a lab produces becomes a memory record, an audited action, or a write anyone downstream relies on. That's what lets us actually poke at something instead of only theorising about it.
Before we start, we write down what would actually convince us either way, the same idea as a pre-registered study, so a good result can't quietly become “well, we always liked it anyway” after the fact. Every lab has an expiry date and exactly two ways to end: promoted into the real stack through a proper decision record, or retired. A clear “no” closes a question just as well as a “yes” opens one. Either way, we stop guessing.
What we've actually found so far
Both labs opened August 2026 and run through February 2027. Here's what running them for real has turned up so far: the findings that complicate the original pitch, right alongside the ones that support it.
LAB-001 — EverOS
Local-first agent-memory runtime. Its lifecycle model (episode, consolidation, reflection, skill distillation) is already canonical; its runtime is what's on trial.
- A real, reproducible upstream bug broke 100% of agent-memory calls as shipped. We found it, diagnosed it, and fixed it with a documented dependency override.
- A structural limitation, not a config gap: its vector tables hardcode a dimension its own settings claim to control. Confirmed by reading the schema files directly.
- Once both were addressed, the real memorize-to-extract cycle worked end to end.
LAB-002 — Semantica
A GraphRAG and knowledge-graph toolkit, evaluated one capability at a time rather than as a single adoption question.
- Its structural parser extracted a real 14×7 table perfectly, and fragmented a different, clean six-row table into three pieces. Recorded as a real limit.
- A retrieval-quality harness, run against the real knowledge base on both chunkers: recall@10 on multi-hop questions went from 0.25 to 0.75, on four questions per category: a first signal that still needs a larger, repeated run. Simple lookups showed no difference either way.
- Its decision-graph export produces genuine, standards-compliant PROV-O output: real
prov:Entity/prov:Activitytriples, not a private vocabulary dressed up as one.
What's in the playground today
| Component | Role | Evaluated under |
|---|---|---|
| LanceDB | Embedded vector index (EverOS) | LAB-001 |
| EverOS | Local-first agent-memory runtime | LAB-001 |
| Semantica | GraphRAG / knowledge-graph toolkit | LAB-002 |
| Docling | Structural document parsing | LAB-002 |
Real third-party software under time-boxed evaluation in the governed lab lane (ADR-OS-004), in structural isolation from every production system. None of it counts among the twelve platforms, belongs to the tech stack, or can be cited as authority by a canonical document. Each charter’s only two outcomes at expiry are promotion through a normal ADR, or retirement.
Synedrix OS is in active development. Everything above actually runs on the same DGX infrastructure as the rest of this site. None of it is a mockup or a someday plan.