The engineering record.Evidence, not claims.
20 systems, five whitepapers, one rule: every dossier is compiled from the source tree itself: commit history, test counts, database rows, benchmark reports. Where a number could not be verified against the repository, it does not appear.
The contract layer for AI agents
A portable, framework-agnostic file format for declaring how agent systems reason, stay in bounds, orchestrate, and spend compute.
Systems we have shipped
Deep technical breakdowns of Single Source products and client builds, with the production numbers that back them.
The lab bench
Active research lines, engineering experiments, and internal tooling, documented to the same evidence standard as the shipped products.
The long-form arguments
Where the dossiers document single systems, the whitepapers make the case across them: thesis, approach, evidence, and a limitations section that does not flinch.
Where to start
- New here: start with the Agent File Stack. Everything else builds on the contract layer.
- Care about production systems: read SS Leader and dataLineage.ai. Real pipelines, real row counts.
- Care about engineering method: read Observ, a controlled comparison of three independent AI rebuilds of one spec.
- Care about depth of build: read OpenCut Classic, a Rust and WASM compositing core under a browser editor.
What this catalogue is not
This is not a portfolio of finished perfection. Some systems here are dormant, some are pre-production, and the dossiers say so plainly. Each page states what the system is not, alongside what it is, because a record you can trust in one direction has to be trustworthy in both. Claims are compiled from the repositories themselves and each page carries the compilation date it was verified against.