Testing ideas before they become products.

Exploring

Can a generation pipeline preserve world state across 50,000+ words of narrative without drift — where every state change requires extracted text evidence before becoming canon?

Why it matters

Longform fiction and script work often breaks at continuity rather than creativity. A shared state record and checks before each change help keep a mistaken generated fact out of the story.

Current status

Research is ongoing. We have completed end-to-end runs for episodes 1 and 2 from the same series. Draft completion and release readiness are tracked separately.

Exploring

Does routing between quantized runtimes by workload class beat any single 'best' runtime for agent work on consumer GPUs — given that raw decode TPS does not predict agent throughput?

Why it matters

A 27B model at interactive speed on one 32GB card expands what self-hosted agents can do. The measurements so far show that no single runtime wins every workload, making routing the central engineering question.

Current status

We have measured three runtimes and recorded their sources and test conditions. A production ranking still requires comparable end-to-end timings across multiple seeds.

Exploring

How far can a read-only parser take operational forensics — parsing journal formats to byte offset with hash-verified fixtures, without loading product JARs or starting the system under investigation?

Why it matters

A diagnostic tool should leave the system under review untouched. This project tests that boundary with a read-only design and repeatable fixtures.

Current status

v0.4.0 Message Trace is being tested as a release candidate. We have checked fixtures generated by three ActiveMQ versions.