Research Preview
ReconCoda Lab Console
A private, single-GPU workbench that turns model choice from a chat-demo opinion into a reproducible decision: fixed datasets, exact model digests, corpus snapshots, telemetry, held-out checks, and one queue controlling the card.
Who it is for — and why
Teams choosing a local model or inference configuration for sensitive work and needing to show exactly why one candidate was promoted over another.
What was verified
- The final node environment passes 269 tests with 24 expected skips, including the sqlite-vec paths that local-only checks skip
- A deterministic findings engine compares only like-for-like runs, uses paired exact McNemar tests with Benjamini–Hochberg false-discovery correction, and reserves REPLICATED for development plus untouched-holdout evidence
- The current breadth suite carries 168 cases; immutable dataset and corpus hashes prevent a historical run from silently using today's evidence
- A real API → queue → GPU → artifact → telemetry → unload canary completed cleanly, and a two-job contention canary proved that only one workload owns the GPU lease
Limits
- The repository and raw model outputs are private; the public artifact is the evidence-led case study
- The workbench runs on one consumer GPU, so its scheduling and throughput evidence does not establish multi-GPU or production-service behavior
- Several results use operator-authored gold and held-out sets, which test consistency within one authoring method rather than independent external validity
- This platform selects and compares models and runtime variants; the separate weight-intervention work is documented in its own project entry
Access and evidence
Related evidence →Source remains private; the public-safe evidence is linked above.