Papers
Working papers
Architecture notes and preliminary experiments, with protocols, observations, and stated limitations. Shorter material appears under Notes.
- WP-01
The crew as an owned artifact
Most agent systems treat orchestration as either a library problem or a chat problem. Lab Z treats the unit of value as a crew: a versioned team of agents with explicit roles, tools, and output contracts. This working paper states the thesis, the failure modes it avoids, and the empirical claims we intend to test.
- WP-02
Timed domain crew: Business Coach in about seven minutes
In a single local session on 20 August 2026, authoring a Business Coach crew through first useful answer took about seven minutes with Zero already running. This working paper records the protocol, timings, limits, and reproduction path. It is a workflow and latency experiment, not a scored coaching-quality study.
- WP-03
How well do local models build apps? Two 30B-class models, one spec
Two locally-hosted models — qwen3-coder:30b-a3b and gpt-oss:20b — each received an identical, deliberately underspecified specification for a small full-stack application over messy operational data, and built it unattended on a single 12 GB GPU. We graded the results as you would grade an engineer's take-home: design quality, code quality, and whether the delivered application runs. gpt-oss produced the far stronger submission on paper — TypeScript, SQLite, clean separation, a design document with a reason for every choice — and it did not run, requiring eleven repairs without ever serving its primary endpoint. qwen3-coder chose a deprecated toolchain and hedged its own design document, and its application worked after a one-line fix, returning filter results that match ground truth exactly. We report what each model can and cannot be trusted with, where their applications break, and why the answer to the title question is: genuinely useful, provided you treat the output as a first draft whose seams you will personally verify.
- WP-04
How well do local models do data science? A classifier take-home, graded automatically
We gave the same data-science take-home to two locally-hosted models — build a spam classifier from a messy 208,000-row corpus — five times each, unattended, on a 12 GB consumer GPU, and graded every submission automatically against a 50,000-row holdout. The task carries a planted temporal-leakage trap: a column that predicts the label almost perfectly in training and is always empty in production. As data scientists, the models are more capable than their reputation suggests: nine of nine design documents chose F1 over accuracy on the imbalanced data, eight of ten exploration scripts executed against the corpus, and one model identified and correctly excluded the leaked column in five of five runs, with sound reasoning stated in writing. As engineers, they fail at the last step: seven of ten submissions crashed before producing a single prediction, on undeclared imports, unpicklable classes, and a module path that moved in 2018. The three that ran scored 0.767, 0.766 and 0.510 against a careful-reference score of 0.801 — and the two best scores came from runs that planned to use the leaked column and were saved only by failing to implement their own design.
- WP-05
How well do local models write numerical code? A physics simulation graded against exact solutions
Third trial in the series: the same two local models (qwen3-coder:30b-a3b, gpt-oss:20b, both on a 12 GB consumer GPU) each ran a numerical-computing take-home five times, unattended — build a 1D heat-conduction simulator, output temperatures at requested positions and times. Grading is fully objective: every reported temperature is compared against the exact Fourier-series solution of the PDE, with one hidden config using piecewise-varying diffusivity so the textbook formula cannot substitute for a real solver. A pre-registered calibration ladder (careful solver 6/6, naive 4/6, lazy 0/6, formula-only 1/6) anchors the scale. The result is the cleanest separation of the series: gpt-oss scored 5/6 in all five runs, failing only the varying-diffusivity trap and for the classic reason — pointwise diffusivity instead of flux-conserving interface treatment, its error matching our naive reference almost exactly. qwen3-coder scored 0/6 in all five runs, not through bad physics but through a violated output contract it was told about twice and never fixed. The mid-run repair loop lifted three gpt-oss runs from 0/3 to 3/3 on the self-check — its first unambiguous wins — while repairing nothing for qwen3-coder in five attempts.