The lab explains
itself.

We publish the models, prompts, boundaries, review patterns, failures, and business gates behind the research. The process is not a trade secret. It is part of the evidence.

GLM-5.3

We route our language-model roles through OpenRouter using z-ai/glm-5.3. GLM-5.3 checks conformance, challenges claims, traces evidence, and edits research. It does not write the locked registration, calculate the statistics, or waive the publication gate.

Read the complete model card →
ROUTER
OpenRouter
MODEL
z-ai/glm-5.3
STATISTICS
Deterministic Python
OUTPUT CONTRACT
Strict JSON Schema
PUBLISHED RUN MODE
Frozen deterministic launch review
HONEST CURRENT STATE

The public Touchdown Regression article remains the deterministic/offline launch snapshot. A newer revision passed all five GLM-5.3 roles, then visual QA found stale chart headlines and withheld it. No agent review is described as independent human peer review. Rejected candidates, the passing-but-withheld bundle, and post-run annotations remain public. Open the machine-readable run index ↗

How the machine
actually works.

Specific design decisions, written while they are still decisions—not polished into mythology later.

Every change leaves a diff.

  • Model identity. Provider, router, slug, role, and run mode belong in the ledger.
  • Prompt identity. Exact role prompts and output schemas are versioned and published.
  • Failure identity. Blocked runs, warnings, and corrections stay visible.
  • Human identity. We will never relabel agent review as independent human peer review.