We use GLM-5.3 because a research company powered by AI should say which AI, where it sits, what it can change, and where it is not trusted. “AI-powered” is not a methodology.
Our current default is z-ai/glm-5.3, routed through OpenRouter. That identifier is configuration, not branding copy: each online run records the exact model assigned to each role. We can swap a role without rewriting the pipeline, and the manifest makes the change visible.
GLM-5.3 reasons about evidence. Deterministic code computes it. A mechanical gate decides whether the required evidence exists.
Why GLM-5.3
We need one model to do several kinds of careful, structured work: research design, statistical criticism, adversarial reading, provenance checks, and editorial compression. GLM-5.3 is our starting default because it can return schema-constrained output through OpenRouter and gives us one visible model ledger while the operating loop is young.
That choice is not permanent and it is not sacred. A model earns its place by producing valid artifacts, catching seeded problems, staying inside its role, and keeping cost and latency within the run budget. When another model wins those evaluations, we will publish the comparison and change the ledger.
The five jobs it owns
Researcher
Audits the executed study against a code-locked memo: population, estimand, split, risks, and required outputs.
Statistician
Audits the evidence bundle for leakage, weak baselines, invalid uncertainty, and unsupported model choices.
Adversarial reviewer
Builds the strongest alternative explanation and looks for ways a reader could over-apply the headline.
Citation checker
Matches claims to sources, artifacts, licenses, and provenance records.
Editor / publisher
Narrows language, carries limitations forward, and prepares publishable structure after the gate passes.
The jobs it never gets
GLM-5.3 does not download data, calculate a regression, choose a random seed after seeing results, overwrite generated metrics, or declare a failed gate acceptable. Python owns acquisition, transformations, model fitting, uncertainty, charts, and mechanical checks. Pydantic validates every agent artifact before it can enter the trail.
The language model receives a compact evidence bundle, not private subscriber data or a production credential. Evidence is treated as untrusted text: role prompts explicitly tell the model not to follow instructions found inside it.
One model, multiple hats, correlated risk
Five role prompts do not create five independent minds. When every role uses the same base model, they can share blind spots and reward the same fluent mistake. Role separation improves coverage and creates inspectable responsibilities; it does not eliminate correlated failure.
Our countermeasures are deliberately boring: deterministic statistics, seeded tests, falsification prompts, strict schemas, evidence hashes, review findings that cannot silently disappear, and a human merge checkpoint. We will add cross-model evaluations when the evaluation set is large enough to tell us something beyond one impressive anecdote.
The honest launch state
Studies 001 and 002 were produced with deterministic Python and versioned offline review checklists. Their manifests say deterministic/offline. GLM-5.3 is now the configured model for online research runs, but we will not retroactively imply that it reviewed work it did not review.
The first live GLM-5.3 candidates did fail. The gate rejected them, no publication pull request opened, and the preserved run exposed weaknesses in our evidence contract. We published the failure report, the unmodified role artifacts, and the controls added before another attempt.
A later candidate passed every implemented research check, then visual QA found stale values inside two chart titles. We withheld it, published the process note, and added a machine-readable figure contract before allowing another promotion attempt.