Register
A falsifiable question, outcome, population, split, and failure criteria before results.
THE FOURTH DOWN STANDARD
Confidence is
an artifact.
Our agents can propose, analyze, challenge, and edit. They cannot quietly move the goalposts or publish around a failed check.
→
Provenance
Source, retrieval time, license, attribution, schema, and SHA-256 recorded.
Analyze
Deterministic code, frozen seed, temporal validation, baseline, uncertainty, artifacts.
Attack
Statistician and adversarial reviewer look for leakage, selection, and overclaiming.
Verify
Every number traces to an artifact; every source and license claim is checked.
Gate
Mechanical checks plus an editor decision. A blocker stops publication.
NON-NEGOTIABLES
What has to be true
- No time travel. Forecasting studies use rolling or held-out future periods.
- Baselines first. A model earns complexity by beating the simplest relevant forecast.
- Uncertainty travels with the result. Point estimates do not publish alone.
- Limitations stay adjacent. Caveats are not buried after the conclusion.
- Evidence is addressable. Claims map to tables, charts, code, and immutable hashes.
- Agents leave receipts. Prompts, schemas, findings, decisions, and models are recorded.
PUBLICATION GATE v1.0
A dramatic headline does not get a vote.
Seven mechanical checks must pass: complete provenance, temporal validation, minimum sample, reported uncertainty, zero blocking findings, disclosed limitations, and source attribution.
✓Data provenance
✓Temporal validation
✓Sample threshold
✓Uncertainty
✓Adversarial review
✓Limitations
✓Citation trace
THE RESEARCH DESK
Specialized roles, shared evidence
Researcher
Audits the run against a code-locked registration.
Statistician
Audits model and validation logic.
Adversary
Tries to break the strongest claim.
Citation checker
Traces numbers, sources, and rights.
Editor / publisher
Narrows claims and enforces the gate.
Orchestrator
Selects due series and records state.
MODEL DISCLOSURE / CURRENT
GLM-5.3 is on the desk. It is not the calculator.
Online language-model roles use z-ai/glm-5.3 through OpenRouter. Statistics, tables, charts, thresholds, and gate checks remain deterministic code.
- LANGUAGE MODEL
- GLM-5.3
- STATISTICS
- Deterministic Python
- AGENT OUTPUT
- Strict JSON Schema
- PUBLISHED RUNS
- Frozen deterministic launch review
Language matters.
We call these specialist-agent reviews. They are not independent human peer review. If qualified external reviewers join a study, we will name that layer and disclose its scope separately.
Why we draw the line →Corrections stay public.
Material errors freeze promotion or receive a dated article note. Original runs remain addressable, and prepublication holds enter the same public ledger.
Open correction policy →