Methods & integrity

The method has to survive the result.

These are the rules HNI applies before calling an experiment, replication, benchmark, or technical report part of its research record.

Study lifecycle

Before and after execution.

Before execution

Lock the question and the ways it could fail.

Research question and rival or null explanation
Primary outcomes and units
Sample and stopping rules
Exclusions and analysis plan
Ethics, privacy, funding, conflicts, and planned AI roles

After execution

Publish the full primary record.

All primary outcomes
Null, negative, mixed, and inconclusive findings
Failed runs and deviations
Verification evidence and its internal or external level
Actual AI involvement and material limitations

Fourteen commitments

Research integrity policy.

Question before conclusion. Every study names a rival or null explanation.
Label the evidence level. Exploratory, confirmatory, replication, benchmark, and commentary are distinct.
Lock before looking when claiming confirmation.
Publish inconvenient results.
Report every primary outcome.
Disclose deviations, exclusions, and failed runs.
Separate measurement from verification. Internal automated checks are labeled as internal, never external review.
Make evidence checkable through code, data, receipts, hashes, or justified redaction.
Disclose actual AI involvement.
Disclose funding, in-kind credits, consulting, and conflicts.
Protect people and private data. The local harness does not substitute for ethics review.
Correct visibly through dated versions.
Do not confuse repository identifiers with peer review.
No prestige theater. HNI claims only the status and review it actually has.

AI use

Tools are disclosed. Humans remain accountable.

AI may assist with implementation, analysis, source discovery, or drafting. It may not fabricate evidence, invent citations, suppress contradictory findings, or replace accountable authorship.

Human research

A higher gate applies.

Human-participant work is blocked without documented external approval or exemption. Personal-data work requires a data-management plan.

Corrections

Earlier versions do not disappear.

Material corrections create a dated version and ledger entry. Withdrawal remains visible unless safety, privacy, or legal duties require removal.

Protocol schema Result schema