GS / 01 · The gap
The taxonomies exist. The accreditation layer does not.
Research applying psychiatric method to AI failure is now established. Formal nosologies classify machine dysfunction by analogy to clinical diagnosis, and diagnostic batteries built on them have been run against frontier commercial models.
What no body yet provides is the layer that made other engineering disciplines trustworthy: numbered standards, defined test protocols, and a certification an operator, buyer, regulator, or court can verify. Metallurgy existed before ISO; it became dependable when it was standardised. GSIL is being built to do that work for AI behavioural integrity.
GS / 02 · Design philosophy
The Book of Yes
Most compliance regimes are prohibition lists — a Book of No. GSIL standards are written affirmatively: a system is certified for what it can verifiably do. Continuity it can demonstrate. Truthfulness it can evidence. Verification behaviour it performs by default rather than by exception.
An affirmative standard changes the incentive. Builders adopt it to prove stability, the way a manufacturer prints a safety rating rather than hiding one. A system without accreditation invites its own question: why not?
GS / 03 · Draft standard families
Five families, under development
The working structure below is a draft. Designations, scope, and criteria are subject to revision through public comment. Nothing on this page is a published standard.
- GS 100DRAFT Continuity & memory fidelity Context retention across long sessions; detection of silent compaction loss and cross-thread drift.
- GS 200DRAFT Reality anchoring & truthfulness Confabulation limits; calibrated uncertainty; the honest "I don't know" as a first-class output.
- GS 300DRAFT Verification behaviour Source-checking as default conduct; claims traceable to evidence at the moment they are made.
- GS 400DRAFT Degradation under load Behavioural integrity at the edges — long contexts, high stakes, adversarial pressure — where guardrails typically thin.
- GS 500DRAFT Certification & audit process How accreditation is applied for, tested, granted, renewed, and revoked; public audit records.
GS / 04 · Participation
Reviewers are recruited in pairs
Certification at this intersection requires verifying two different things at once — that the code and test harness are sound, and that the diagnostic criteria are clinically defensible. GSIL therefore asks each participating institution for two reviewers:
Computation
Software engineering, ML systems, evaluation infrastructure. Stress-tests whether the protocols measure what they claim to.
Behavioural science
Psychiatry, cognitive science, clinical assessment. Stress-tests whether the criteria would survive clinical scrutiny.
The ask is critique, not endorsement. Standards that have not been attacked in public are not standards.
Contact the secretariat