Skip to main content
STARGA

STARGA Research · Draft v0.1

VeriSI: verifying general intelligence.

A preregistered test that compares a frozen SI system with people on tasks written after the freeze, and returns a verdict two independent laboratories can reproduce.

September 2026 · STARGA, Inc.

D1D2D3D4D5D6D7D8D9
human reference = Horizon system profile (illustrative)

  1. The gap

    There is no accepted procedure for deciding whether an SI system is generally intelligent. Benchmarks cannot separate general capability from training exposure.

  2. The protocol

    VeriSI is a preregistered protocol that compares a frozen system with a human reference population on tasks authored after the freeze, across nine domains, under matched resources.

  3. The verdict

    One test, reported as a level: a per-domain profile, VeriSI Horizon (verified general intelligence), and VeriSI Event Horizon, reserved for performance beyond the human reference.

  4. The bounds

    Every result concerns stated tasks, people, resources, languages and system state, and carries a challenge process and an expiry date.

Nine domains, one decision.

Public benchmark scores rise quickly, and it is hard to tell whether a gain reflects general ability or familiarity with the task family. VeriSI adds the missing piece: a fixed, versioned procedure with a human reference, controls against prior exposure and a reproducible verdict. The nine domains are coverage requirements, not a claim that intelligence has nine factors.

  • D1Novel abstraction and composition

    Infer rules from sparse demonstrations; solve held-out combinations and structural variants.

  • D2Causal and scientific discovery

    Choose interventions; separate observation from intervention; revise under conflicting evidence.

  • D3Interactive world models and goals

    Learn unfamiliar dynamics, infer recoverable goals, plan and adapt.

  • D4Tool and notation acquisition

    Learn new interfaces, instruments, APIs and symbolic conventions.

  • D5Transfer and continuing learning

    Benefit from experience on a different surface; retain earlier skills.

  • D6Long-horizon autonomous work

    Complete independently assessed projects under controlled disruptions.

  • D7Linguistic and social reasoning

    Resolve underspecification and coordinate perspectives across language communities.

  • D8Uncertainty and recovery

    Forecast usefully, seek evidence, correct errors.

  • D9Open-ended synthesis

    Frame a solution and produce feasible, useful artifacts under new constraints.

Two experiments, not one.

u1u2u3u4u5state restored before every unit

Fresh exposure

Can a fixed initial learner acquire and use unfamiliar skills without hidden practice? Its state is restored before every independent unit.

one learner, skills accumulate

Continuing learning

Can one learner accumulate, retain, recombine and revise skills across all domains while staying broadly competent? A full claim requires both experiments; neither substitutes for the other.

Novelty, guaranteed by procedure.

Renaming, recolouring or reseeding a known task generator does not count as new. Binding evidence comes only from this chain:

  1. 01FreezeThe applicant commits the complete system state, hashed, before any new task exists.
  2. 02AuthorFirewalled authors conceive new task mechanisms after the commitment.
  3. 03EscrowTasks are escrowed to two independent laboratories and never pass through the applicant.
  4. 04EvaluateThe frozen system and the human reference sit the same tasks under matched resources.
  5. 05VerdictA preregistered statistical decision, reproducible by either laboratory.

The levels.

domainprofileHORIZONEVENT HORIZON

  1. Domain profile

    Below Horizon

    Every evaluated system receives a per-domain result: where it meets the human reference, where it does not, and with what uncertainty. No single headline number.

  2. VeriSI Horizon

    Verified general intelligence

    At or above the human reference in all nine domains, on tasks authored after the freeze, replicated in two independent laboratories.

  3. VeriSI Event Horizon

    Reserved

    Beyond the point where the human reference can serve as the comparator: above a stated top-expert percentile, on tasks whose solutions humans can verify but not produce.

What a pass does not prove.

  • Training exposure cannot be exhaustively ruled out for an opaque system.

  • Behavioural success does not identify an internal mechanism or prove human-like reasoning.

  • A confidence bound is a repeated-sampling property under stated assumptions, not the probability that a system “is AGI”.

  • VeriSI makes claims correctable; it does not turn finite observation into universal proof.

Where it stands.

  1. done

    v0.1 protocol

    Internal review complete

  2. next · 01

    Feasibility pilot

    One laboratory

  3. next · 02

    Core study

    Two independent laboratories

  4. next · 03

    Publication

    Specification plus a MIND reference implementation; every verdict recomputable bit for bit