An original, executable-design attempt to build generalized machine cognition using independent internal learning systems linked by a machine corpus callosum.
Julia cognitive runtime · native Ada/SPARK action authorization · persistent opposing models · continuously coupled numerical activity · independent prediction evaluation.
State of evidence: real source code is supplied, but the new native Julia/Ada suite was not executable in this authoring environment. Its new CI jobs must pass before describing this initial release as verified. True AGI has not been demonstrated.
The package integrates the actual IX-BlackFox-AGI Julia source and its compiled Ada authorization boundary, not a mocked substitute. A second independently learning subsystem models observed effects including effects on external persistent objects which the original BlackFox local-effect learner refuses to learn. A delayed, signed, sparse numerical interface feeds internal activity across the two systems. Training feedback changes both route gains and weight matrices based on actual predictive error. It never passes natural-language messages or hidden world rules between the systems.
The runtime can ACT, INVESTIGATE or HOLD. An unauthorized proposal cannot dispatch through the standard path. The second hemisphere cannot authorize itself. Checkpoints persist and are SHA-256 verified.
Julia 1.10+; GNAT and GPRbuild for a compatible native Ada compiler; x64 Windows or Linux. No paid model APIs, no pretrained language model, and no Python runtime in the cognitive system. Python is not required to run IX-AGI.
gprbuild -p -P vendor/IXBlackFoxAGI/ada/ix_guard.gpr
julia --project=. test/runtests.jl
julia --project=. bin/ixagi.jl
julia --project=. experiments/comparison.jlOr use tools/run-checks.ps1 on Windows, tools/run-checks.sh on Linux.
The demo uses a host-controlled task that includes an effect involving an object outside the action argument. BlackFox's original learner rejects the unsupported effect, while the independent learner can acquire it after repeated observations. A held-out object then tests transfer. The callosal mode transports numerical predicted-state-change activations, never the right hemisphere's proposed symbolic future state. A late-fusion control deliberately receives those symbolic predictions to test whether the callosal pathway is necessary. The exact same proposal is denied with permission=0 and can execute with permission=1 through native Ada; dispatches are counted. A severed-channel control checks the loss of cross-hemisphere influence. The matched comparison deliberately includes a reversed rule to expose failure; it does not constitute AGI evidence.
The standalone numerical projection calibration passed 1,000 independently seeded sparse-channel checks, but it is a Python-standard-library mathematical sanity check, not a Julia integration test or evidence of AGI. The native suite remains pending.
The project keeps the donor's original evaluation-only license under vendor/IXBlackFoxAGI/LICENSE. Top-level LICENSE specifies evaluation-only access. See NOTICE.md and docs/STATUS.md for exact provenance and verification limits.
- The original IX-Sally full source is not included; its cognitive routines have not been executed or merged. See
docs/SALLY_INTEGRATION.md. - The current learner has bounded symbolic perception; it does not exhibit unconstrained concept invention or broad lifelong generality.
- No complete correctness proof of Ada/SPARK and no AGI claim.
- Internal signs, delays and numerical routing are computational analogues, not proof that human brains function identically.
AI proposes. Humans decide.