Compatibility note: the canonical executable truth-loop contract is now
COGNITIVE_TRANSACTION_LOOP_MODEL.md. The cortex agi command and the paths
under md-os/ops/agi/ remain compatibility surfaces during migration; they do
not define a separate AGI layer.
MD-OS does not define AGI as consciousness, identity, unrestricted autonomy, or an internal change to model weights.
Within MD-OS, AGI-like progress is defined as verified cross-domain task acquisition:
the system can solve new tasks, detect failure, learn procedures, validate
them, promote reusable skills, and improve future performance without
corrupting epistemic state or operational safety
The Semantic Operational Compiler is the substrate. The Verified AGI Loop is the learning cycle that uses that substrate.
The bounded loop is now grounded by a typed cognitive transaction:
TaskSpec with executable acceptance tests
-> bounded connector action
-> ActionReceipt with observed state delta
-> independent postcondition verifier
-> formal episode
-> failure analysis
-> skill candidate
-> eval
-> promotion gate
-> runtime compiler rebuild
The preferred first implementation surface is a single replayable transaction:
cortex cognition run-once --task-spec md-os/ops/tasks/<task_spec_id>.json
It is not a daemon and it does not start continuous autonomous operation.
A plain task string may still be recorded through the compatibility surface,
but without executable acceptance tests its verdict is unverified.
The kernel is separated into explicit roles:
context_compiler
task_controller
planner_search
tool_executor
verifier
episode_memory
failure_analyzer
skill_distiller
eval_runner
promotion_gate
risk_controller
runtime_compiler
The planner proposes. The executor acts through bounded tools or filesystem state. The verifier judges. The learner writes episodes, failures, candidates, eval results, and promotion readback.
Every run must leave a formal episode under:
md-os/ops/episodes/
An episode records:
task
task_type
context_pack_id
plan
actions
observations
errors
artifacts
verifier_results
verdict
lessons
candidate_claim_updates
candidate_skills
regressions
Without an episode, a run is only a log. With an episode, MD-OS can compare similar tasks, detect failure patterns, distill procedures, evaluate them, and promote reusable skills.
Reusable operational intelligence accumulates as skills under:
md-os/ops/skills/candidates/
md-os/ops/skills/promoted/
md-os/ops/skills/skill_registry.json
A promoted skill must be:
structured
source-bound
executable by bounded tools
testable
versionable
evaluated
revocable
rollback-aware
Promotion is not allowed just because a procedure sounds plausible.
The hard rule is:
no promotion without eval
no eval without artifact
no artifact without readback
no readback without deterministic manifest
The gate requires:
schema_valid
source_bound
verifier_passed
eval_passed
no_regression
risk_reviewed
rollback_available
The gate blocks when:
imported_unverified_high_impact_claim
unsafe_tool_required
missing_readback
semantic_contradiction_open
eval_contamination_detected
The planner must not reason in a void. The world model is generated under:
md-os/ops/world/world_model.json
It indexes operational entities such as workspace, compiler, connectors, tools, constraints, risks, readback surfaces, and linked capabilities.
Every learning rebuild emits:
md-os/ops/evals/agi_eval_report.json
Metrics include:
task success rate
mean steps to success
failure recovery rate
autonomy horizon
semantic drift
claim contradictions
skill reuse
regressions
cost
The first bounded horizon is a single safe cycle. Longer horizons should be introduced only after the single-cycle gate is reliable.
The generic episode-to-skill loop does not by itself prove that learning
improves future performance. The bounded accelerator in
NEUROMORPHIC_LEARNING_ACCELERATOR_MODEL.md adds a causal before/after protocol:
same provider without skill on sealed holdouts
-> independently verified development episodes
-> competitive hypothesis elimination
-> sparse parameterized skill
-> distinct validation case
-> same provider with skill on the same sealed holdout cohort
The primary metric is:
change in verified holdout success / verified source episodes
Promotion additionally requires equal holdout attempt budgets, no evaluation case in induction memory, no oracle access by the provider, no regression, and an external verifier. Failed candidate hypotheses are retained as prediction errors and information-gain events; they are not mislabeled as successful task episodes or runtime failures.
The canonical bounded command is:
cortex agi accelerate --experiment-id <append_only_id>
Its report must keep agi_achieved and agi_claim_supported false. The
experiment can support a narrow learning-transfer claim only.
This model does not claim:
consciousness
unbounded autonomy
hidden self-modification
continuous self-running agency
parametric training of the base model
It defines the operating path for external behavioral and procedural learning: episodes, verifier readback, skill distillation, evals, promotion gates, and runtime recompilation.