AGI→ASI Implementation Guide¶

What was built from the "From AGI to ASI" gap analysis (arXiv:2606.12683), where it lives, and how to use it. The pieces compose into one loop — see Self-Improving Reasoning Substrate for the architecture. The gap analysis itself is in reports/agi-to-asi-gap-analysis-2026-06-13.md.

What it is¶

The paper frames the AGI→ASI transition as four pathways — scaling, paradigm shifts, recursive self-improvement, multi-agent collectives — gated by frictions, with the recurring rule: for every friction, build a countermeasure or a way to measure whether it binds. These concepts make AU measure its own dynamics, reason in plural paradigms, bound cost, and stay corrigible — the substrate for moving along that continuum.

Concept reference¶

Concept	Module	What it does	How to use
KG-2.68	`knowledge_graph/core/reasoner.py`	Outcome-learning paradigm router (keystone)	`kg.reason(ReasoningTask(goal, tags, payload))`
KG-2.69	`harness/program_synthesis.py`	Inductive program synthesis + MDL/Occam prior	`synthesize(primitives, examples)`; `select_top_k(..., method="mdl")`
KG-2.67	`knowledge_graph/core/world_model.py`	Action-conditioned world model	`WorldModel().observe(...); wm.rollout(start, policy, horizon)`
AHE-3.22	`knowledge_graph/research/code_synthesis.py`	Autonomous single-file code generation in the evolution loop	`governed_publish(...)` (default-on, sandbox + ActionPolicy gated)
AHE-3.23 / 3.24	`knowledge_graph/research/capability_ratchet.py`	Verified apply→verify→rollback + monotone capability ratchet	runs post-publish in `governed_publish`; consulted by promotion governance
AHE-3.26 / SAFE-1.3	`knowledge_graph/research/improvement_ledger.py`	RSI velocity ledger (improving/stalling)	`ImprovementLedger(engine).summarize()`; auto-recorded each golden-loop cycle
SAFE-1.1	`harness/frontier_scorers.py`	Non-saturating progress: compression, Elo, saturation detector	`CompressionScorer` (in reliability suite); `saturation_detector(pass_rates)`
SAFE-1.5	`core/corrigibility.py`	Corrigibility + irreversibility aversion + knowledge-seeking reward	wired into the goal loop; `ACTION_IRREVERSIBILITY_AVERSION=1` for the policy gate
OS-5.35	`orchestration/cost_governor.py`	Throughput-per-dollar scale-up cap	`FLEET_SCALE_BUDGET_USD_PER_HOUR` (opt-in; unset = unchanged)

Using the reasoning router¶

from agent_utilities.knowledge_graph.facade import KnowledgeGraph
from agent_utilities.knowledge_graph.core.reasoner import ReasoningTask

kg = KnowledgeGraph()

# Inductive: learn the shortest program fitting examples (routes to KG-2.69)
kg.reason(ReasoningTask(
    goal="learn the mapping", tags=("induction",),
    payload={"primitives": {"double": lambda x: x*2}, "examples": [(1, 2), (3, 6)]},
))

# Symbolic: forward-chain to a goal fact (routes to the deductive paradigm)
kg.reason(ReasoningTask(
    goal="derive C", tags=("deduction",),
    payload={"facts": ["A"], "rules": [(("A",), "B"), (("B",), "C")], "goal_fact": "C"},
))

Each call routes to the paradigm whose learned reward EMA + capability tags best fit, runs it, and feeds the result's score back — so the router self-improves. Register a new paradigm with get_reasoner_router().register(my_reasoner); it needs only a name, capability_tags, and reason(task) -> ReasoningResult(score).

Safety & cost envelope¶

Corrigibility (SAFE-1.5): autonomous goal loops checkpoint and yield to a supervisor shutdown signal without resisting; set ACTION_IRREVERSIBILITY_AVERSION=1 to route irreversible actions (delete/destroy/merge/deploy) to a human even under an auto tier.
Cost (OS-5.35): set FLEET_SCALE_BUDGET_USD_PER_HOUR to cap autoscaler scale-ups at a throughput-per-dollar budget; every scaling action carries an est_cost_usd_per_hour.
Capability ratchet (AHE-3.24): a published evolution branch that measurably regresses capability is abandoned; the verdict feeds promotion governance.

Measuring the loop¶

ImprovementLedger(engine).summarize() reads the loop's own audit streams (EvolutionCycle, ProposalPublication, CapabilityRatchetResult) into a velocity reading — cycle cadence, the genotypic-vs-prose mechanism split, capability pass-rate, and an improving/stalling verdict whose signals flag the paper's "research-gets-harder" mode. It is recorded automatically each golden-loop cycle as an ImprovementVelocity node.

Status & roadmap¶

Implemented + merged (local main): KG-2.67/2.68/2.69, AHE-3.22/3.23/3.24/3.26, SAFE-1.1/1.3/1.5, OS-5.35. Extending the same loop next: OS-5.34/AHE-3.25 (distil winning reasoning traces into training data, model-collapse-guarded by SAFE-1.4) and ORCH-1.46/47/48 (the router at population scale — market allocation, emergent specialists, hierarchical coordination). Tracked in reports/agi-to-asi-gap-analysis-2026-06-13.md.