Skip to content

Cog Arch Dgm

Discovering Gödel Machine (DGM) cognitive architecture component.

Cluster: Experience & Autonomy | Type: component | MCP Tools: 42

Overview

Discovering Godel Machine (DGM) cognitive architecture that implements self-improving program search: programs can be defined, mutated, crossed over, benchmarked, safety-checked, evolved through a lineage-tracked archive, and inspected through deterministic backend-native capability discovery. Delegates 37 operations to an inner DGMMCPBlock backed by SQLite, storing programs, evaluation results, proof attempts, and evolution history.

Launch caveat: Treat this as an advanced or experimental value-add, not a first-10-minutes onboarding feature for vibe coders. It is functional and useful for users who understand benchmarks, archive search, and verification loops, but most end users should encounter it through simple workflow templates or guided examples rather than raw DGM concepts. Do not position it as autonomous production self-improvement without domain-specific evaluation and review.

Safety caveat: Public evolution can run without caller-supplied invariants, but those runs are qualified-draft and heuristic/benchmark-only. Treat verified safety as requiring supplied invariants and proof evidence.

When to use:

  • Building self-improving agents that propose and verify code rewrites before execution
  • Maintaining a lineage of evolving program versions with evidence-backed improvement checks
  • Integrating self-modification into the G6 pipeline under explicit safety constraints

Example:

from mvp.cog_arch_dgm import CogArchDGMBlock, DGMInput

block = CogArchDGMBlock(name="dgm")
defined = block.infer(DGMInput(
    op="program_define",
    program_json='{"source_repr":"def policy(x): return x.get(\\"a\\", 0) + x.get(\\"b\\", 0)"}',
))
result = block.infer(DGMInput(op="program_analyze", program_id=defined.value.programs[0]["id"]))
# result.ok -> True; result.value -> DGMOutput with records and diagnostics

Works well with: cegis, formal_methods, recursive_architect

Public API

CogArchDGMBlock(AIBlock[DGMInput, DGMOutput, dict])

Top-level DGM block — delegates to DGMMCPBlock.

Field Type Default
name str 'cog_arch_dgm'
state dict \| None None
db_path str ':memory:'
resource_bounds ResourceBounds \| None _DEFAULT_RESOURCE_BOUNDS
usage ResourceUsage field(default_factory=ResourceUsage)

Methods:

infer(data: DGMInput) -> Result[DGMOutput]

DGMInput(BaseModel)

Input for CogArchDGMBlock.

Field Type Default
op str required
program_id str ''
program_json str '{}'
benchmark_id str ''
config_json str '{}'
query str ''

DGMOutput(BaseModel)

Output from CogArchDGMBlock.

Field Type Default
op str required
ok bool required
programs list[dict] Field(default_factory=list)
evaluations list[dict] Field(default_factory=list)
proofs list[dict] Field(default_factory=list)
message str ''
error str ''
data_json str ''
degraded bool False
degradation_reason str ''
confidence float 0.0
next_steps list[str] Field(default_factory=list)

Functions

compute_weighted_fitness(foundational: float, intermediate: float, advanced: float, weights: tuple[float, float, float] = (0.5, 0.3, 0.2)) -> float

Aggregate per-stage benchmark scores into a single fitness value.

MCP Tools

Operation Source
archive_create cog_arch_dgm_mcp
archive_get cog_arch_dgm_mcp
archive_list cog_arch_dgm_mcp
archive_branch cog_arch_dgm_mcp
archive_prune cog_arch_dgm_mcp
program_define cog_arch_dgm_mcp
program_mutate cog_arch_dgm_mcp
program_crossover cog_arch_dgm_mcp
program_analyze cog_arch_dgm_mcp
benchmark_define cog_arch_dgm_mcp
benchmark_run cog_arch_dgm_mcp
benchmark_stage cog_arch_dgm_mcp
benchmark_compare cog_arch_dgm_mcp
verify_safety cog_arch_dgm_mcp
check_invariant cog_arch_dgm_mcp
prove_improvement cog_arch_dgm_mcp
proof_status cog_arch_dgm_mcp
evolve_step cog_arch_dgm_mcp
evolve_run cog_arch_dgm_mcp
evolve_status cog_arch_dgm_mcp
evolve_history cog_arch_dgm_mcp
find_analogy cog_arch_dgm_mcp
realtime_bridge cog_arch_dgm_mcp
dgm_search cog_arch_dgm_mcp
dgm_info cog_arch_dgm_mcp
list_patterns cog_arch_dgm_mcp
list_capabilities cog_arch_dgm_mcp
describe_engine cog_arch_dgm_mcp
describe_mutation_backends cog_arch_dgm_mcp
describe_evaluation_policy cog_arch_dgm_mcp
describe_sandbox_policy cog_arch_dgm_mcp
probe_formal_backends cog_arch_dgm_mcp
describe_proof_status cog_arch_dgm_mcp
describe_llm_proposal_backends cog_arch_dgm_mcp
describe_store_schema cog_arch_dgm_mcp
describe_archive_policy cog_arch_dgm_mcp
describe_pattern_mechanisms cog_arch_dgm_mcp
proved cog_arch_dgm_mcp
disproved cog_arch_dgm_mcp
unknown cog_arch_dgm_mcp
solver_unavailable cog_arch_dgm_mcp
fallback_only cog_arch_dgm_mcp