GENERAL_ENVIRONMENT_AUTONOMY.md

General Environment Autonomy

This is the current, code-grounded checklist for an Aura architecture that can handle NetHack-scale complexity without becoming NetHack-specific. NetHack is a stress adapter; the reusable system is the environment OS.

Required Capabilities

The executable audit lives in core/environment/capability_matrix.py.

Capability	Implemented by
Bounded adapter contract	core/environment/adapter.py
Observation normalization	core/environment/state_compiler.py, adapter compilers
Shared ontology	core/environment/ontology.py, ParsedState
Persistent belief graph	core/environment/belief_graph.py
Canonical spatial/topological memory	EnvironmentBeliefGraph.spatial, upsert_spatial()
Resource homeostasis	core/environment/homeostasis.py
Modal state machine	core/environment/modal.py
Startup prompt policy	core/environment/startup_policy.py
Candidate generation	core/environment/policy/candidate_generator.py
Tactical simulation	core/environment/simulation.py
Action ranking	core/environment/policy/action_ranker.py
Strategic HTN planning	core/environment/strategy/htn_planner.py
Semantic command compilation	core/environment/command.py, environment compilers
Closed-loop action semantics	core/environment/action_semantics.py
Semantic action budgets	core/environment/action_budget.py via EnvironmentKernel
Effect gate	core/environment/action_gateway.py
Governance bridge	core/environment/governance_bridge.py
Receipt chain	core/environment/receipt_chain.py
Semantic outcome diff	core/environment/outcome/semantic_diff.py
Outcome attribution	core/environment/outcome_attribution.py
Cross-run outcome ledger	core/environment/outcome/ledger.py
Procedural memory connection	core/memory/procedural/store.py via EnvironmentKernel
Competence tracking	EnvironmentKernel.competence_tracker
Hindsight replay / causal rules	core/environment/experience_replay.py
Autonomous abstraction discovery	core/environment/abstraction_discovery.py
Open-ended curriculum tasks	core/environment/curriculum.py
Black-box trace + replay	core/environment/blackbox.py, core/environment/replay.py
Run lifecycle/postmortem	core/environment/run_manager.py, core/environment/postmortem.py
Benchmark mode separation	core/environment/benchmark_runner.py, BoundaryGuard
External task proof gate	core/environment/external_validation.py
Runtime activation proof bridge	core/runtime/activation_audit.py, core/runtime/proof_kernel_bridge.py
Async/concurrency health	core/runtime/concurrency_health.py, core/utils/task_tracker.py, core/resilience/lock_watchdog.py, core/tasks/dead_letter_queue.py
Corrective grounding	core/brain/grounding_guard.py, core/evaluation/self_evaluator.py
Robust context budgeting	core/brain/llm/context_gate.py
Structured knowledge distillation	core/learning/formalizer.py

Current Integration State

• EnvironmentKernel is the canonical loop for autonomous environment runs.
• The kernel starts a durable run record, seeds the same HTN planner used by policy, observes through the adapter, compiles state, updates belief/spatial memory, selects an intent, simulates it, routes governance, gates effects, compiles a command, validates closed-loop action semantics, records action budget pressure, executes, diffs pre/post state, attributes outcome, updates replay/abstraction/curriculum/ledger/procedural stores, emits trace rows, and closes the run with a postmortem when terminal failure/success is detected.
• Generic policies emit semantic intents. Environment-specific code is limited to adapters, parsers/compilers, and command compilers.
• Authority-required actions now fail closed if no authority gateway is connected.
• Spatial memory is metadata-rich but remains backward-compatible with legacy kind-string lookups.
• External proof now distinguishes strict-real, simulated-canary, and fixture evidence, and refuses placeholder/stub adapters as capability proof.
• Activation audit now starts and samples the live proof-kernel bridge, lock watchdog, and concurrency health monitor. The proof bridge reports an explicit claim scope and does not treat engineered proxies as subjective consciousness proof.
• The generic terminal-grid compiler parses ASCII/ANSI-like screens into the shared ontology. NetHack-specific code remains an adapter/parser/command translation layer, not shared strategy.
• Repository repair scripts are archived under archive/repair_scripts/ and are no longer root-level canonical runtime artifacts.

NetHack Stress Adapter

Run the strict real stress loop:

python challenges/nethack_challenge.py --mode strict_real --steps 5000

Run the deterministic safe canary:

python challenges/nethack_challenge.py --mode simulated --steps 100

NetHack-specific modules are not strategy organs. They provide terminal-grid observation and command translation so the general kernel can be tested in a harsh, long-horizon environment.

Verification

Current focused suite:

python -m pytest \
  tests/test_final_general_hardening.py \
  tests/test_environment_general_integration.py \
  tests/environment/final_blockers \
  tests/environments/terminal_grid/test_nethack_audit_comprehensive.py \
  tests/environments/terminal_grid/test_terminal_grid_live_canary.py \
  tests/environments/terminal_grid/test_nethack_adapter_preflight.py \
  tests/environments/terminal_grid/test_terminal_grid_contract.py \
  tests/architecture \
  tests/test_embodied_cognition_runtime.py \
  tests/test_runtime_stability_edges.py \
  tests/test_runtime_service_access.py \
  tests/nethack_crucible.py -q

Focused final-pass result: 271 passed, 1 subtests passed.

Stress canary:

python challenges/nethack_challenge.py --mode simulated --steps 20 \
  --trace artifacts/test_nethack_kernel_trace.jsonl --log-level ERROR

Result: passed, with 40 hash-chained trace rows emitted to artifacts/test_nethack_kernel_trace.jsonl.

Strict-real smoke:

python challenges/nethack_challenge.py --mode strict_real --steps 40 \
  --trace /tmp/aura_strict_probe_after_threat_response.jsonl --log-level INFO

Result: reached a live dlvl_1 run, resolved startup modals, opened a door, moved, handled information modals, and stayed alive through 40 steps. This is not an ascension and is not counted as task mastery.

No successful strict-real ascension is recorded in this document.